CoreWeave CRWV NASDAQ

CoreWeave's record describes a company that used its public listing less as a source of equity and more as the entry point to the debt markets. The IPO raised $1.5 billion at $40 (claim #19821), with net proceeds of $1.4 billion (claim #4967). Since then, nearly all of the capital has come from secured delayed-draw facilities and high-yield notes. The DDTL sequence grows from $2.6 billion (claim #26685) to $3.0 billion (claim #4277) to $8.5 billion (claim #11218). The company's own aggregate figures, more than $25 billion over 18 months (claim #26689) and $28 billion in commitments over twelve months (claim #11221), show how small the IPO is within the total.

The mix looks deliberate rather than improvised. Contract-backed GPU spending goes into secured DDTLs, which earned A3/A(low) ratings on the latest structure (claim #11219). Unsecured notes and strategic equity fill the gaps around them. That equity comes from Nvidia at $2 billion (claim #17594), Jane Street at $1 billion (claim #10305) and a customer issuance of $350 million (claim #4514).

The unsecured layer is getting more expensive. Coupons went from 9.250% (claim #29433) to 9.000% (claim #26755) and then to 9.750% (claim #23816). Dilution is coming more through convertibles than through direct issuance. The December 2025 convertible was $2.25 billion (claim #18733), partly offset by a $295.7 million capped call (claim #18738). A proposed $3.0 billion convertible due 2032 followed in April 2026 (claim #10325). Earlier claims put that deal at $2.6 billion (claim #11765), so the sizing was revised and requires explanation.

The record does not show where the money was actually spent, only stated intentions. One stated use repeats. Both the 2030 and 2031 note offerings name a $450 million revolver repayment (claims #29778 and #26759), even though $444 million of additional revolver borrowing appears in between (claim #29780). That suggests the revolver is being cycled rather than retired.

The item that most deserves follow-up is the January 2026 covenant amendments. These postponed debt-service-coverage testing (claim #3912) and contract-realization testing (claim #3913), and adjusted minimum liquidity (claim #3911). A reader should watch for several things: when that testing resumes; whether draws on DDTL 4.0 track the Meta commitment of $21 billion (claim #2124) and the OpenAI commitment of $6.5 billion (claim #4297); whether the next unsecured coupon prices above 9.75%; and how the convertibles are settled. No insider transactions are on record, so that dimension cannot be assessed.

CoreWeave's record shows a company financed mainly through debt, with equity used sparingly. After the $40 IPO (claim #11616) and $1.4B of net proceeds (claim #4967), new equity came mostly from strategic holders rather than public follow-ons: a $350M customer share issuance (claim #4514), NVIDIA's $2B Class A purchase (claim #17594), and Jane Street's $1B investment (claim #10305). Dilution from straight equity has been modest relative to the balance sheet. The more meaningful future dilution sits in the convertibles: the December 2025 issue, recorded at $2.25B (claim #18733), and the proposed $3.0B 2032 convert with a $450M option (claim #10325, #10326). The figures recorded for the first convert do not agree with each other, at $2B (claim #18739), $2.25B and $2,587.5M (claim #3980), and net proceeds of $2,210.6M versus $2,542.2M (claims #18737, #3981). These require reconciliation, most likely to option exercise. The $295.7M capped call (claim #18738) shows the company actively managing that overhang. The mix looks deliberate. Each senior note deal was upsized, from $1.5B to $2.0B at 9.250% (claims #29777, #29433) and to $1.75B at 9.000% (claims #26755, #4698), while asset-level delayed-draw facilities grew from $2.6B at SOFR+4% (claim #26686) to $8.5B (claim #11218) with A3/A(low) ratings (claim #11219). That progression suggests contract-backed borrowing is meant to replace high-coupon unsecured paper. The April 2026 notes at 9.750% (claim #23816) show that the unsecured cost of capital has not yet come down. On use of proceeds, the 2024 term loan repayment (claim #29779) is documented. The $450M revolver paydown, however, is cited for both the May and July 2025 note deals (claims #29778, #26759), which requires explanation. The postponed debt service coverage ratio testing (claim #3912), along with a contract realization postponement and a liquidity amendment that are not yet verifiable (claims #3913, #3911), indicate that some covenant timelines have been revised. Readers should watch whether DDTL 4.0 draws track the Meta commitment (claim #2124), whether coverage testing resumes on its revised schedule, and how the Series C put obligation (claim #4276) is resolved. No insider transactions appear in the record provided, and the claim list is truncated, so no conclusions are drawn on insider selling.

CoreWeave's record describes a company that has financed almost entirely through debt since listing, with equity arriving mainly from strategic counterparties rather than from the public market. After an IPO of $1.5B gross (claim #19821), or $1.4B net (claim #4967), the company issued high-yield notes at roughly two-month intervals, and each deal was upsized at pricing. The 2030 notes went from $1.5B proposed (claim #29777) to $2.0B priced at 9.250% (claim #29433). The 2031 notes went from $1.5B (claim #26765) to $1.75B at 9.000% (claim #26755). A later 2031 tranche of $2.75B carried 9.750% (claim #23816), so unsecured borrowing costs have not come down over the period. Direct dilution has come from convertibles and strategic placements: the December convertible of $2.25B base (claim #18733), the proposed $3.0B 2032 convertible (claim #10325), NVIDIA's $2B purchase of Class A stock (claim #17594), and Jane Street's $1B (claim #10305). The record contains no insider transactions, so nothing can be said about insider selling.

The overall mix follows a consistent pattern. Asset-level delayed-draw facilities fund contracted GPU deployments, culminating in the $8.5B DDTL 4.0 with A3/A(low) ratings (claims #11218, #11219), while unsecured and convertible paper sits at the holding company. That reads as a deliberate structure rather than patchwork. However, the January 2026 postponement of debt-service-coverage and contract-realization testing, together with the amendment to the minimum liquidity covenant (claims #3912, #3913, #3911), suggests the underlying contracts have not been converting to cash as the original terms assumed. That gap requires explanation.

The stated uses of proceeds are hard to trace through the record. A $450M revolver paydown appears in both the May and July note offerings (claims #29778, #26759), and the filings do not show whether these were separate repayments or the revolver was drawn again in between. The December convertible's figures are also internally inconsistent. Net proceeds appear as $2,210.6M (claim #18737) and as $2,542.2M (claim #3981), and capped-call costs appear as $295.7M (claim #18738) and as $340M (claim #3986). The likely explanation is pre- versus post-option figures, but the disclosures do not reconcile them.

Readers should next watch for three things: when deferred covenant testing resumes, how much of the 2032 convertible is upsized, and whether revenue is collected on the Meta commitment of $21B (claim #2124) and the OpenAI commitment of $6.5B (claim #4297). These contracts are what ultimately service a stack of more than $25B raised in 18 months (claim #26689).

CoreWeave's record shows a company that has funded itself mostly with debt rather than common equity, and it has done so at an accelerating pace. The IPO raised $1.5B gross (claim 19821) and $1.4B net (claim 4967). Later equity came in large single-counterparty blocks rather than broad issuance: NVIDIA's $2B (claim 17594), Jane Street's $1B (claim 10305) and a $350M strategic customer issuance (claim 4514). Direct dilution from common stock is therefore modest relative to total capital raised. The larger dilution question sits in the convertibles. The December 2025 deal grew from a $2.25B base (claim 18733) to $2.5875B (claim 3980), with a $340M capped call (claim 3986) bought to soften conversion dilution. An April 2026 proposal of $3.0B plus a $450M option (claims 10325, 10326) followed within four months.

The mix looks deliberate rather than patched together. Contract-backed delayed-draw facilities fund GPU capex against named commitments: DDTL 3.0 at $2.6B (claim 26685) and DDTL 4.0 at $8.5B (claim 11218), the latter carrying A3/A(low) ratings (claim 11219). Unsecured high-yield and convertible paper cover corporate needs. Two features complicate that reading. First, corporate coupons have drifted upward, from 9.250% (claim 29433) to 9.000% (claim 26755) to 9.750% (claim 23816). Second, the stated totals compound quickly, from $17.2B (claim 4968) to over $25B in 18 months (claim 26689) to $28B in twelve months (claim 11221).

On use of proceeds, both the 2030 and 2031 notes cite repaying the revolver, at $450M each (claims 29778, 26759). An additional $444M RCF borrowing appears alongside (claim 29780). This pattern suggests the revolver is being cycled as a bridge rather than retired, and that warrants explanation. Several aggregates in the record are recorded at inconsistent scales or conflict across claims, for example the $2.6B versus $3.0B convertible figures (claims 11765, 10325). Exact reconciliation therefore requires the primary filings.

A sophisticated reader should watch the covenant posture first. Debt service coverage and contract realization testing were postponed, and the minimum liquidity covenant was amended (claims 3912, 3913, 3911). Whether those tests resume on schedule or are revised again will say more than any new issuance. Beyond that, the reader should track how much of the $21B Meta commitment (claim 2124) and the $6.5B OpenAI commitment (claim 4297) converts into DDTL draws, and whether the next unsecured coupon prints above 9.75%. The record contains no insider transactions, so no read on insider behavior is possible from this data.

CoreWeave's record shows a company that went public mostly to open the capital markets, not to fund itself with equity. The IPO raised $1.4B net (claim #4967), and within two months the company was back in the market with debt. Holdco unsecured notes came in quick succession: $2.0B at 9.250% due 2030 (claim #29433), $1.75B at 9.000% due 2031 (claim #26755), and $2.75B at 9.750% due 2031 (claim #23816). Asset-level delayed-draw facilities grew alongside them, from $2.6B (claim #26685) to $3.0B (claim #4277) to $8.5B (claim #11218). The company's own cumulative figures rise from $17.2B (claim #4968) to over $25B in 18 months (claim #26689) to $28B in twelve months (claim #11221).

The mix looks deliberate rather than improvised, but it has split in two. The asset-level structure now carries A3/A(low) ratings (claim #11219), while the holdco unsecured coupon has not compressed and the latest print is the highest. Equity has come in as strategic placements rather than broad offerings: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). It has been paired with convertibles: $2.59B including the option (claim #3980), then a proposed $3.0B due 2032 (claim #10325). Capped calls (claim #3986) limit near-term dilution from the converts but do not remove it. The record contains no insider transactions, so nothing can be said about insider selling.

Proceeds are hard to trace. Use-of-proceeds claims exist (claims #29776, #26766), but the record only shows debt repayment. Two separate note deals each repaid the revolver by $450M (claims #29778, #26759), while additional revolver borrowing of $444M is recorded (claim #29780). This looks like recycling short-term debt, and it requires explanation alongside the stated capex purpose (claim #2138).

The January 2026 covenant amendments deserve close attention. Debt service coverage testing was postponed (claim #3912), and the minimum-liquidity and contract-realization changes are not yet verifiable (claims #3911, #3913). Readers should watch when those tests resume, whether the cost of new holdco notes moves toward the investment-grade asset-level debt, and how the convertibles' conversion terms (claim #3984) compare with the share price. The record also has duplicates, mixed units, and a truncated final entry (claim #9316), which limits precision.

CoreWeave's record describes a company that has financed itself overwhelmingly through debt since listing, with equity used sparingly and mostly placed with strategic counterparties. After the $1.5B IPO (claim 19821), common-stock dilution has come mainly from negotiated placements: the $350M strategic customer issuance (claim 4514), NVIDIA's $2B investment (claim 17594) and Jane Street's $1B (claim 10305). The convertibles carry dilution forward rather than avoiding it. The December 2025 convertible was sized at $2.25B at announcement (claim 18733) and $2,587.5M at completion (claim 3980), consistent with the full $337.5M option (claim 18735). Capped calls partly offset the conversion exposure. The proposed $3.0B 2032 convertible (claim 10325) would roughly double that latent share overhang within four months. The record contains no insider transactions, so nothing can be said about management selling against this backdrop.

The mix reads as a deliberate two-track structure rather than patching. Contract-backed delayed-draw facilities have grown from $2.6B (claim 26685) to $8.5B (claim 2137), the latter carrying investment-grade ratings (claim 11219). Unsecured senior notes are layered alongside them. The cost of the unsecured track has not improved. The 9.000% 2031 notes (claim 26755) were followed by 9.750% 2031 notes (claim 23816). Several offerings were upsized from their announced size, for example from $1.5B to $2.0B (claims 11921, 29433). This suggests demand is available, though at a price that has moved up rather than down.

Use of proceeds tracks deployment in the instances the record allows us to check. The 2024 term loan was repaid in full (claim 29779), and revolver paydowns were disclosed. The identical $450M revolver repayment appears in both the 2030 and 2031 note offerings (claims 29778, 26759), and that repetition requires explanation. Capped-call allocations differ between filings, at $295.7M (claim 18738) versus $340M (claim 3986), and those figures have not been reconciled.

A careful reader should watch the covenant amendments. Debt service coverage testing was postponed (claim 3912), the minimum liquidity covenant was amended (claim 3911), and the contract realization test is not yet verifiable (claim 3913). Whether these tests resume on schedule matters more than any single new facility. Two further items bear on that. Projected secured debt needed to service contracts was $8.9B in May 2025 (claim 29783), and draws against the DDTL 4.0 facility should be read against the Meta and OpenAI commitments (claims 2124, 4297). The final claim on record (claim 9316) is truncated and may duplicate the January NVIDIA investment.

CoreWeave's record shows a company that has financed itself mostly through debt since listing, with equity arriving in occasional large blocks. After the $1.5B IPO (claim #19821, net $1.4B per claim #4967), common-equity dilution came mainly through strategic placements rather than marketed follow-ons: the $350M strategic customer issuance (claim #4514), NVIDIA's $2B purchase (claim #17594), and Jane Street's $1B (claim #10305). The dilution that matters more over time is contingent. It sits in two convertible offerings roughly four months apart: December 2025, sized at $2.25B base and $2,587.5M with the option (claims #18733, #3980), and a proposed $3.0B due 2032 in April 2026 (claim #10325), which a parallel claim records as $2.6B (claim #11765). That discrepancy requires explanation. Capped calls offset some of the exposure, but the recorded cost differs, at $295.7M (claim #18738) versus $340M (claim #3986).

The debt mix looks more like a layered strategy than ad hoc patching. Contract-backed delayed-draw facilities scale steadily, from DDTL 3.0 at $2.6B (claim #26685) to DDTL 4.0 at $8.5B with A3/A(low) ratings (claims #11218, #11219). Unsecured notes fill the gaps between them. Their pricing has not improved: 9.250% (claim #29433), 9.000% (claim #26755), then 9.750% on the 2031 notes (claim #23816). The secured channel is getting cheaper and larger while the unsecured channel is not. The company's own aggregate figures, $17.2B raised (claim #4968), over $25B in 18 months (claim #26689), and $28B committed in twelve months (claim #11221), show the scale of that reliance.

On deployment, stated uses are partly circular. Two separate notes offerings each name a $450M revolver repayment (claims #29778, #26759), and the 2024 term loan was refinanced in full (claim #29779). New unsecured debt is therefore retiring prior borrowings as well as funding capex. That is not unusual, but it is worth tracking. The January 2026 amendments postponing debt-service-coverage and contract-realization testing and revising the minimum liquidity covenant (claims #3912, #3913, #3911) show that covenant terms were renegotiated within months of issuance.

The record contains no insider transactions, so nothing can be said about insider selling. A sophisticated reader should watch four things: when deferred covenant tests resume, whether the April convertible size is reconciled, the status of the Series C put repurchase obligation (claim #4276), and whether unsecured coupons stay near 9-10% as maturities cluster in 2030-2032. The final claim in this dataset is truncated, so later events may be missing.

CoreWeave's record after the listing is dominated by debt, not equity. The IPO raised $1.5B gross at $40, per claims #19821 and #11616, with net proceeds of $1.4B per claim #4967. Since then, direct equity dilution has come mainly from strategic placements: NVIDIA's $2B purchase of Class A stock (claim #17594) and Jane Street's $1B investment (claim #10305). The larger potential dilution sits in the convertibles. The December 2025 issue grew from $2.25B to $2.5875B (claims #18733 and #3980) and references roughly 30.0M underlying shares (claim #3984). A proposed $3.0B convertible due 2032 followed in April 2026 (claim #10325). Capped calls of about $340M (claim #3986) reduce but do not remove the overhang. The record contains no insider transactions, so that dimension cannot be assessed here.

The mix does show a structure. Contract-backed delayed-draw facilities fund GPU deployment, and the progression runs from DDTL 3.0 at SOFR+4% (claim #26686) to the $8.5B DDTL 4.0 carrying A3/A(low) ratings (claim #11219). Unsecured notes and convertibles sit at the corporate level above them. The unsecured layer is the part that requires explanation. Coupons moved from 9.250% (claim #29433) to 9.000% (claim #26755) and then to 9.750% on the April 2026 notes (claim #23816), so corporate-level borrowing costs have risen even as asset-level financing earned investment-grade treatment. Several offerings were upsized shortly after announcement, which reads as demand-led sizing rather than fixed planning.

On use of proceeds, the May and July 2025 notes both earmarked $450M to repay the revolver (claims #29778 and #26759). The revolver was drawn again in between (claim #29780), which suggests note proceeds are partly refinancing a facility that keeps being redrawn. In January 2026, debt service coverage ratio testing was postponed (claim #3912) and the minimum liquidity covenant was amended (claim #3911). Both changes sit uneasily beside the volume of new commitments, stated at $28B over twelve months (claim #11221).

A careful reader should watch for several things next: whether postponed covenant tests resume on schedule, whether the coupons on unsecured debt stabilize, how much of the convertible stack ends up settled in shares, and whether projected contract-secured debt needs, last stated at $8.8B (claim #26760), keep growing faster than disclosed revenue commitments. The supplied record is truncated after claim #9315, so later events are not reflected here.

CoreWeave's record is mostly a debt story, not a dilution story. Common equity enters at only a few points: the $1.5B IPO at $40 (claims #19821, #11616), netting $1.4B (claim #4967); a $2B NVIDIA purchase of Class A stock (claim #17594); and a $1B Jane Street investment (claim #10305). Dilution beyond that depends on two convertibles: the December 2025 offering, upsized from $2B (claim #18739) to $2.25B (claim #18733), and the proposed $3.0B 2032 notes (claim #10325). Both include capped calls, $295.7M per claim #18738, which suggests management is trying to limit conversion dilution rather than ignoring it.

The mix follows a coherent pattern, though each step has cost more. Unsecured coupons rose from 9.25% on the upsized $2.0B 2030 notes (claim #29433) to 9.0% on the 2031 notes (claim #26755) and then to 9.75% on the $2.75B 2031 notes (claim #23816). Over the same period, asset-level financing grew from the $2.6B DDTL 3.0 at SOFR+4% (claim #26686) to the $8.5B DDTL 4.0 (claim #11218), which carries A3/A(low) ratings at the facility level (claim #11219). The structure separates contract-backed secured debt from expensive corporate paper, and the company's own figures of $25B raised in 18 months (claim #26689) and $28B in 12 months (claim #11221) show the pace accelerating.

The record confirms some stated uses of proceeds. The 2024 term loan was repaid in full (claim #29779). Revolver paydowns of $450M appear after both the May and July note deals (claims #29778, #26759), which points to the revolver being drawn and refinanced repeatedly. Deployment of the remaining proceeds into capex is not documented here. The December convertible figures also need reconciling: net proceeds appear as $2.21B (claim #18737) and $2.54B (claim #3981), and the capped call as $295.7M and $340M (claim #3986). This likely reflects exercise of the purchasers' option, but the record does not say so.

The January 2026 amendments require explanation. They postponed debt-service-coverage and contract-realization ratio testing (claims #3912, #3913) and revised the minimum liquidity covenant (claim #3911). A sophisticated reader should watch for when that testing resumes, how much of DDTL 4.0's $7.5B initial capacity (claim #11220) is actually drawn, and whether coupons keep rising. No insider transactions are present in this record, and the claim list is truncated after May 2026, so later events are not covered here.

CoreWeave's record is overwhelmingly a debt story, and the reader should weigh it that way. Common equity dilution is concentrated in a few events: the IPO at $40 (claim #11616) with $1.5B gross (claim #19821) and about $1.4B net (claim #4967), a $350M strategic customer issuance (claim #4514), and two later private placements, NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Equity has not been the main funding channel. Most capital has come through senior notes, delayed-draw term loans and, since December 2025, convertibles. The convertibles are where dilution now builds. The December deal grew from $2B proposed (claim #18739) to $2.5875B (claim #3980), partly offset by capped calls of $295.7M (claim #18738) or $340M (claim #3986); those two figures differ and require reconciliation. A further $3.0B convertible due 2032 was proposed in April 2026 (claim #10325).

The mix looks deliberate rather than patched together. Asset-level DDTL facilities finance specific customer contracts (claim #2138), culminating in the $8.5B DDTL 4.0 with A3/A(low) ratings (claims #11218, #11219). Unsecured notes and convertibles sit at the parent level. The cost of that unsecured layer has not fallen. Coupons ran 9.25% on the 2030 notes (claim #29433), 9.0% on the 2031 notes (claim #26755) and 9.75% on the April 2026 notes (claim #23816). Repeated upsizing, from $1.5B to $2.0B (claims #29774, #29433) and from $1.5B to $1.75B (claims #26765, #26755), shows strong demand but not cheaper money.

The stated uses of proceeds can be partly traced. The $1B repayment of the 2024 term loan (claim #29779) and the $450M revolver paydowns (claims #29778, #26759) are disclosed. Most other use-of-proceeds claims are generic, and the record has no deployment data against them.

Three items warrant attention. First, the January 2026 postponement of debt-service-coverage and contract-realization ratio testing, together with an amended minimum-liquidity covenant (claims #3912, #3913, #3911), requires explanation, given that the March 2026 facility carries its own coverage covenant (claim #2141). Second, the cumulative raise claims rise from $17.2B (claim #4968) to over $25B (claim #26689) to $28B in twelve months (claim #11221). They use different windows and cannot be compared directly. Third, conversion of the 2025 and 2026 convertibles is the main path to future share dilution.

The record contains no insider transactions, and it is cut off in May 2026, so these conclusions cover only that window.

CoreWeave's record is dominated by debt rather than common equity, so dilution has accumulated slowly and mostly through equity-linked paper. After the $1.5B IPO (claim #19821), which netted roughly $1.4B (claim #4967), new shares came mainly from strategic placements: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). The more consequential dilution sits in the convertibles. The December 2025 issue grew from $2.25B (claim #18733) to $2.5875B (claim #3980), with conversion terms referencing roughly 30.0M shares (claim #3984). A further $3.0B convertible due 2032 followed in April 2026 (claim #10325). That is two large equity-linked issues within about four months.

The mix reads as a deliberate structure more than patchwork. Contract-backed delayed-draw facilities fund GPU capex tied to specific customers (claim #2138). DDTL 4.0 at $8.5B carries A3/A(low) ratings (claims #11218, #11219). Unsecured notes and converts sit above that layer. The terms, however, have not improved with scale. Coupons went from 9.250% (claim #29433) to 9.000% (claim #26755) and then up to 9.750% on the April 2026 notes (claim #23816). The January 2026 covenant changes also require explanation: debt service coverage testing was postponed (claim #3912), and amendments to contract realization testing and minimum liquidity were recorded (claims #3913, #3911). Both sit alongside an investment-grade rating at the facility level.

Use of proceeds can be traced in places. Both 2025 note offerings earmarked $450M to repay the revolver (claims #29778, #26759), and the 2024 term loan was repaid in full (claim #29779). Revolver repayment appearing twice in two months suggests the revolver works as a bridge that is drawn and then refinanced. The capped call allocation is stated as $295.7M (claim #18738) and also as $340M (claim #3986), which likely reflects the base deal versus the upsized deal. The July 2025 claims on convertible-note conversion and a $270 termination fee (claims #4737, #4739) are no longer discussed in later records.

The disclosed record contains no insider transactions, so that dimension cannot be assessed. Unit conventions vary across claims, and several events appear in duplicate.

Going forward, the most informative items are the following. First, whether covenant testing resumes on the revised schedule. Second, whether the April 2026 $1B notes closed as expected (claim #9699). Third, whether claim #9316 records a second NVIDIA investment or restates the first. Fourth, how much of the 2031–2032 maturity wall is ultimately settled in shares rather than cash.

CoreWeave's record describes a company that has financed itself mostly through debt, with equity held in reserve. After the $1.5B IPO at $40 per share (claim #19821, #15974), the issuer turned almost immediately to high-yield unsecured paper. Those issues ran from the 9.250% notes due 2030, upsized from $1.5B to $2.0B (claims #29774, #29433), through the 9.000% notes due 2031, upsized to $1.75B (claim #26755), to a 9.750% $2.75B tranche (claim #23816). Each of the first two offerings was upsized between announcement and pricing, which points to strong demand rather than difficulty placing the paper. Coupons have not compressed, however, and the most recent tranche on record carries the highest rate. Direct equity dilution has come mainly through strategic placements: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Potential dilution sits in the convertibles: the December issue, whose size is reported as $2.0B, $2.25B and $2.5875B across claims #18739, #18733 and #3980, and a proposed $3.0B 2032 convert (claim #10325). Capped calls offset part of that exposure (claim #18738).

The mix reads as deliberate rather than improvised. Secured delayed-draw facilities tied to customer contracts scaled from DDTL 3.0 at $2.6B (claim #26685) to DDTL 4.0 at $8.5B with A3/A(low) ratings (claims #11218, #11219). Those facilities sit against backlog such as the OpenAI and Meta commitments (claims #4297, #2124). Unsecured notes and converts fund the parent. Use of proceeds is traceable in the early period, since notes were applied to the revolver and term loan as stated (claims #29778, #29779). For later issues the record carries only 'intended use' language, and actual deployment cannot yet be confirmed.

The items that most require explanation are the covenant changes dated January 2026. Debt-service-coverage and contract-realization testing were postponed, and the minimum liquidity covenant was amended (claims #3911, #3912, #3913). Readers should watch whether that testing resumes on schedule and how DDTL draws track contracted revenue. They should also watch the share count once the converts approach their conversion terms (claim #3984) and whether coupons on new unsecured paper ease. No insider transactions appear in the record provided, and the claim list ends mid-entry, so this read covers only the disclosed portion.

CoreWeave's record is overwhelmingly a debt story, with equity used sparingly. After the IPO at $40 (claim 11601) raising $1.5B gross (claim 19821) and roughly $1.4B net (claim 4967), the company did not return to straight common equity markets. It instead layered high-yield notes, delayed-draw term loans and convertibles. The only subsequent primary equity came from strategic counterparties: NVIDIA's $2B (claim 17594) and Jane Street's $1B (claim 10305). Whether the May 2026 NVIDIA entry (claim 9316) is a second tranche or a restatement of the January investment is not clear from the record and requires explanation. Dilution so far has therefore been modest and concentrated, with the larger latent source being the convertibles: the December 2025 issue at $2.25B base (claim 18733), $2,587.5M with the option (claim 3980), and a proposed $3.0B 2032 convertible (claim 10325). The capped calls (claim 18738) partly offset that overhang.

The mix reads as coherent in its architecture but increasingly heavy in its pace. Stated capital raised moved from $17.2B (claim 4968) to over $25B in 18 months (claim 26689) and then $28B in commitments over 12 months (claim 11221). Unsecured coupons have not compressed. The 2030 notes priced at 9.250% (claim 29433), the 2031 notes at 9.000% (claim 26755), and a later 2031 series at 9.750% (claim 23816). Against that, the $8.5B DDTL 4.0 (claim 11218) carried A3/A(low) ratings at the facility level (claim 11219). That suggests contract-backed secured financing is becoming the cheaper channel. Repeated upsizing, $1.5B to $2.0B (claims 29774, 29433) and $1.5B to $1.75B (claims 26758, 26755), indicates demand rather than strain.

On use of proceeds, the record shows the refinancing legs clearly: a $450M revolver paydown (claim 29778) and the full repayment of the 2024 Term Loan (claim 29779). It does not show actual capex deployment against the projected $8.9B of secured debt for contracts (claim 29783), so that link cannot be confirmed here. The record contains no insider transaction data.

The January 2026 postponement of debt service coverage ratio testing (claim 3912) and contract realization ratio testing (claim 3913) requires explanation. It sits awkwardly beside the rapid issuance that followed. Readers should watch whether those tests resume on schedule. They should also watch how the conflicting April 2026 sizes are reconciled: $3.0B versus $2.6B for the convertible (claims 10325, 10327), and $1.25B versus $1B for the notes (claims 2125, 9697). Finally, watch whether the coupon trend on unsecured paper stabilizes.

CoreWeave's record shows a company that has rarely gone more than six weeks without a capital event since its listing. The direct equity component has stayed modest relative to the whole. The IPO raised $1.5B at $40 (per claims #19821 and #11616), followed by a $68M over-allotment (claim #29782) and a $350M strategic customer issuance (claim #4514). The heavier equity layers came later: NVIDIA's $2B Class A investment (claim #17594) and Jane Street's $1B (claim #10305). Dilution from these is steady rather than abrupt. The more consequential overhang sits in the convertibles. A $2.25B base offering grew to roughly $2.59B with the option (claims #18733 and #3980). A further $3.0B convertible due 2032 was proposed in April 2026 (claim #10325), though claim #10327 records that offering at $2,600, a discrepancy that requires explanation.

The mix reads as a deliberate layering strategy rather than patching. Asset-level delayed-draw facilities fund contract capex, culminating in the $8.5B DDTL 4.0 with A3/A(low) ratings (claims #11218 and #11219). Above them sits corporate high-yield and convertible paper. The corporate tier has a cost trajectory worth noting. Coupons moved from 9.250% on the 2030 notes (claim #29433) to 9.000% on the 2031s (claim #26755), then to 9.750% on a further 2031 tranche (claim #23816). Both 2025 note deals were upsized from their announced sizes (claims #29777 versus #29433, and #26758 versus #26755). Upsizing suggests demand, though it also enlarged the fixed-charge base.

On deployment, the disclosed uses largely track. The 2024 Term Loan was repaid in full at $1B (claim #29779), and revolver paydowns of $450M were stated for both the May and July note deals (claims #29778 and #26759). The capped call allocation, however, appears as both $295.7M (claim #18738) and $340M (claim #3986), likely reflecting pre- and post-option sizing. The April 2026 senior notes are recorded as $1.25B (claim #2125), $1B completed (claim #2056), and $2.75B (claim #23816), which is not reconciled in the record.

The item a careful reader should weigh most is the January 2026 amendment. It postponed debt service coverage and contract realization ratio testing and revised a minimum liquidity covenant (claims #3911 through #3913), and the latter two remain not yet verifiable. Readers should watch for when that testing resumes. They should also watch how drawn DDTL balances compare with the backing commitments, such as Meta's $21B (claim #2124) and OpenAI's $6.5B (claim #4297). The record contains no insider transactions, so no read on insider positioning is possible.

CoreWeave's record since its March 2025 listing shows a company that has relied mainly on debt rather than on share issuance. The equity it has taken in comes from a small number of large blocks: the IPO at $40 (claim 11601), which raised about $1.4B net (claim 4967), NVIDIA's $2B Class A purchase (claim 17594), and Jane Street's $1B investment (claim 10305). A further NVIDIA $2B entry in May 2026 (claim 9316) needs to be explained. It is unclear whether it is a second tranche or a restatement of the January investment. Most dilution risk instead sits in the convertibles: roughly $2.59B issued in December 2025 (claim 3980) and a proposed $3.0B due 2032 (claim 10325). Both are partly offset by capped calls, which the record sizes at $340M (claim 3986) in one place and $295.7M (claim 18738) in another.

The mix looks like a deliberate plan rather than patchwork. Unsecured high-yield notes, asset-level delayed-draw facilities and convertibles each serve a separate purpose. The DDTL 4.0 facility carries A3 ratings (claim 11219), which suggests contract-backed borrowing is getting cheaper. Unsecured debt has not followed: coupons went from 9.250% (claim 29433) to 9.000% (claim 26755) and then 9.750% (claim 23816). Several deals grew after launch. The $1.5B 2030 notes became $2.0B (claims 29777, 29433), and the $1.5B 2031 notes became $1.75B (claims 26758, 26755). That shows strong investor demand, but it also means the company took on more debt than it first planned.

The record supports some stated uses of proceeds better than others. Repayment of the $1B 2024 term loan is recorded (claim 29779), as is a $450M revolver paydown (claim 29778). However, the same $450M revolver repayment appears again in the July offering (claim 26759), which suggests the revolver is being drawn and repaid in cycles. The record does not show where the remaining note proceeds were actually spent.

The most important items to watch are the covenant changes. Debt service coverage testing was postponed (claim 3912), contract realization ratio testing was also postponed (claim 3913), and the minimum liquidity covenant was amended (claim 3911). Readers should check when these tests resume, how the contract backlog tied to Meta's $21B commitment (claim 2124) is converted into revenue, and whether further convertible issuance moves dilution from a contingent risk to an actual one.

The record shows a balance sheet built mainly with debt. Equity dilution is real but secondary. After the $40 IPO (claim 11616), which raised net proceeds of about $1.4B (claim 4967), nearly every subsequent financing was secured or unsecured debt. Equity-linked issuance appears only twice. The first is the December 2025 convertible, sized at $2,587.5M including the option (claim 3980) and referencing roughly 30.0M conversion shares (claim 3984). The second is a proposed $3.0B 2032 convertible in April 2026 (claim 10325). Strategic equity placements make up the rest of the dilution: NVIDIA's $2B (claim 17594) and Jane Street's $1B (claim 10305). Insider transactions do not appear in the supplied record, so no read on insider selling is possible here.

The mix looks like a deliberate program, not patchwork. High-yield notes were upsized twice, from $1.5B to $2.0B at 9.25% (claims 29774, 29433) and from $1.5B to $1.75B at 9.0% (claims 26765, 26755). In parallel, the company stacked contract-backed delayed-draw facilities, culminating in the $8.5B DDTL 4.0 with A3 ratings on the HPC financing (claims 11218, 11219). Pricing has not improved on the unsecured side, however. The $2.75B 2031 notes carry 9.750% (claim 23816), above both 2025 coupons.

Use of proceeds tracks partially. The 2030 notes were directed to repaying the 2024 term loan in full (claim 29779) and $450M of revolver borrowings (claim 29778). The 2031 notes cite the same $450M revolver repayment (claim 26759). This suggests the revolver was redrawn between offerings, which requires explanation. The capped-call allocation also appears as both $340M (claim 3986) and $295.7M (claim 18738).

The item most in need of follow-up is the January 2026 postponement of debt service coverage ratio testing (claim 3912). Alongside it sit an amended minimum liquidity covenant (claim 3911) and a contract realization ratio deferral (claim 3913), both not yet verifiable. Readers should watch for when that testing resumes and whether it is deferred again. They should also track the pace of DDTL draws against the $8.8–8.9B of projected secured debt (claims 26760, 29783), further drift in unsecured coupons, and how the 2030–2032 maturities stack up. The supplied record is truncated after claim 9315, so later events are not reflected here.

CoreWeave's record describes a company financing itself primarily through debt, with equity serving as a smaller supplement. The $1.5B IPO, per claim #19821, which netted $1.4B per claim #4967, is modest beside the more than $25B in total capital raised over 18 months cited in claim #26689 and the $28B in commitments over 12 months cited in claim #11221. Direct dilution has come in discrete, strategic tranches rather than a steady drip: NVIDIA's $2B, per claim #17594, and Jane Street's $1B, per claim #10305. The more consequential dilution channel is contingent. The December convertible, sized at $2.25B per claim #18733 and $2.5875B including the option per claim #3980, was followed by a proposed $3.0B convertible due 2032, per claim #10325, alongside an earlier-reported $2.6B figure in claim #11765. The record does not reconcile those two sizes, and that requires explanation. Capped calls, at $295.7M per claim #18738 or $340M per claim #3986, show an attempt to manage conversion dilution, though the two figures differ.

The mix looks coherent rather than improvised. Unsecured notes fund corporate needs and repay revolver balances, delayed-draw term loans fund contract-specific capex, and convertibles and strategic equity fill in between. Upsizing has been routine. The 2030 notes moved from $1.5B, per claim #29777, to $2.0B, per claim #29433, and the 2031 notes from $1.5B, per claim #26758, to $1.75B, per claim #26755. Pricing has not improved with scale. Coupons ran 9.250%, per claim #29433, then 9.000%, per claim #26755, then 9.750%, per claim #23816. The investment-grade rating on the DDTL 4.0 structure, per claim #11219, suggests the secured layer is where cost relief is being pursued.

Stated uses of proceeds broadly track deployment. The $1B 2024 term loan repayment, per claim #29779, and the $450M revolver paydowns, per claims #29778 and #26759, match their announced purposes. The record contains no disclosed insider sales to assess.

The items to watch are the covenant changes. Debt service coverage testing was postponed, per claim #3912. Contract realization testing was also postponed and the minimum liquidity covenant amended, per claims #3913 and #3911, though both remain unverified. When these tests resume, they will show whether the anchor contracts convert to cash on schedule: OpenAI's $6.5B, per claim #4297, and Meta's $21B, per claim #2124. Readers should also watch whether the $8.9B of projected secured debt, per claim #29783, stays near that level.

CoreWeave's record describes a company that has financed itself mostly through debt rather than through dilution. The equity base was set by the $1.5B IPO at $40 (claims 19821, 11601), which netted roughly $1.4B (claim 4967) plus a $68M over-allotment (claim 29782). After that, new shares came only through negotiated placements: the $2B NVIDIA purchase (claim 17594) and the $1B Jane Street investment (claim 10305). Dilution so far is episodic and tied to strategic counterparties, not to a steady market program. No ATM program and no insider sales appear in the record provided. The record also cuts off after claim 9315, so later events may be missing.

The debt layering is coherent in structure but rising in cost. Delayed-draw facilities fund specific customer contracts, growing from $2.6B (claim 26685) to $3.0B (claim 4277) to the $8.5B DDTL 4.0 (claim 11218), which carries A3/A(low) ratings on the HPC financing (claim 11219). Unsecured notes sit above that layer, and their coupons have moved upward: 9.000% on the 2031s (claim 26755), 9.250% on the 2030s (claim 29433), and 9.750% on a later 2031 tranche (claim 23816). Convertibles were added starting December 2025 at $2,587.5M including the option (claim 3980). Upsizing is routine: the 2030 notes went from $1.5B to $2.0B (claims 29774, 29433), and the 2031 notes rose by $250M (claim 4698).

Stated use of proceeds mostly tracks what was reported. The term loan was repaid in full (claim 29779). However, the same $450M revolver repayment appears for both the May and July note offerings (claims 29778, 26759), and it is unclear whether these are separate draws. The April 2026 records also require reconciliation. The convertible appears as $3.0B (claim 10325) and as $2.6B (claim 10327). The senior notes appear as $1.25B, $1B, and $2.75B (claims 2125, 2056, 23816).

The item most in need of explanation is the January 2026 covenant package. It amended the minimum-liquidity covenant (claim 3911) and postponed both the debt-service-coverage and contract-realization ratio tests (claims 3912, 3913). This came one month before the DDTL 4.0 closed and while backlog was growing, including the $21B Meta commitment (claim 2124).

A careful reader should track four things next: when the postponed covenant tests resume, how much of DDTL 4.0 is drawn beyond its $7.5B initial capacity (claim 11220), whether coupons on new notes keep rising, and how much potential share issuance is embedded in the two convertible series. Cumulative raises reported at $17.2B (claim 4968), then over $25B (claim 26689), then $28B (claim 11221) show the pace of capital intake continuing to accelerate.

CoreWeave's record describes a company that has financed itself mostly through debt. Common-equity dilution has been modest relative to the total capital raised, and the more consequential dilution risk sits in the convertible layer. After the $1.5B IPO (claim #19821) and the $68M over-allotment (claim #29782), new common equity came mainly from strategic holders: a $350M customer issuance (claim #4514), NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Meanwhile, management describes more than $25B raised in 18 months (claim #26689) and $28B in commitments over the past 12 months (claim #11221). The pace of financing has quickened rather than slowed.

The mix looks deliberate, not reactive. Asset-level delayed-draw facilities fund contracted GPU deployments, scaling from $2.6B (claim #26685) to $8.5B (claim #11218), with the last carrying A3/A(low) ratings (claim #11219). Unsecured high-yield notes and convertibles sit above that layer. The cost of that upper layer requires attention. Coupons ran from 9.000% (claim #26717) to 9.250% (claim #29421) to 9.750% (claim #23816). Convertibles then followed: $2.25B in December (claim #18733) and a proposed $3.0B due 2032 (claim #10325). Moving from straight debt to convertibles can be read as an effort to contain cash interest, at the cost of future share count. The capped call (claim #3986) is meant to soften that.

Where use of proceeds can be traced, it broadly matches what was announced. The notes were described as repaying the revolver (claim #29778), and the 2024 term loan was reported repaid in full (claim #29779). This is refinancing rather than new capacity, which readers should weigh when adding up headline totals. Three items require explanation. The record shows DSCR testing postponed (claim #3912) and a minimum-liquidity amendment that is not yet verifiable (claim #3911). A $270M termination fee payable to CoreWeave (claim #4739) has no recorded follow-through. No insider transactions appear on the record, so no conclusion can be drawn there.

The record itself is uneven. Figures are mixed between millions and full dollars, and the convertible sizing is reported inconsistently: net proceeds appear as $2.21B (claim #18737) and as $2.54B (claim #3981), probably before and after the option exercise.

Readers should watch for four things: when deferred covenant testing resumes, how far draws on DDTL 4.0 track the Meta (claim #2124) and OpenAI (claim #4297) commitments, whether the next unsecured coupon prices above 9.750%, and how much convertible share count is outstanding against the capped-call strikes.

CoreWeave's record describes a company that has financed itself mostly through debt, with equity arriving in a few large, discrete events. After the $1.5B IPO (claim #19821, net $1.4B per claim #4967), every subsequent quarter added at least one debt instrument. These include the $2.0B 9.250% notes due 2030 (claim #29421), the $1.75B 9.000% notes due 2031 (claim #26717), the $2.6B DDTL 3.0 (claim #26685), the upsized convertible of up to $2.5875B (claim #3980), the $8.5B DDTL 4.0 (claim #11218), the $2.75B 9.750% notes (claim #23816), and a further $1B of notes (claim #2056). Direct share dilution has been modest by comparison and has come from strategic holders rather than the open market: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). The convertibles, however, carry deferred dilution, which the company has partly hedged with capped calls ($295.7M of proceeds allocated per claim #18738).

The mix looks deliberate rather than improvised. Asset-level delayed-draw facilities fund contracted GPU deployments, and the stated basis for DDTL 4.0 is customer contract capex (claim #2138). That facility drew A3/A(low) ratings (claim #11219), which suggests contract-backed collateral is reaching cheaper capital. At the corporate level, though, unsecured coupons have not compressed: they went from 9.250% to 9.000% and then to 9.750% over roughly ten months.

On use of proceeds, the early notes went to the stated purposes, including repaying the 2024 term loan (claim #29779) and the revolver (claim #29778). A $450M revolver repayment appears again as the use of proceeds for the 2031 notes (claim #26759), after an additional $444M RCF draw (claim #29780). That sequence suggests the revolver is being cycled as a bridge, which is reasonable but worth tracking. The record contains no insider sales, so that dimension cannot be assessed here.

The items that require explanation sit in covenant terms. The record shows a postponement of debt service coverage ratio testing (claim #3912), a postponement of contract realization ratio testing (claim #3913), and a minimum liquidity amendment (claim #3911), though two of these are not yet verifiable. Looking ahead, a reader should check whether those tests resume on schedule. They should also look for a closing record for the proposed $3.0B 2032 convertible (claim #10325), since only the proposal appears so far. Two other open points are whether the May 2026 NVIDIA entry (claim #9316) is a second investment or a restatement of the January one, and whether unsecured coupons move toward the investment-grade pricing achieved at the facility level.

CoreWeave's record shows a company that used its listing as a credential for borrowing rather than as its main source of capital. The IPO priced at $40 (claim 11616) and raised $1.5B gross (claim 19821), about $1.4B net (claim 4967). That is small next to the company's own aggregate figures: over $25B raised in 18 months (claim 26689), $28B committed in the twelve months to March 2026 (claim 11221), and more than $20B year-to-date by May 2026 (claim 9123). Dilution has been modest in direct share issuance and concentrated in strategic placements and convertibles: NVIDIA's $2B (claim 17594), Jane Street's $1B (claim 10305), and convertible programs whose stated conversion terms reference roughly 30.0M shares (claim 3984), with a further $3.0B due 2032 proposed (claim 10325). The record contains no ATM program and no insider sales, so neither can be assessed.

The debt mix looks deliberate rather than patched together, but it has grown more expensive in some layers. Unsecured coupons moved from 9.250% (claim 29433) to 9.000% (claim 26755) and then to 9.750% on the April 2026 tranche (claim 23816). Meanwhile, the contract-backed delayed-draw facilities moved toward investment-grade treatment: the A3/A(low) ratings on DDTL 4.0 (claim 11219), and DDTL 5.0, described as the first publicly syndicated HPC infrastructure-backed financing (claim 9116, claim 9117). This is a two-track structure: cheaper secured capital tied to customer contracts, and costly unsecured paper filling the gaps. Repeated upsizing, such as the $250M increase on the 2031 notes (claim 4698), suggests investor demand has been available.

Use of proceeds is only partly traceable. The $450M revolver repayment from note proceeds (claim 26759) and the DDTL financing of contract capex and GPU servers (claim 2138) are specific. Most other use-of-proceeds claims are stated without detail, so deployment cannot be checked against intent from this record. The December 2025 convertible figures also need reconciling. The offering appears as $2.25B (claim 18733) and $2.5875B (claim 3980), and the capped call as $295.7M (claim 18738) and $340M (claim 3986), most likely before and after the purchasers' option.

The January 2026 postponement of debt-service-coverage and contract-realization testing (claim 3912, claim 3913) is the item that most needs explaining. A sophisticated reader should watch when those tests resume, how much of the $8.8B in projected secured debt (claim 26760) is actually drawn, whether unsecured coupons keep rising, and whether convertible settlement turns into real share issuance.

CoreWeave's record shows a company that went public mainly to unlock the debt markets. The equity raised at the March 2025 listing was modest, about $1.4B net per claim #4967. Debt then followed at a steady pace. The $2.0B 9.250% 2030 notes were upsized from $1.5B (claims #11921, #29433). The $1.75B 9.000% 2031 notes were also upsized from $1.5B (claims #26758, #26755). The delayed-draw facilities grew from $2.6B for DDTL 3.0 (claim #26685) to $8.5B for DDTL 4.0 (claim #11218), followed by $3.1B for DDTL 5.0 (claim #9116). The company's own aggregate figures accelerate over time: over $25B in 18 months (claim #26689), $28B over twelve months (claim #11221), and more than $20B year-to-date by May 2026 (claim #9123).

The financing mix reads as a coherent strategy rather than patching. Contract-backed secured facilities fund customer capex (claim #2138), and that structure earned A3/A(low) ratings at the facility level (claim #11219). Unsecured notes and convertibles serve as corporate-level filler. Equity dilution has come mostly through negotiated placements rather than an ATM; none appears on record. The placements were $2B from NVIDIA (claim #17594) and $1B from Jane Street (claim #10305).

Two threads complicate that reading. The January 2026 covenant amendments postponed debt service coverage testing (claim #3912) and contract realization testing (claim #3913), and those postponements require explanation. Unsecured pricing has also not improved: the April 2026 notes carry 9.750% (claim #23816), compared with 9.000% the prior July. Within the April notes, the $2.75B figure (claim #23816) and the $1.25B and $1B figures (claims #2125, #9697) are not reconciled in the record.

On use of proceeds, the claims are announcements, not evidence of deployment. Two items can be partly traced: the $450M revolver repayment (claim #26759) and the capped-call allocation of $295.7M (claim #18738), later cited as $340M (claim #3986). Beyond those, the record cannot confirm whether proceeds were spent as stated.

A careful reader should watch several things. The first is when, and whether, the deferred covenant tests resume. The second is the conversion overhang from the December 2025 convertible (claim #18733) and the 2032 convertible (claim #10325). The third is the 2030 refinancing window. The fourth is how heavily the collateral depends on a small number of commitments, such as OpenAI (claim #4297) and Meta (claim #2124). The record contains no insider transactions and appears truncated after May 2026, so conclusions about insider behavior or later events are not supported.

CoreWeave's record describes a company that went public with a modest equity raise and has since financed itself mainly through debt. The IPO priced at $40 (claim #11616) and raised $1.5B gross (claim #19821), or about $1.4B net (claim #4967). After that, common equity appears only as two strategic placements: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Direct dilution has therefore been episodic and tied to counterparties rather than spread through the market, and no at-the-market program or insider sales appear in this record. The dilution question shifts to the convertibles. The December 2025 offering is cited as $2.25B (claim #18733) and also as $2,587.5M (claim #3980), with net proceeds of $2,210.6M (claim #18737) and $2,542.2M (claim #3981). The difference likely reflects the purchasers' option, but it requires reconciliation. A further $3.0B convertible due 2032 was proposed (claim #10325) with a $450M option (claim #10326). Capped calls of $295.7M (claim #18738) limit, but do not remove, eventual share issuance.

The mix looks more structured than opportunistic. Unsecured notes were upsized twice, from $1.5B to $2.0B (claims #11921, #29433) and from $1.5B to $1.75B (claims #26758, #26755). Asset-backed delayed-draw facilities grew from $2.6B at SOFR+4% (claims #26685, #26686) to an $8.5B facility rated A3 (claims #11218, #11219) and then a publicly syndicated $3.1B tranche (claims #9116, #9117). Against that, unsecured coupons have not compressed. They moved from 9.25% to 9.0% and then to 9.750% on the $2.75B 2031 notes (claim #23816). In January 2026 the company also postponed debt-service-coverage and contract-realization testing and amended a minimum liquidity covenant (claims #3911–#3913). None of these changes is explained in the record.

The record states the purposes of the borrowing but not how the money was spent. The delayed-draw facilities are tied to customer-contract capex and GPU servers (claim #2138). However, part of the July 2025 notes went to repaying $450M of revolver borrowings (claim #26759), so some new debt refinanced existing debt rather than funding new capacity. Large customer commitments, from OpenAI at $6.5B (claim #4297) and Meta at $21B (claim #2124), are the stated collateral logic, but the record contains no draw-down or capex reconciliation against them.

A careful reader should watch four things: whether the postponed covenant tests resume on schedule, whether unsecured coupons start to fall toward the investment-grade facility pricing, how the conversion terms in claim #3984 translate into share count, and whether disclosed contract revenue keeps pace with roughly $28B of financing in twelve months (claim #11221). This record ends mid-entry on the DDTL 5.0 claims, so later events are not reflected here.

CoreWeave's record describes a company that financed itself overwhelmingly with debt after a modest equity debut. The IPO raised $1.5 billion (claim #19821). Debt facilities followed at an accelerating pace, and management's own cumulative tallies climbed from over $25 billion in 18 months (claim #26689) to $28 billion in commitments over 12 months (claim #11221) and more than $20 billion year-to-date by May 2026 (claim #9123). Direct dilution came mainly through negotiated placements rather than market issuance: the $2 billion NVIDIA investment (claim #17594), the $1 billion Jane Street investment (claim #10305), and a $350 million strategic customer issuance (claim #4514). Potential dilution is building in the convertibles, the $2.25 billion December 2025 offering (claim #18733) and the proposed $3.0 billion 2032 notes (claim #10325). The record shows no ATM program and no insider sales, so neither can be assessed here.

The mix reads as a deliberate two-track structure rather than patching. Contract-backed delayed-draw facilities fund customer-specific capex; the $8.5 billion DDTL 4.0 carries A3/A(low) ratings (claim #11219) and is described as financing customer contract capital expenditures and GPU servers (claim #2138). Unsecured notes and converts sit at the corporate level. The unsecured layer has stayed expensive, however. Coupons moved from 9.250% (claim #29433) to 9.000% (claim #26755) and then 9.750% on the 2031 notes (claim #23816), so the investment-grade progress at the facility level has not yet reached the parent's cost of capital.

The record does not allow a meaningful test of whether stated use of proceeds tracked actual deployment. Use-of-proceeds claims appear repeatedly (claims #29436, #26757, #18743), but the only quantified allocations are a $450 million revolver repayment (claim #26759) and capped-call costs. Those capped-call costs require reconciliation: $295.7 million (claim #18738) against $340 million (claim #3986), alongside net proceeds of $2.21 billion (claim #18737) versus $2.54 billion (claim #3981), most plausibly reflecting pre- and post-upsize figures.

A sophisticated reader should watch three things. The first is the January 2026 postponement of debt-service-coverage and contract-realization testing (claims #3912, #3913), and whether that testing resumes on schedule under the newer facilities that carry the same covenant (claims #2141, #1324). The second is whether the projected $8.8 billion of additional secured debt (claim #26760) keeps pace with commitments such as Meta's $21 billion (claim #2124). The third is conversion economics on the 2032 notes. The record ends mid-entry on DDTL 5.0's use of proceeds, so that facility's purpose remains undisclosed here.

CoreWeave's financing record since its March 2025 listing describes a company that used equity sparingly and debt continuously. The IPO itself was modest relative to what followed, roughly $1.4B net per claim #4967. Corporate-level dilution after that came mainly from strategic placements, NVIDIA's $2B per claim #17594 and Jane Street's $1B per claim #10305, rather than from marketed follow-ons. The heavier dilution is latent: the $2.5875B convertible per claim #3980 and the proposed $3.0B convertible due 2032 per claim #10325 carry conversion exposure. The capped calls, funded with $295.7M of proceeds per claim #18738, are meant to reduce it.

The mix reads as a deliberate two-tier structure rather than patching. Contract-backed delayed-draw facilities fund GPU capex at the asset level, with proceeds tied to customer contracts per claim #2138. The program scaled from DDTL 3.0 at $2.6B per claim #26685 to DDTL 4.0 at $8.5B with A3/A(low) ratings per claim #11219. High-coupon unsecured notes and low-coupon converts sit above them at the corporate level. Coupons have not compressed with scale, however. They ran from 9.250% per claim #4824 to 9.000% per claim #26755 and then 9.750% per claim #23816, which requires explanation alongside the company's description of its growing scale. DDTL 5.0 was rated Ba2/BB+ per claim #8732, a step below its predecessor's investment-grade designation. The record does not show why.

Deployment can only be partly checked against stated uses. Part of the 2031 notes was earmarked to repay $450M of revolver borrowings per claim #26759, which means some new debt is refinancing prior debt. The January 2026 postponement of debt-service-coverage and contract-realization testing, per claims #3912 and #3913, indicates some facilities were not ramping on their original schedule. The record does not show actual drawdowns against any facility, so stated use cannot be fully reconciled with realized spending.

Several figures overlap without clear reconciliation. Convertible amounts appear as $2B, $2.25B and $2.5875B (claims #18739, #18733, #3980), and as both $2.6B and $3.0B (claims #10327, #2126). The aggregate capital-raised statements in claims #26689, #11221 and #9123 cover overlapping windows. The record contains no insider transactions, and the claim list ends at claim #8729, so it may be incomplete.

A sophisticated reader should watch for four things. The first is whether the deferred covenant tests resume on schedule. The second is the gap between roughly three-year underlying customer contracts per claim #8731 and loan maturity of about 5.5 years per claim #9120, which places weight on re-contracting. The third is whether unsecured coupons begin to decline. The fourth is how the share price trades relative to the convertibles' conversion terms.

The dominant feature of CoreWeave's record is that debt, not common equity, carried the build-out after the IPO. The listing itself was modest relative to what followed: roughly $1.5 billion gross at $40 (claim #19821, #11601), and about $1.4 billion net (claim #4967). Direct equity dilution has since come mainly from strategic placements, namely NVIDIA's $2 billion Class A purchase (claim #17594) and Jane Street's $1 billion (claim #10305). Those were negotiated with known counterparties rather than taken through open-market issuance, which mutes the immediate dilution optics. The more meaningful dilution vector is the convertibles: the December 2025 issue upsized to $2.59 billion (claim #3980) and the proposed $3.0 billion 2032 issue with a $450 million option (claim #10325, #10326). Both were paired with capped calls intended to soften conversion. The two capped-call figures on record, $295.7 million (claim #18738) and $340 million (claim #3986), differ and require explanation.

The mix reads as a deliberate two-track design rather than patching. Asset-level delayed-draw facilities are matched to customer contracts (claim #2138). DDTL 4.0 reached A3/A(low) ratings (claim #11219), while DDTL 5.0 was rated Ba2/BB+ (claim #8732). Corporate unsecured notes fund the remainder. The cost of that corporate layer has not improved: coupons of 9.25% (claim #4824), 9.0% (claim #26755) and then 9.750% (claim #23816). Repeated upsizing, from $1.5 billion to $2.0 billion (claim #11921, #11917) and from $1.5 billion to $1.75 billion (claim #26765, #26755), indicates demand was available at those prices, not that pricing tightened.

Stated uses of proceeds are partly traceable. Revolver repayment of $450 million is specified (claim #26759), and convertible proceeds partly fund capped calls (claim #18743). The record does not show actual deployment against the cumulative totals, which escalate from $17.2 billion (claim #4968) to over $25 billion (claim #26689) to $28 billion over twelve months (claim #11221).

A careful reader should watch three items. First, the postponed debt-service-coverage and contract-realization tests and the amended minimum-liquidity covenant (claim #3912, #3913, #3911). Second, whether customer contracts averaging about three years (claim #8731) can support loans of roughly 5.5 years (claim #9120). Third, the Series C put repurchase obligation (claim #4276). No insider transactions appear in the record, and the excerpt ends mid-entry, so conclusions on insider behavior are not possible.

CoreWeave's record shows a capital structure built mostly from debt, with dilution so far modest in direct terms and largely deferred. Direct equity issuance is limited to the IPO, with net proceeds of $1.4B per claim #4967, the $350M strategic customer issuance per claim #4514, NVIDIA's $2B investment per claim #17594, and Jane Street's $1B per claim #10305. The larger potential dilution sits in two convertible issues: the December 2025 offering, cited at $2.25B per claim #18733 and later at $2.5875B per claim #3980, and the proposed $3.0B 2032 convertible with a $450M option per claims #10325 and #10326. The capped calls, with $295.7M of proceeds allocated per claim #18738, suggest the company is trying to soften conversion dilution rather than ignore it. The mix reads as a coherent strategy, not patching. Asset-level delayed-draw facilities fund contract capex, as with the stated purpose of DDTL 4.0 per claim #2138 and DDTL 5.0 per claim #1320. Corporate unsecured notes and converts sit above them. The trajectory of terms requires explanation, however. Unsecured coupons moved from 9.000% on the 2031 notes per claim #26755 to 9.750% on a later 2031 issue per claim #23816. Meanwhile DDTL 4.0 achieved A3/A(low) ratings per claim #11219, while DDTL 5.0 was rated Ba2/BB+ per claim #8732. On use of proceeds, the July 2025 notes were tied to a $450M revolver repayment per claim #26759, yet the revolver was later expanded to $2.5B per claim #11766. That pattern points to cyclical terming-out of short-term borrowing more than deleveraging. The January 2026 postponement of debt service coverage and contract realization ratio testing per claims #3912 and #3913, alongside a minimum liquidity amendment per claim #3911, has not been revisited in subsequent claims on record. A sophisticated reader should watch whether those deferred covenant tests resume on schedule. They should also track the gap between underlying customer contracts averaging roughly three years per claim #8731 and loan maturities near 5.5 years per claim #9120, which implies dependence on renewals. A third item is whether aggregate raising, above $20B year-to-date per claim #9123, continues to outpace disclosed deployment. No insider transactions appear in this record, and the claims list ends mid-August 2026, so conclusions about insider behavior or recent activity cannot be drawn.

The record describes a company that has financed itself mostly through debt, with equity serving as a smaller secondary channel. After the $1.5B IPO at $40 (claim #19821), with net proceeds of about $1.4B (claim #4967), nearly every subsequent event is a secured term loan or a high-coupon note. The coupons are rising rather than compressing. They run from 9.000% on the 2031 notes (claim #26755) to 9.25% on the 2030 notes (claim #4824) to 9.750% on a later 2031 tranche (claim #23816), and the most recent DDTL 5.5 is priced at SOFR plus 5.50% (claim #8733). Direct dilution has come from strategic placements, namely NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). These placements are notable in size but small against stated cumulative raises above $30B year-to-date (claim #8735). The larger latent dilution sits in two convertible programs, the December 2025 notes (claim #18733) and the proposed $3.0B 2032 notes (claim #10325). The first was partly offset by capped calls costing $295.7M (claim #18738). The record does not show whether the second carried a similar hedge. The mix reads as a coherent template rather than patching. Contract-backed delayed-draw facilities (DDTL 3.0 through 5.5) fund GPU capex against named customer commitments (claim #2138), and the corporate-level notes and converts fill the gaps. The investment-grade A3 rating on DDTL 4.0 (claim #11219) suggests the asset-level structure is gaining acceptance with lenders. One pattern requires explanation. In January 2026, debt service coverage ratio testing was postponed (claim #3912), and contract realization ratio testing and the minimum liquidity covenant were also reportedly revised (claims #3913, #3911, not yet verifiable). Some stated uses are traceable, such as the $450M revolver repayment (claim #26759) and the capped call funding (claim #18743). Actual capex deployment against the facilities is not disclosed in this record, so tracking cannot be confirmed. Convertible sizing is reported inconsistently across claims #18733, #18739 and #3980, and the record is truncated after August 2026. It also contains no insider transactions, so no conclusions about insider behavior are possible. Readers should watch four things: whether the postponed covenant tests resume on schedule, the draw deadlines on DDTL 5.0 (claim #1321), whether coupons on new issuance keep rising, and the conversion price of the 2032 notes relative to the trading price.

CoreWeave's record is primarily a debt-financing story, with equity used sparingly. After the $1.5B IPO (claim #19821), which netted $1.4B (claim #4967), the company drew on nearly every debt channel within about eighteen months. These included the $2B 9.250% notes due 2030 (claim #11917), the upsized $1.75B 9.000% notes due 2031 (claims #4698, #26717), and successive delayed-draw facilities up to the $8.5B DDTL 4.0 (claim #11218) and $3.1B DDTL 5.0 (claim #9116). By August 2026 management described more than $30B raised year-to-date (claim #8735). Direct common dilution is modest relative to that scale. It consists of a $350M strategic share issuance (claim #4514), NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Prospective dilution sits mainly in two convertibles: the December 2025 offering of up to $2.5875B (claim #3980) and the April 2026 $3.0B due 2032 with a $450M option (claims #10325, #10326). Both were partly offset by capped calls (claim #18738). The record contains no insider sales, so that dimension cannot be assessed.

The mix reads as a coherent architecture. Contract-backed delayed-draw loans finance GPU capacity tied to specific customers (claims #2138, #1320), while unsecured notes and converts fund the corporate layer. The pricing trajectory requires explanation. Unsecured coupons moved from 9.000% (claim #26755) to 9.750% (claim #23816). Delayed-draw spreads widened from SOFR+4% on DDTL 3.0 (claim #26686) to SOFR+5.50% on DDTL 5.5 (claim #8733), even as DDTL 4.0 earned A3/A(low) ratings (claim #11219). January 2026 also brought postponed testing of the debt service coverage and contract realization ratios and an amended minimum liquidity covenant (claims #3912, #3913, #3911). These are revised terms, not breaches, but they are not discussed further in the record.

Stated uses of proceeds are specific in places. Examples include the $450M revolver repayment (claim #26759) and the capped call allocations. Actual deployment of the larger facilities cannot be verified from this data. The July 2025 projection of $8.8B in additional secured debt (claim #26760) was later exceeded by DDTL 4.0 and 5.0 alone.

Readers should watch several things. One is whether the postponed covenant tests resume on schedule. Another is whether pricing on new notes keeps rising. A third is how drawdowns track the $21B Meta (claim #2124) and $6.5B OpenAI (claim #4297) commitments. The last is where the stock trades relative to the convertible conversion terms. The record provided ends mid-entry on DDTL 5.5, so its full terms remain unassessed.

CoreWeave's record describes a company that used its public listing mainly as a gateway to the debt markets. The IPO raised $1.5B gross at $40 per share (claims #19821, #11616), with net proceeds of $1.4B (claim #4967). Within about fourteen months the company reported more than $30B of debt and equity secured year-to-date (claim #8735), after earlier tallies of $25B over 18 months (claim #26689) and $28B over 12 months (claim #11221). The IPO is therefore a small share of the capital stack. Direct equity dilution has come mostly from strategic placements: NVIDIA's $2B (claim #17594), Jane Street's $1B (claim #10305), and a $350M strategic customer issuance (claim #4514). The larger potential dilution sits in two convertibles, the $2.25B December 2025 notes (claim #18733) and the proposed $3.0B notes due 2032 (claim #10325). The first was partly hedged with a $295.7M capped call (claim #18738).

The mix looks layered rather than improvised. Asset-backed delayed-draw facilities (DDTL 3.0 at $2.6B, claim #26685; DDTL 4.0 at $8.5B with A3/A(low) ratings, claims #11218 and #11219; DDTL 5.0 at $3.1B, claim #9116) fund contracted GPU buildouts. High-yield unsecured notes and converts sit at the parent level. Pricing has drifted wider, however. Unsecured coupons moved from 9.000% (claim #26755) to 9.750% (claim #23816), and DDTL 5.5 is priced at SOFR + 5.50% (claim #8733) against SOFR + 4% on DDTL 3.0 (claim #26686). The January 2026 postponement of debt-service-coverage testing (claim #3912), and a reported contract-realization testing postponement that is not yet verifiable (claim #3913), sit awkwardly beside that drift and require explanation.

The filings state use of proceeds but show little about where the money actually went. The one traceable item is the $450M revolver repayment from the 2031 notes (claim #26759), which is refinancing rather than buildout. The December convertible's net proceeds appear as both $2.21B (claim #18737) and $2.54B (claim #3981), most likely before and after the option exercise, and should be reconciled.

The record contains no insider transactions, so no reading of insider behavior is possible. The final claim in the supplied data is truncated.

Watchers should look next for three things. The first is whether the postponed covenant tests resume on schedule. The second is whether draw deadlines on the DDTL 5.0 facility (claim #1321) are met with contracted capex. The third is whether a third convertible follows, which would make equity-linked funding a recurring tool rather than an occasional one.

CoreWeave's record describes a company that has funded itself overwhelmingly through debt since listing. Common-equity dilution has been modest by comparison. After the IPO, which netted $1.4B per claim #4967, the only new primary equity on record is a $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A purchase (claim #17594) and Jane Street's $1B investment (claim #10305). Potential dilution sits instead in the convertible layer: the December 2025 notes at 1.75% (claims #18733, #18736) and the proposed $3.0B 2032 converts (claim #10325). The first issue was paired with capped calls costing $295.7M of proceeds (claim #18738), which suggests management is deliberately deferring and limiting share-count growth. The mix looks more like a strategy than ad hoc patching. Asset-level delayed-draw facilities have grown from $2.6B at SOFR+4% (claims #26685, #26686) to an $8.5B investment-grade-rated DDTL 4.0 (claims #11218, #11219). Unsecured notes and converts sit at the corporate level. Pricing on that unsecured layer has not improved. Coupons have moved from 9.25% (claim #25585) to 9.0% (claim #26717) to 9.75% (claim #23816), even as secured contract-backed debt achieved investment-grade ratings. On the record, the corporate credit is not refinancing more cheaply as the company scales. Stated use of proceeds tracks deployment only partially. The July 2025 notes were earmarked in part to repay $450M of revolver borrowings (claim #26759), so new debt was partly retiring old debt. Contract-specific capex is the stated purpose of the DDTL 5.0 (claim #1320), but the claims include no deployment data to confirm drawdowns against it. Two items require explanation. The first is the postponement of debt service coverage and contract realization ratio testing, along with an amended minimum liquidity covenant (claims #3912, #3913, #3911). The second is a $270M termination fee payable to CoreWeave tied to a convertible-to-equity arrangement (claim #4739), whose resolution is not discussed further. A reader should watch whether the postponed covenant tests are reinstated on schedule and whether unsecured coupons start to converge with secured pricing. Reported cumulative raises have climbed from over $25B in 18 months (claim #26689) to more than $30B year-to-date by August 2026 (claim #8735). Each time that figure is restated, it is worth checking against contracted revenue such as the $21B Meta commitment (claim #2124). Several facility sizes on record also conflict, such as the December convertible at $2.25B versus $2.5875B (claims #18733, #3980). No insider transactions appear in the record, and the claim set ends mid-entry in August 2026.

CoreWeave's record describes a company financing itself mainly through debt, with equity used sparingly and often strategically. After the IPO, which raised $1.5B gross and $1.4B net per claims #19821 and #4967, nearly every subsequent event is secured or unsecured borrowing. The scale and pace both increase over time. Company-stated cumulative totals rise from $17.2B (claim #4968) to over $25B across 18 months (claim #26689), and then to more than $30B in 2026 year-to-date alone (claim #8735). Direct equity dilution has been modest by comparison. The $350M strategic customer issuance (claim #4514), NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305) are placements with named counterparties rather than market offerings. Latent dilution sits in two convertibles: the December 2025 notes, with roughly 30.0M underlying shares per claim #3984, and the proposed $3.0B 2032 notes (claim #10325). Capped calls partly offset this exposure.

The mix looks more like a structured program than ad hoc patching. Successive numbered delayed-draw facilities (DDTL 3.0 at $2.6B per claim #26685, DDTL 4.0 at $8.5B per claim #11218, DDTL 5.0 at $3.1B per claim #9116) finance specific contracts, and claim #2138 states their proceeds are tied to customer capex and GPU servers. Pricing has not clearly improved, however. Unsecured coupons moved from 9.25% (claim #25585) to 9.0% (claim #26717) and then to 9.75% (claim #23816). A facility priced at SOFR+5.50% (claim #8733) compares with SOFR+4% on DDTL 3.0 (claim #26686). Both data points sit alongside the A3 rating on DDTL 4.0 (claim #11219).

On deployment, the record states use of proceeds but rarely confirms it. The one traceable item is the $450M revolver repayment (claim #26759). Capped-call cost is given as $295.7M (claim #18738) and also as $340M (claim #3986). Convertible size appears as $2.0B, $2.25B and $2.5875B across claims #18739, #18733 and #3980. These figures are probably explained by upsizing and the option exercise, but they require reconciliation. The April 2026 convertible was described as $2.6B (claim #11765) before it appeared as $3.0B (claim #2126).

The most consequential item for a careful reader is the postponement of debt service coverage and contract realization ratio testing, along with a minimum liquidity amendment (claims #3912, #3913 and #3911). Watch whether testing resumes on schedule, whether coupons on future unsecured issues ease or keep rising, and how counterparty concentration (claims #4297 and #2124) is reflected in borrowing-base terms. The record contains no insider transactions, and the claim list is truncated after claim #8731, so the most recent events cannot be assessed.

CoreWeave's record describes a company that has funded itself overwhelmingly through debt, with equity arriving in a few large blocks rather than a steady drip. After the IPO, sized at $1.5B gross (claim #19821) and $1.4B net (claim #4967), the next equity came from two strategic placements: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #2080). The more meaningful source of potential dilution is the convertible program. The December 2025 convertible was first described at $2B (claim #18739) and settled at $2.5875B (claim #3980). A proposed $3.0B convertible due 2032 followed in April 2026 (claim #10325), with a $450M option attached (claim #10326). The capped calls blunt some of that dilution, but their stated size moves from $295.7M (claim #18738) to $340M (claim #3986) across disclosures, and that difference requires explanation.

The mix looks structured rather than improvised. Delayed-draw term loans are tied to customer contracts and GPU servers (claim #2138), unsecured notes repay or backstop the revolver (claim #26759), and converts sit on top. Two observations complicate that picture. First, unsecured coupons have drifted upward, from 9.25% on the 2030 notes (claim #4824) to 9.000% on the first 2031 tranche (claim #26755) and 9.750% on a later 2031 issue (claim #23816). Second, the DDTL 4.0 facility carried A3/A(low) ratings (claim #11219), while DDTL 5.5 is rated Ba2/BB+ (claim #8732). Several offerings were also upsized between announcement and pricing, which points to strong demand rather than distress.

Use-of-proceeds language is consistent across filings, but the record contains no deployment data, so whether actual spending tracked those statements cannot be assessed here. The covenant amendments matter more. Debt-service-coverage testing was postponed (claim #3912), contract-realization-ratio testing was postponed (claim #3913), and the minimum liquidity covenant was amended (claim #3911). None of these has been revisited in later claims on record.

No insider transactions appear in the data, and the record ends mid-entry in August 2026. A careful reader should watch for when the postponed covenant tests resume, how the 2031 and 2032 maturity stacks are refinanced, whether the projected $8.8B of additional secured debt (claim #26760) is actually drawn, and how much conversion overhang becomes live dilution.

CoreWeave's record describes a company that went public mainly to gain access to the debt markets, not to raise equity. The IPO raised $1.5B at $40 (claims #19821, #11616), about $1.4B net (claim #4967). Debt followed within weeks. Direct equity dilution after the listing has come in a few large strategic blocks: the $350M strategic customer issuance (claim #4514), NVIDIA's $2B (claim #17594), and Jane Street's $1B (claim #10305). The larger dilution question sits in two convertibles. The first is the December 2025 issue of up to $2.5875B (claim #3980). The second is the $3.0B due-2032 proposal, with a $450M option (claims #2126, #10326). Capped calls blunt part of their effect, but neither has converted yet. The pace of raising has accelerated. The company cited $25B over 18 months (claim #26689), more than $20B year-to-date by May 2026 (claim #9123), and more than $30B by August (claim #8735).

The mix looks deliberate rather than patched together. Delayed-draw term loans tied to specific customer contracts finance GPU capex (claim #2138). Unsecured high-yield notes, converts, and strategic equity fund the parent. The DDTL 4.0 facility's A3/A(low) rating (claim #11219) suggests the contract-backed structure is being accepted. Pricing has not improved in step. Unsecured coupons moved from 9.250% (claim #4824) to 9.000% (claim #26755) and then to 9.750% (claim #23816). DDTL spreads widened from SOFR+4% (claim #26686) to SOFR+5.50% (claim #8733).

For stated use of proceeds, the record supports only partial verification. The 2031 notes were to repay $450M of revolver borrowings (claim #26759). For the December convertible, the capped-call allocation is given as both $340M (claim #3986) and $295.7M (claim #18738), and the gap needs explanation. The record contains no evidence of actual deployment against these stated uses. It also contains no insider transactions, so no read on insider selling is possible.

Readers should watch the covenant terms. Debt service coverage testing was postponed and the minimum liquidity covenant was amended (claims #3912, #3911), while new facilities carry the same coverage ratio tests (claims #2141, #1324). A further postponement would matter more than any single new issuance. Readers should also confirm that the deals marked not yet verifiable have closed, and track the projected $8.8B of additional secured debt (claim #26760) against the anchor customer commitments it depends on, such as Meta's $21B (claim #2124) and OpenAI's $6.5B (claim #4297).

CoreWeave's record is overwhelmingly a debt story with equity used selectively. After the $1.5B IPO at $40 (claims #19821, #11616), net proceeds of roughly $1.4B (claim #4967) were quickly dwarfed by borrowing. The company cites over $25B raised in 18 months (claim #26689), $28B in commitments over twelve months (claim #11221), and more than $20B year-to-date by May 2026 (claim #9123). Direct common-stock dilution came in discrete blocks: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). The larger dilution question sits in the convertibles: roughly $2.59B gross in December 2025 (claim #3980) and a proposed $3.0B due 2032 (claim #10325). The December issue was partly offset by capped calls of about $340M (claim #3986). The record contains no insider-sale entries, so that dimension cannot be assessed here.

The mix reads as a deliberate two-track structure rather than patching. Asset-level delayed-draw facilities fund contracted GPU capacity, with DDTL 4.0 at $8.5B carrying A3/A(low) ratings (claims #11218, #11219) and DDTL 5.0 at $3.1B (claim #1319). Unsecured notes and convertibles fund the corporate layer. Two features complicate that reading. Unsecured coupons have stayed high and drifted upward, from 9.25% (claim #4824) to 9.0% (claim #26755) to 9.75% (claim #23816). And in January 2026, debt-service and contract-realization ratio testing was postponed (claims #3912, #3913) alongside a minimum-liquidity amendment (claim #3911). The postponement requires explanation relative to the pace of new issuance.

The stated uses of proceeds are specific where disclosed: $450M to repay the revolver (claim #26759), $295.7M for capped calls (claim #18738), and GPU servers tied to customer contracts (claim #2138). The record contains no data on actual deployment, so whether proceeds tracked those uses cannot be verified from these claims.

A sophisticated reader should watch four things. The first is the gap between underlying customer contracts averaging about three years (claim #8731) and the DDTL 5.5 maturity of about five years (claim #8730), which carries Ba2/BB+ ratings (claim #8732) rather than investment grade. The second is whether the covenant tests resume on schedule. The third is the pricing of the next unsecured issue against the 9%-plus trend. The fourth is the share-count effect if the 2032 convertible closes with its $450M option exercised (claim #10326). Several claims repeat the same events in inconsistent units, and the record is cut off in August 2026, so later events are not reflected.

CoreWeave's capital history is mostly debt with relatively little equity, and the sequence is fast. The IPO priced at $40 (claim #11616) and raised $1.5B gross (claim #19821), or roughly $1.4B net (claim #4967). That is small next to what followed. The company's own running totals rise from over $25B raised in 18 months (claim #26689) to more than $30B secured in 2026 year-to-date alone (claim #8735). Direct equity dilution has come mainly through strategic placements, not public follow-ons: the NVIDIA $2B investment (claim #17594) and the Jane Street $1B investment (claim #10305). The larger potential dilution sits in two convertible issues. The first was a $2.25B offering in December 2025 (claim #18733), with $295.7M of proceeds allocated to capped calls (claim #18738). The second was a proposed $3.0B convertible due 2032 (claim #10325). The record shows the December size inconsistently, at $2.25B, $2.0B and $2.5875B (claims #18733, #18739, #3980), and this should be reconciled.

The mix looks like a deliberate layered structure, not opportunistic patching. Delayed-draw facilities secured against customer contracts sit at the asset level, culminating in the $8.5B DDTL 4.0 rated A3/A(low) (claims #11218, #11219). Unsecured notes and convertibles sit at the corporate level. Two features of that structure need explaining. First, unsecured coupons have not come down as scale increased. They ran from 9.25% on the 2030 notes (claim #4824) to 9.000% on the first 2031 notes (claim #26755) to 9.750% on the April 2026 issue (claim #23816). Second, the January 2026 amendments postponed debt service coverage ratio testing (claim #3912). Related changes to the minimum liquidity covenant and the contract realization ratio are not yet verifiable (claims #3911, #3913).

The data does not allow a clean test of whether use of proceeds tracks actual deployment. Most use-of-proceeds claims are general. The one specific item is the $450M revolver repayment from the 2031 notes (claim #26759), and no claim on record confirms that it was carried out.

The record also contains no insider transactions, so no view on insider selling is possible. It ends mid-entry in August 2026, so the most recent activity is incomplete.

A careful reader should watch four things. The first is when deferred covenant testing resumes. The second is the gap between the roughly five-year DDTL 5.5 maturity (claim #8730) and underlying customer contracts that average about three years (claim #8731). The third is whether conversion pressure builds on the 2025 and 2032 convertibles. The fourth is whether the company's next unsecured pricing comes in above or below the 9.750% mark.

CoreWeave's record describes a company financing itself mainly through debt, with equity in a secondary role. Ordinary-share dilution after the IPO (net proceeds of $1.4B per claim #4967) comes from strategic placements rather than broad follow-ons: a $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). The heavier potential dilution sits in the convertibles: the December 2025 deal, reported at $2.0B (claim #18739), $2.25B (claim #18733) and $2.5875B including the option (claim #3980), followed by a proposed $3.0B 2032 issue with a $450M option (claims #10325, #10326). The roughly $2.6B figure in claims #11765 and #10327 does not cleanly reconcile with either and requires explanation. Capped calls are described as both $295.7M (claim #18738) and $340M (claim #3986), which likely reflects pre- and post-option sizing.

The mix is coherent rather than patched together. It forms a repeated ladder of delayed-draw term loans: $2.6B (claim #26685), $3.0B (claim #4277), $8.5B (claim #11218), $3.1B (claim #9116) and DDTL 5.5 (claim #8730). Unsecured notes, an expanded revolver (claim #11766) and convertibles sit alongside these loans. The cost of capital has not visibly declined. Coupons run from 9.25% (claim #4824) to 9.0% (claim #26755) to 9.75% (claim #23816). Ratings have also fallen, from A3/A(low) on DDTL 4.0 (claim #11219) to Ba2/BB+ on DDTL 5.5 (claim #8732).

On use of proceeds, the record shows stated intentions, such as repaying the revolver with $450M of 2031 note proceeds (claim #26759) and funding contract capex and GPU servers (claim #2138), but it does not show actual deployment. Tracking therefore cannot be assessed. The repayment of revolver draws with term notes is itself a form of liability recycling worth noting.

A sophisticated reader should watch the covenant postponements on debt service coverage and contract realization (claims #3912, #3913) and the minimum liquidity amendment (claim #3911) to see whether they recur. The DDTL 5.5 facility's roughly five-year tenor against underlying contracts averaging about three years (claims #8730, #8731) also requires explanation of refinancing or renewal assumptions. Cumulative raise figures rose quickly, from $17.2B (claim #4968) to more than $30B (claim #8735). The record contains no insider transactions, so no conclusion can be drawn on that dimension.

The disclosed record is overwhelmingly a debt story, and direct equity dilution has been modest and episodic. Common equity came through the IPO, about $1.5B gross (claim 19821) and roughly $1.4B net (claim 4967), followed by a $350M strategic customer issuance (claim 4514), NVIDIA's $2B Class A investment (claim 3891) and Jane Street's $1B (claim 10305). The more consequential dilution channel is the convertibles. These are the December 2025 notes, upsized to about $2.59B (claim 3980) and carrying conversion terms tied to roughly 30.0M shares (claim 3984), and the proposed $3.0B 2032 tranche (claim 10325). The record shows no closing for that tranche, and its relationship to the $2.6B figure in claims 11765 and 10327 requires explanation. No insider transactions appear in the record, so that dimension cannot be assessed. The mix reads as a coherent strategy rather than patching. Delayed-draw term loans fund specific contract capex: DDTL 3.0 at $2.6B (claim 26685), DDTL 4.0 at $8.5B with A3/A(low) ratings (claims 11218, 11219), and DDTL 5.0 at $3.1B (claim 9116). Unsecured notes and convertibles fill corporate needs around them. Several deals were upsized, including the 2030 notes from $1.5B to $2B (claims 11921, 11917) and the 2031 notes from $1.5B to $1.75B (claims 26765, 26755), which indicates demand rather than strain. However, unsecured coupons have not compressed. They moved from 9.25% (claim 4824) to 9.0% (claim 26755) and then 9.75% (claim 23816). The investment-grade pricing has been confined to the asset-backed layer. Stated use of proceeds is partly traceable. The 2031 notes earmarked $450M for revolver repayment (claim 26759), and the convertible allocated $295.7M to capped calls (claim 18738), rising to $340M after upsizing (claim 3986). Deployment of DDTL draws against the OpenAI and Meta commitments (claims 4297, 2124) is not documented in the record. Cumulative raise figures escalate quickly: $17.2B (claim 4968), over $25B in 18 months (claim 26689), $28B in twelve months (claim 11221) and over $20B year-to-date in 2026 (claim 9123). A careful reader should watch the January 2026 covenant amendments, which postponed debt-service-coverage and contract-realization testing and revised minimum liquidity (claims 3911, 3912, 3913). They should also watch whether the 2032 convertible closed, and whether unsecured coupons follow the secured layer down. The record ends mid-entry on the DDTL 5.5 facility (claim 8733), so the most recent terms remain incomplete.

CoreWeave's record is overwhelmingly a debt story, and the equity side is comparatively modest. Following the IPO (the $1.5B offering per claim #19821, $1.4B net per claim #4967), new equity has come mainly through strategic placements rather than marketed follow-ons: NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305). Dilution so far has therefore been episodic and counterparty-driven. The larger latent source is the convertible layer. The December 2025 notes grew from $2.25B (claim #18733) to $2.5875B (claim #3980), and a $3.0B 2032 convertible with a $450M option was proposed in April 2026 (claim #10325). Other entries cite $2.6B for that program (claims #10327, #11765), and that discrepancy requires explanation.

The debt mix looks like a deliberate architecture rather than patching. Contract-backed delayed-draw facilities (DDTL 3.0 through 5.0) are paired with recurring unsecured notes. The cost trajectory is less settled. Unsecured coupons have moved from 9.25% (claim #4824) to 9.000% (claim #26755) and then to 9.750% (claim #23816). DDTL spreads widened from SOFR+4% (claim #26686) to SOFR+5.50% (claim #8733), even though DDTL 4.0 carried A3/A(low) ratings (claim #11219). The January 2026 postponement of debt-service-coverage testing (claim #3912), alongside a minimum-liquidity amendment that is not yet verifiable (claim #3911), sits awkwardly beside the scale of new issuance and is not explained in the record.

Use-of-proceeds disclosure is mostly generic. One specific allocation, repaying $450M of revolver borrowings from the 2031 notes (claim #26759), has no matching deployment claim. Capped-call funding was stated as $295.7M (claim #18738) and later as $340M (claim #3986), consistent with the upsize. Beyond that, the record does not let a reader trace actual deployment against stated intent. Insider transaction data is absent from the claims provided, so no read on insider selling is possible here.

A sophisticated reader should watch several items. The first is whether the deferred coverage tests resume on schedule. The second is whether coupons and spreads continue to drift upward. The third is the gap between underlying customer contracts averaging roughly three years (claim #8731) and loan tenors near 5.5 years (claim #9120). The last is the cumulative convertible overhang as share prices approach conversion terms. Aggregate-raise figures ($25B over 18 months per claim #26689, $28B over 12 months per claim #11221) overlap and should not be summed. Some claims also appear in mixed units, and the record provided is truncated at claim #8729.

CoreWeave's record since its March 2025 listing is overwhelmingly a debt story. Equity dilution has been modest and episodic: the $1.5B IPO (claim 19821, net $1.4B per claim 4967) was followed by strategic placements rather than marketed follow-ons, namely NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 10305). The convertibles carry the main latent dilution. The December 2025 notes were sized variously at $2B, $2.25B and $2.5875B across claims 18739, 18733 and 3980, which is most plausibly the upsize plus option. They were followed in April 2026 by a proposed $3.0B convertible due 2032 (claim 10325). Capped calls of $295.7M to $340M (claims 18738, 3986) suggest the company is deliberately managing that overhang.

The mix reads as a coherent architecture rather than patching. Asset-level delayed-draw facilities are tied to customer contracts (claim 2138), and they progressed from DDTL 3.0 at SOFR+4% (claim 26686) to an investment-grade-rated $8.5B DDTL 4.0 (claims 11218, 11219) and a publicly syndicated $3.1B DDTL 5.0 (claim 9117). Unsecured high-yield notes and converts sit above that layer. Two features complicate the picture. Coupons have not compressed: 9.25% (claim 25585), then 9.0% (claim 26755), then 9.75% (claim 23816). The August 2026 facility priced at SOFR+5.50% with Ba2/BB+ ratings (claims 8733, 8732), wider than DDTL 3.0. Separately, the January 2026 postponement of debt-service-coverage and contract-realization ratio testing, and the amendment to the minimum liquidity covenant (claims 3912, 3913, 3911), require explanation alongside the pace of new issuance.

Use of proceeds is only partially traceable. The 2031 notes were earmarked partly to repay $450M of revolver borrowings (claim 26759), and convertible proceeds funded capped calls (claim 18743). Deployment of the DDTL draws against stated contract capex cannot be verified from this record. Several April 2026 figures, such as $1.25B versus $2.75B for notes due 2031 (claims 2125, 23816), are not reconciled here. No insider transactions appear in the data provided, so no read on insider behavior is possible.

A sophisticated reader should watch four things. The first is whether the deferred covenant tests resume on schedule. The second is whether investment-grade asset-level pricing extends beyond DDTL 4.0. The third is whether unsecured coupons stabilize or keep widening. The fourth is how the cumulative capital figures, $25B in 18 months (claim 26689) and more than $20B year-to-date (claim 9123), compare with recognized revenue from the large customer commitments (claims 4297, 2124).

CoreWeave's capital history since its March 2025 listing reads less as a sequence of discrete raises than as a continuous financing program, and its trajectory has steepened. The company cited more than $20B raised year-to-date by May 2026 (claim #9123) and more than $30B by August (claim #8735), after describing $28B in commitments over the prior twelve months (claim #11221). That figure dwarfs the IPO's $1.4B in net proceeds (claim #4967). Direct equity dilution has come mainly through strategic placements rather than open-market issuance: NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305). Latent dilution sits in two convertible tranches, the December 2025 notes (claim #18733) and the $3.0B notes due 2032 (claim #10325). The December convertible is recorded at $2.0B, $2.25B and roughly $2.59B across claims #18739, #18733 and #3980, and its capped call at $295.7M and $340M (claims #18738, #3986). Upsizing and option exercise likely explain this, but it requires reconciliation.

The mix shows a coherent architecture. Asset-backed delayed-draw facilities fund contracted GPU capacity (claim #2138), unsecured high-yield notes and convertibles supply corporate liquidity, and strategic equity anchors the balance sheet. Within that structure, the terms have moved unevenly. Senior note coupons rose from 9% (claim #4697) to 9.25% (claim #11917) to 9.75% (claim #23816). After the investment-grade A3 rating on DDTL 4.0 (claim #11219), DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732). In January 2026 the company postponed debt service coverage and contract realization testing and amended a minimum liquidity covenant (claims #3912, #3913, #3911). These are revised terms that merit continued attention.

The stated use of proceeds for the term loans is customer-contract capex. The record contains no deployment data that would let a reader confirm whether proceeds track that purpose, so that question remains open. No insider transactions appear in the record provided, and the dataset itself is truncated.

A careful reader should watch four things next. The first is whether Q2 interest expense of $640M (claim #7374) on $2.58B of revenue (claim #8183) compresses as facilities draw down. The second is how roughly three-year customer contracts (claim #8731) are refinanced against facilities with roughly five-year maturities (claim #8730). The third is when the postponed covenant tests resume. The fourth is whether future tranches price closer to DDTL 4.0 or DDTL 5.5.

CoreWeave's record since the March 2025 listing reads less as a dilution story than as a leverage story. The IPO raised $1.5B at $40 (claims #19821, #17815), with net proceeds of $1.4B (claim #4967). Nearly everything after that came as debt. The company itself frames the pace: $28B of commitments in twelve months (claim #11221), more than $20B year-to-date by May 2026 (claim #9123), and more than $30B by August (claim #8735).

The mix looks coherent. Asset-backed delayed-draw facilities are tied to customer contracts and stated to fund GPU servers and contract capex (claims #2138, #1320). Unsecured high-yield notes and converts sit alongside them, and strategic equity from NVIDIA ($2B, claim #3891) and Jane Street ($1B, claim #10305) added the only meaningful new common since the IPO.

Two threads complicate that reading. First, the cost of capital has not clearly fallen. DDTL 4.0 carried A3/A(low) ratings (claim #11219), but the later DDTL 5.5 is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). Unsecured coupons went from 9.25% (claim #11917) to 9.750% on the 2031 notes (claim #23816). Second, in January 2026 the company postponed debt service coverage and contract realization ratio testing and amended its minimum liquidity covenant (claims #3912, #3913, #3911). Larger raises followed within weeks.

Convertible issuance is where equity dilution will build up. The December 2025 notes carried roughly 30.0M underlying shares (claim #3984), and a $3.0B 2032 convert followed with a $450M option (claims #2126, #2127). The capped-call figure for the December deal is $295.7M in one disclosure (claim #18738) and $340M in another (claim #3986). That gap requires explanation.

The record does not let us check whether stated use of proceeds tracks actual deployment, because it contains no capex or drawdown figures. No insider transactions appear in the visible portion of the record, which is truncated.

Three things are worth watching next. One is when the postponed covenant tests resume. Another is the mismatch between DDTL 5.5's roughly five-year maturity and underlying contracts averaging about three years (claims #8730, #8731). The last is whether the $104B backlog (claim #1493), including the Meta and OpenAI commitments (claims #2124, #4297), turns into revenue fast enough to cover quarterly interest expense, which already stands at $640M (claim #7341).

CoreWeave's record describes a company that has financed itself almost continuously since listing. Direct equity dilution has been episodic rather than steady. The IPO raised $1.5B at $40 (claims #19821, #11616). That was followed by strategic placements: NVIDIA's $2B (claim #17594) and Jane Street's $1B (claim #10305). Two convertible issues add latent dilution. The December 2025 deal grew from $2B to $2.25B to $2.5875B as it was marketed (claims #18739, #18733, #3980). A proposed $3.0B due 2032 followed, with a $450M upsize option (claims #10325, #10326). The capped calls funded from the first deal, $295.7M of net proceeds per claim #18738, suggest management is attentive to conversion dilution, though they do not eliminate it.

The mix reads less as patching than as a deliberate two-track structure. Asset-level delayed-draw facilities (DDTL 3.0 through 5.5) fund contracted GPU deployments, while corporate unsecured notes and converts fund the parent. Within that structure, terms have moved. DDTL 4.0 carried A3/A(low) ratings (claim #11219). The later $2.6B facility is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). Senior note coupons have run between 9% and 9.75% (claims #4697, #4824, #23816). Several note offerings were upsized before pricing, for example $1.5B to $2B (claims #11921, #11917). That indicates demand, but at high cost.

On use of proceeds, the stated purposes are consistent: customer-contract capex and GPU servers (claim #2138), plus capped calls (claim #18743). The record contains no disclosure tracing actual deployment against those purposes, so the alignment cannot be confirmed from this data. The postponement of debt-service-coverage and contract-realization testing (claims #3912, #3913) requires explanation alongside any deployment narrative.

No insider transactions appear in the claims provided, and the record is truncated after the Q2 2026 results.

A careful reader should watch several things next. One is whether the postponed covenant tests resume on schedule. Another is the gap between facility maturities of roughly five years and underlying customer contracts averaging about three years (claims #8730, #8731), which raises renewal risk. A third is interest expense, reported at $640 for Q2 2026 (claim #7341; the record gives no unit), relative to operating income. The last is whether concentrated commitments, such as Meta's $21B (claim #2124) and OpenAI's $6.5B (claim #4297), continue to anchor new facilities at investment-grade terms or migrate further toward high-yield pricing.

CoreWeave's post-listing capital history is defined less by equity dilution than by a steady escalation of leverage. The $1.5B IPO at $40 per share (claims #19821, #11616), which netted roughly $1.4B (claim #4967), was quickly dwarfed by debt. Within about two months the company had expanded its revolver by $1.5B (claim #11923), issued $2B of 9.25% notes due 2030 (claims #11917, #4824), and arranged a $2.6B delayed-draw term loan (claim #11922). That cadence then accelerated: the company cites more than $30B of debt and equity secured year-to-date in 2026 (claim #8735), against Q2 2026 revenue of $2.58B (claim #8183) and a reported debt-to-equity ratio near 6 (claim #3058).

The mix looks coherent rather than opportunistic, built around asset-level delayed-draw facilities tied to customer contracts. The $8.5B DDTL 4.0 carried A3/A(low) ratings (claims #11218, #11219), which suggests contract-backed structures can borrow on better terms than the parent. Corporate unsecured paper tells a different story, with coupons moving from 9% (claim #4697) to 9.25% to 9.75% on the 2031 notes (claim #23816). Equity has come mostly from strategic holders, namely NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305), and from convertibles sized at $2.59B (claim #3980) and $3.0B (claim #10325), with capped calls of $340M partly offsetting the first (claim #3986). Several claims require explanation. Covenant testing on debt service coverage and contract realization was postponed and a minimum liquidity covenant amended (claims #3911–#3913). The DDTL 5.5 has a roughly five-year maturity against underlying contracts averaging about three years (claims #8730, #8731), and it priced at SOFR + 5.50% with sub-investment-grade ratings (claims #8732, #8733).

Stated uses of proceeds point to customer-contract capex and GPU servers (claim #2138), but the record contains no disclosure of actual deployment, so whether spending tracks those statements cannot be assessed here. The record also contains no insider transactions, so that dimension cannot be evaluated.

A sophisticated reader should watch four things. The first is whether the postponed covenant tests are reinstated on schedule. The second is how the gap between financing tenor and contract duration is handled at renewal. The third is whether corporate coupons keep rising. The fourth is whether the reported $104B backlog (claim #1493) converts to revenue fast enough to narrow losses that are, by the company's own account, widening (claim #1495).

CoreWeave's record since the March 2025 listing is overwhelmingly a debt story. Direct equity issuance is a small share of the total: the $1.5B IPO (claim #19821), a $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A purchase (claim #3891), and Jane Street's $1B investment (claim #10305). Dilution so far has come in a few large placements rather than a steady drip, and the record shows no ATM program. The larger latent dilution sits in the convertibles: the December 2025 issue, sized at $2.25B (claim #18733) or $2.59B with the option (claim #3980), references roughly 30.0M underlying shares (claim #3984), partly offset by capped calls, while a further $3.0B due 2032 was proposed in April 2026 (claim #10325).

The mix looks structured rather than improvised. Asset-backed delayed-draw term loans fund GPU capex against customer contracts (claim #2138), and the progression from DDTL 3.0 to the $8.5B investment-grade DDTL 4.0 (claims #11218, #11219) suggests an attempt to lower the cost of that layer. Two developments complicate the picture. The January 2026 amendments postponed debt service coverage and contract realization ratio testing and amended the minimum liquidity covenant (claims #3911–3913). The later DDTL 5.5 also came at Ba2/BB+ and SOFR+5.50% (claims #8732, #8733), well below DDTL 4.0's rating. Unsecured coupons drifted from 9% (claim #4697) and 9.25% (claim #4824) to 9.75% (claim #23816), which suggests rising rather than falling marginal funding costs.

Use of proceeds is stated in general terms, and the record contains no deployment data to check it against. The December convertible's net proceeds and capped-call figures differ across filings ($2.21B and $295.7M per claims #18737 and #18738, versus $2.54B and $340M per claims #3981 and #3986), presumably reflecting option exercise, but the record does not reconcile them. No insider transactions appear in the record provided.

A careful reader should watch five things. First, whether the postponed covenant tests resume on schedule. Second, how Q2 2026 interest expense of $640M (claim #7341) compares with operating income alongside a negative 25% net margin (claim #10221). Third, the gap between DDTL 5.5's roughly five-year maturity and underlying contracts averaging about three years (claims #8730, #8731). Fourth, conversion activity on the 2032 notes. Fifth, whether issuance continues at the stated pace of more than $30B year-to-date (claim #8735).

CoreWeave's record describes a company that used its public listing as a gateway to debt markets far more than as an equity-raising event. The IPO priced at $40 (claim #11616) and raised $1.5B gross (claim #19821), about $1.4B net (claim #4967). It was quickly dwarfed by a sequence of high-coupon unsecured notes. These include the 9.250% notes due 2030 at $2B (claim #11917), the 9% notes upsized to $1.75B (claims #4697, #4698), the 9.750% notes due 2031 at $2.75B (claim #23816), and $3.5B due 2032 (claim #8927). By August 2026 the company described more than $30B of debt and equity secured year-to-date (claim #8735).

Direct equity dilution has been comparatively measured. Beyond the IPO, the record shows a strategic customer issuance of $350M (claim #4514), NVIDIA's $2B Class A investment (claim #17594), and Jane Street's $1B (claim #10305). The heavier latent dilution sits in two convertible programs: $2.25B in December 2025 (claim #18733), partly offset by capped calls costing $295.7M (claim #18738), and a proposed $3.0B due 2032 (claim #10325). The December figures conflict across claims (#3980, #3981, #18737), likely reflecting pre- and post-option sizing; this requires reconciliation against the final indenture.

The mix reads as deliberate rather than opportunistic. Asset-level delayed-draw facilities are tied explicitly to customer-contract capex and GPU servers (claims #2138, #1320). Corporate notes and converts fund the balance sheet more broadly. The DDTL 4.0 achieved A3 ratings (claim #11219). The later DDTL 5.5, however, carries Ba2/BB+ ratings at SOFR + 5.50% (claims #8732, #8733), so not every contract pool commands the same credit quality.

Whether deployment tracks stated use cannot be verified from this record. The capex and revenue figures (claims #18918, #7340) lack consistent units, and no facility-level draw data is disclosed. The January 2026 entries require explanation: a minimum liquidity covenant amendment and postponed debt service coverage and contract realization testing (claims #3911, #3912, #3913). They indicate that original covenant schedules were revised within a year of issuance.

The record contains no insider transactions, and the Q2 2026 data is truncated. Going forward, a careful reader should watch four things. The first is quarterly interest expense, reported at $640 (claim #7341), relative to revenue. The second is whether the postponed covenant tests resume on their revised dates. The third is the final sizing of the 2032 convertible. The fourth is whether backlog of roughly $104B (claim #18917) converts to contracted cash flow on the roughly three-year average contract terms underlying recent facilities (claim #8731).

The record describes a company whose funding rests mainly on secured and unsecured debt, with equity added at intervals rather than the other way around. After the March 2025 IPO, which raised $1.5B gross per claim #19821 and roughly $1.4B net per claim #4967, CoreWeave issued senior notes almost every quarter. These include the 9.250% notes due 2030 (claim #4824), the 9% notes upsized to $1.75B (claims #4697, #4698), and further 2031 and 2032 tranches in 2026 (claims #23816, #8927). Alongside the notes ran a sequence of delayed-draw term loans, from DDTL 3.0 through DDTL 5.5. Direct equity dilution has come mostly from strategic placements, namely NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305). Convertibles are the main source of potential future dilution. The December 2025 convertible totaled $2.5875B including the option (claim #3980), and the April 2026 $3.0B convertible due 2032 carries a $450M option (claim #10326). Both appear to have been paired with capped calls, which moderate effective dilution at a cost of $295.7M to $340M on the first deal (claims #18738, #3986). These two capped-call figures differ, as do the expected and reported net proceeds ($2.2106B versus $2.5422B, claims #18737, #3981), and the difference is not reconciled in the record.

The mix reads as deliberate, since the company itself describes asset-backed DDTLs tied to customer contracts as its infrastructure financing approach (claims #2138, #3557). The financing terms, however, are uneven. DDTL 4.0 earned A3/A(low) investment-grade ratings (claim #11219), while DDTL 5.5 was rated Ba2/BB+ and priced at SOFR plus 5.50% (claims #8732, #8733). That later facility has a maturity of roughly five years against underlying contracts averaging about three (claims #8730, #8731), a duration gap that requires explanation. In January 2026, testing of the debt service coverage and contract realization ratios was postponed and the minimum liquidity covenant was amended (claims #3911 to #3913). The record does not explain why.

The stated use of proceeds, customer-contract capex and GPU servers, is broadly consistent with the reported Q2 2026 capex (claim #18918) and a backlog of about $104B (claim #17802). The record does not let a reader trace deployment facility by facility.

Several things are worth watching next: whether the postponed covenant tests resume on schedule, the trajectory of quarterly interest expense ($640M in Q2 2026 per claim #7341), the share of the backlog that sits with OpenAI and Meta (claims #4297, #2124), and the conversion economics of roughly $5.5B or more in convertibles. The record contains no insider transactions, and the Q2 2026 entries are truncated, so this reading of those quarters is incomplete.

CoreWeave's capital history since its March 2025 listing is dominated by debt rather than equity. The IPO raised roughly $1.5 billion gross (claim #19821) and $1.4 billion net (claim #4967). After that, common-equity dilution came mainly through strategic placements: NVIDIA's $2 billion (claim #3891) and Jane Street's $1 billion (claim #10305). Cumulative raise figures climbed quickly, from $28 billion over twelve months (claim #11221) to more than $30 billion year-to-date by August 2026 (claim #8735). Most of that is leverage, and the stated debt-to-equity ratio of about 6 (claim #3058) reflects it. The record contains no insider transactions, so no assessment of insider selling is possible here.

The instrument mix has an internal logic. Asset-level delayed-draw facilities are tied to specific customer contracts and GPU purchases (claims #2138, #1320). Corporate unsecured notes and convertibles sit above them. Several offerings were upsized, including $1.5 billion to $2 billion (claims #11921, #11918) and $1.5 billion to $1.75 billion (claims #4700, #4697), which indicates demand rather than strain. Pricing, however, has not improved. Unsecured coupons moved from 9.25% (claim #4824) to 9.75% on the 2031 notes (claim #23816). The most recent facility, DDTL 5.5, carries Ba2/BB+ ratings at SOFR plus 5.50% (claims #8732, #8733), compared with the A3 rating on DDTL 4.0 (claim #11219). That divergence requires explanation, likely in the counterparty quality of the underlying contracts.

Use of proceeds is broadly consistent where disclosed. Part of the December 2025 convertible proceeds went to capped calls, recorded as $295.7 million (claim #18738) and $340 million (claim #3986). Those two figures should be reconciled against the base and upsized deal sizes. The January 2026 amendment that postponed debt service coverage and contract realization ratio testing (claims #3912, #3913) is the clearest sign that deployment trailed the original schedule.

Going forward, three items bear watching. First, the underlying customer contracts average about three years (claim #8731) against a roughly five-year facility maturity (claim #8730), so renewal risk sits inside the debt tenor. Second, Q2 2026 interest expense was recorded at $640 million (claim #7341) against negative net margins (claim #7343). Third, the market will look for evidence that the backlog of roughly $100 billion or more (claims #647, #962) converts to cash fast enough to meet covenants once the postponed tests resume.

CoreWeave's record shows little conventional equity dilution and a large, fast-growing debt stack. After the IPO (claim 19821), new common equity came mainly from strategic placements: NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 10305) and a $350M customer issuance (claim 4514). Latent dilution sits in the convertibles. The December 2025 notes were upsized to roughly $2.59B (claim 3980) with capped calls costing $340M (claim 3986). A 2032 convertible followed, described as $3.0B in claim 10325 and $2.6B in claims 10327 and 11765, a discrepancy that requires explanation. The pace is notable: more than $20B raised year-to-date by May (claim 9123) and more than $30B by August (claim 8735). The mix reads as a deliberate two-tier design rather than patching. Asset-level delayed-draw facilities are tied to customer contract capex (claim 2138), and high-coupon unsecured notes sit at the corporate level, running from 9% (claim 4697) to 9.75% (claim 23816). Two developments complicate that picture. First, the minimum liquidity covenant was amended and coverage-ratio testing was postponed (claims 3911–3913). Second, the newest facility, DDTL 5.5, carries Ba2/BB+ ratings at SOFR+5.50% (claims 8732, 8733), well below the A3 rating attached to DDTL 4.0 (claim 11219). Stated uses track the instruments on paper: DDTL proceeds go to contract-linked GPU capex (claim 1320), and convertible proceeds partly fund capped calls (claim 18743). The record contains no capex or draw-down figures, so actual deployment cannot be checked against announcements. Servicing costs are already significant, with Q2 2026 interest expense of about $640M (claim 7341) and a net margin near −25% (claim 7343), against a backlog cited near $100B (claim 647). The record contains no insider transactions, so no read on insider selling is possible. A sophisticated reader should watch several things. One is whether the postponed covenant tests resume on schedule. Another is whether underlying contracts averaging about three years (claim 8731) can support a facility with a roughly five-year maturity (claim 8730). A third is whether rating quality keeps sliding from DDTL 4.0 to later tranches. The last is whether backlog converts to revenue fast enough to cover interest expense before the convertibles approach their conversion economics.

CoreWeave's capital history is overwhelmingly a debt story. Common equity has been a minority input, with the IPO at roughly $1.4B net (claim #4967), a $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). Dilution so far has been moderate, but the more consequential overhang sits in two convertible tranches: the $2.5875B December 2025 deal (claim #3980) and the proposed $3.0B 2032 notes (claim #10325), which carry a $450M upsize option (claim #10326). The record gives differing figures for the December capped call, $295.7M (claim #18738) and $340M (claim #3986), and differing sizes for the April convertible, $2.6B (claim #10327) and $3.0B (claim #10325). Both discrepancies require explanation before anyone models potential share count.

The mix reads as deliberate rather than patched together. Contract-backed delayed-draw facilities form the core: DDTL 3.0 at up to $2.6B (claim #11922), the $8.5B DDTL 4.0 rated A3 (claims #11218, #11219), DDTL 5.0 at $3.1B (claim #1319) and DDTL 5.5 at $2.6B (claim #8729). Unsecured senior notes layer on top of that base, beginning at 9.25% (claim #4824) and 9% (claim #4697) and extending to a $3.5B 2032 issue (claim #8927). The trajectory within that structure deserves attention. DDTL 5.5 priced at SOFR + 5.50% (claim #8733) and was rated Ba2/BB+ (claim #8732), well below DDTL 4.0's investment-grade rating. Its roughly 5-year maturity (claim #8730) also runs against underlying contracts averaging about 3 years (claim #8731). Separately, the minimum liquidity covenant was amended and the debt-service-coverage and contract-realization tests were postponed (claims #3911–#3913). Those are revisions, not routine maintenance.

On use of proceeds, the stated purposes are consistent: customer-contract capex and GPU servers for DDTL 4.0 (claim #2138), and capped calls funded from the convertible (claim #18743). The record does not contain deployment data that would let a reader confirm actual spending against these purposes. What it does show is the cost of the program, with Q2 2026 interest expense of $640M (claim #7341), widening operating and net losses (claim #1495), and a debt-to-equity ratio near 6 (claim #3058). Aggregate raise figures escalate from $17.2B (claim #4968) to more than $30B year-to-date (claim #8735).

A careful reader should watch four things next. The first is whether the postponed covenant tests resume on schedule. The second is whether new facilities return to investment grade or continue at high-yield spreads. The third is how contract tenor is matched to debt maturity. The fourth is customer concentration, since the OpenAI ($6.5B, claim #4297) and Meta ($21B, claim #2124) commitments underpin much of this collateral. The supplied record ends mid-entry in August 2026 and contains no insider transactions, so no conclusions about insider behavior can be drawn from it.

Since its March 2025 listing, CoreWeave has financed itself mostly through debt, and direct equity dilution has been modest relative to the size of the capital stack. Beyond the $1.5B IPO at $40 (claims #19821, #11601), common issuance consists of strategic placements: a $350M customer issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). The larger potential dilution sits in convertibles. The December 2025 offering is reported at sizes ranging from $2.0B to $2.59B across claims #18739, #18733 and #3980, and is partly offset by capped calls (claim #3986). The proposed $3.0B 2032 convertible carries a $450M option (claims #2126, #2127). The record contains no ATM or insider sale activity, so those dimensions cannot be assessed.

The mix reads as a deliberate tiering rather than patching. Contract-backed delayed draw term loans fund the asset build. DDTL 3.0 at $2.6B (claim #11922) was followed by DDTL 4.0 at $8.5B with A3/A(low) ratings (claims #2137, #11219), DDTL 5.0 at $3.1B (claim #9116), and a further $2.6B facility priced at SOFR + 5.50% and rated Ba2/BB+ (claims #8729, #8733, #8732). Unsecured notes at 9.25% and 9% (claims #4824, #4697) and the convertibles fund the corporate layer. The pace is accelerating: company-stated totals move from $17.2B cumulative (claim #4968) to more than $30B raised in 2026 year-to-date alone (claim #8735).

Use-of-proceeds language is on record (claims #11920, #1320, #2138). However, deployment data is limited to a single quarter's capex (claim #18918), so the record cannot confirm whether actual spending tracks those stated uses. Several note closings are tagged as delayed (claims #11919, #4699, #9699). The January 2026 amendments postponing debt service coverage and contract realization testing (claims #3912, #3913) also require explanation, particularly alongside the coverage covenants written into later facilities (claims #2141, #1324).

A reader should watch three things. First, whether the next DDTL keeps investment-grade treatment or follows the Ba2/BB+ path. Second, how interest expense (claim #18919) scales against reported backlog (claim #18917), given that underlying customer contracts average roughly three years while the loans run about five (claims #8731, #8730). Third, whether further covenant relief follows the January 2026 amendments.

CoreWeave's record describes a company that has financed itself mostly with debt rather than equity. The $1.5B IPO at $40 (claims #19821, #11616) netted about $1.4B (claim #4967). Since then, outright equity has come mainly from strategic placements: NVIDIA's $2B (claim #3891), Jane Street's $1B (claim #2080), and a $350M strategic customer issuance (claim #4514). Dilution so far has therefore been moderate. The more meaningful overhang sits in the convertibles, the 2025 notes upsized from $2.25B to $2.5875B (claims #18733, #3980) and a further 2032 tranche described as $2.6B and as $3.0B in different claims (claims #11765, #10325), which requires reconciliation. Capped calls blunt some of this dilution, though the reported cost moved from an estimated $295.7M to $340M (claims #18738, #3986).

The mix looks deliberate rather than improvised. Asset-level delayed-draw facilities fund GPU servers against specific customer contracts (claim #2138), and corporate high-yield and convertible paper fill the gaps around them. Within that structure, credit terms have shifted. DDTL 4.0 carried A3 ratings (claim #11219). DDTL 5.5 was rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733), with a roughly 5-year maturity against underlying contracts averaging about 3 years (claims #8730, #8731). That gap between loan life and contract life requires explanation. Upsizings of the 2025 notes (claims #11918, #4698) suggest demand existed at 9% to 9.25% coupons (claims #4824, #4697). The January 2026 amendment to the minimum liquidity covenant and the postponed debt service and contract realization ratio tests (claims #3911–#3913) are the clearest sign that original terms were revised under pressure.

On use of proceeds, the data supports stated intent but says little about actual deployment. Facility-level purposes are disclosed, but no capex or draw-down figures appear that would let a reader confirm proceeds went where announced. By Q2 2026, interest expense of 640 (claim #7341, presumably millions) set against revenue reported as 3 (claim #7340, presumably billions), widening losses (claim #1495), and a debt-to-equity ratio near 6 (claim #3058) show the financing cost is now material to operating results. No insider transactions appear in the records provided, and the record is truncated after August 2026.

A sophisticated reader should watch four things: whether covenant tests deferred in January 2026 are reinstated on schedule, whether new facilities continue migrating toward non-investment-grade pricing, how the 2032 convertible's final size is settled, and whether the $104B backlog (claim #1493) converts to cash quickly enough to service the more than $30B raised year-to-date (claim #8735).

CoreWeave's record shows a company that financed itself almost continuously after its March 2025 listing. The equity component is comparatively small: the IPO raised $1.5B gross (claim #19821) and about $1.4B net (claim #4967), followed by strategic placements from NVIDIA of $2B (claim #3891) and Jane Street of $1B (claim #10305). Most direct dilution came from these negotiated sales. The larger share-count consideration sits in the convertible layer. The December 2025 convertible appears in the record at $2.0B (claim #18739), $2.25B (claim #18733) and $2,587.5M (claim #3980), with capped call costs of $295.7M (claim #18738) and $340M (claim #3986). The December pricing alone accounts for these progressions, but the figures require reconciliation. A second convertible of $3.0B due 2032 followed in April 2026, with a $450M option (claims #10325, #10326).

The mix reads as a deliberate structure rather than patching. Contract-backed delayed draw term loans carry the asset build: DDTL 4.0 at $8.5B with A3/A(low) ratings (claims #11218, #11219), DDTL 5.0 at $3.1B (claim #9116), and DDTL 5.5 at $2.6B (claim #8729). Unsecured notes and converts fund the corporate layer. The structure did shift over time. DDTL 5.5 was rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733), below the investment-grade tier achieved months earlier. Its roughly five-year maturity (claim #8730) also runs against underlying customer contracts averaging about three years (claim #8731). Separately, the January 2026 amendments postponed debt service coverage and contract realization ratio testing (claims #3912, #3913). The record does not explain these postponements.

The stated use of proceeds for the DDTLs is GPU servers and contract-specific capex (claims #2138, #1320). The record contains no deployment data that would let a reader confirm that. The same gap applies to the unsecured notes, whose proceeds disclosure (claim #11920) lacks detail.

The cumulative pace is steep. The company cited $28B in commitments over twelve months (claim #11221), then more than $30B year-to-date by August 2026 (claim #8735). Over the same period, Q2 2026 interest expense was reported at $640M (claim #7341) and net margin at roughly -25% (claim #7343). The record contains no insider transactions, so no conclusion can be drawn there.

Going forward, the useful items to watch are the first covenant tests after the postponements end, the conversion thresholds on the two convertible tranches, and whether new facilities keep their investment-grade rating. The most important is whether the roughly $100B backlog (claims #647, #962) turns into contracted revenue over the full life of the debt that funds it.

CoreWeave's record shows a company that used its equity listing mainly to open access to debt markets. The IPO at $40 per share raised $1.5 billion gross and roughly $1.4 billion net (claims #19821, #11616, #4967). Direct equity dilution since then has come from a small number of strategic placements: NVIDIA's $2 billion Class A investment (claim #3891), Jane Street's $1 billion (claim #10305) and a $350 strategic customer issuance (claim #4514). Most of the potential dilution sits instead in convertibles. The December 2025 convertible grew from $2 billion to $2.25 billion and reached $2.5875 billion with the option (claims #18739, #18733, #3980). The April 2026 convertible due 2032 was $3.0 billion plus a $450 million option (claims #10325, #10326). Capped calls partly offset this dilution, although the record gives the capped-call cost as both $295.7 million and $340 million (claims #18738, #3986), and that difference requires explanation.

The mix is more systematic than opportunistic. Revolver capacity went from $1.5 billion to $2.5 billion (claims #11923, #11766). Delayed-draw facilities were numbered in sequence: DDTL 3.0, 4.0, 5.0 and 5.5. Unsecured notes came roughly quarterly, and several offerings were upsized, including the 2025 notes from $1.5 billion to $2 billion and the July notes by $250 (claims #11921, #11917, #4698). Three developments complicate that picture. First, in January 2026 debt service coverage and contract realization testing were postponed and the minimum liquidity covenant was amended (claims #3912, #3913, #3911), a few weeks after the convertible closed. Second, DDTL 4.0 carried A3/A(low) ratings (claim #11219), while the August DDTL 5.5 was rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). Third, DDTL 5.5 has about a five-year maturity (claim #8730) but rests on customer contracts averaging about three years (claim #8731).

The stated uses of proceeds are specific for the term loans, which fund customer-contract capex and GPU servers (claims #2138, #1320). They are generic or capped-call directed for the convertibles (claim #18743). The record does not allow a reconciliation of proceeds against deployment beyond the Q2 2026 capex figure (claim #18918).

No insider transactions appear in this record. The data also ends mid-entry after the Q2 2026 results, so later events may be missing. Readers should watch four things: whether the postponed covenant tests resume on schedule; how Q2 interest expense of $640 (claim #18919) grows against widening losses (claim #1495) and a debt-to-equity ratio near 6 (claim #7345); whether new facilities keep pricing below investment grade; and how backlog reported at $104 (claim #1493) turns into cash as the three-year contracts roll off.

CoreWeave's record is overwhelmingly a debt story rather than an equity-dilution story. The March 2025 IPO raised $1.5B at $40 per share (claims #19821, #11601), about $1.4B net (claim #4967). Nearly every subsequent financing has been debt or debt-like. Equity dilution since the listing has come mainly through strategic placements: NVIDIA's $2B Class A investment (claim #17594) and Jane Street's $1B (claim #10305). The other source is contingent share issuance embedded in convertibles: the $2.5875B December 2025 convertible at 1.75% (claims #3980, #18736) and the $3.0B convertible due 2032 (claim #2126). The December convertible was paired with $340M of capped calls (claim #3986), which suggests management is actively trying to limit that contingent dilution.

The mix reads as a coherent strategy that has been under visible strain. The core architecture is recurring delayed-draw term loans secured against customer contracts. It runs from DDTL 3.0 at $2.6B, later $3.0B (claims #11922, #4277), to the investment-grade $8.5B DDTL 4.0 rated A3 (claims #11218, #11219), then DDTL 5.0 at $3.1B (claim #9116) and DDTL 5.5 at $2.6B (claim #8729). High-coupon unsecured notes sit alongside this structure, including the 9.25% notes due 2030 (claim #4824), the upsized 9% $1.75B notes (claims #4697, #4698) and the $3.5B notes due 2032 (claim #8927). Two points require explanation. The January 2026 amendments postponed debt service coverage and contract realization ratio testing and revised the minimum liquidity covenant (claims #3912, #3913, #3911). In addition, the most recent facility priced at SOFR+5.50% with Ba2/BB+ ratings (claims #8733, #8732), a step down from DDTL 4.0's investment-grade terms only months earlier.

Use-of-proceeds claims exist (claims #11920, #1320, #2138), but the record does not show deployment detail, so it cannot confirm whether funds went where they were announced. Q2 2026 capex (claim #18918) and interest expense (claim #7341) are disclosed without clear units in this dataset. Company-stated cumulative totals rose from $28B over twelve months (claim #11221) to more than $30B year-to-date by August (claim #8735).

A careful reader should watch several things. The first is whether postponed covenant tests resume on schedule. The second is how DDTL 5.5's roughly five-year maturity (claim #8730) is serviced against underlying contracts that average about three years (claim #8731). The third is whether future facilities regain investment-grade pricing. The last is the conversion share count as the convertibles season. No insider transactions appear in this record, and the claim list is truncated after the Q2 2026 results, so this read is necessarily incomplete.

CoreWeave's record is overwhelmingly a debt story. Direct equity dilution has come in discrete, strategically placed tranches: the $1.5B IPO (claim #19821, net $1.4B per claim #4967), a $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891), and Jane Street's $1B (claim #10305). Potential dilution now sits mainly in convertibles. The December 2025 offering grew from a stated $2B (claim #18739) to $2.25B (claim #18733) and then $2.5875B with the option (claim #3980). A second $3.0B convertible due 2032 followed (claim #10325), with a further $450M option (claim #10326). Capped calls partly offset this, though the record gives two figures for them, $295.7M (claim #18738) and $340M (claim #3986), which requires explanation.

The mix looks structured rather than improvised. Asset-backed DDTL facilities carry the GPU build, with proceeds tied to customer-contract capex (claims #2138, #1320). High-coupon unsecured notes, at 9.25% (claim #4824) and 9% (claim #4697), plus the revolver, which rose from $1.5B (claim #11923) to $2.5B (claim #11766), supply corporate flexibility. The pace, however, is steep. The company cites $17.2B raised by May 2025 (claim #4968), $28B over twelve months (claim #11221), and more than $30B year-to-date by August 2026 (claim #8735).

The record does not let a reader reconcile stated use of proceeds against actual deployment. Q2 2026 capex is disclosed (claim #18918), but no facility-level draw data appears.

Several terms have shifted. The debt service coverage and contract realization tests were postponed and the minimum liquidity covenant amended (claims #3912, #3913, #3911). The DDTL 4.0 carried A3/A(low) ratings (claim #11219), while DDTL 5.5 is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). DDTL 5.5 also runs about five years (claim #8730) against underlying contracts averaging about three (claim #8731).

A sophisticated reader should watch four things: when the postponed covenant tests resume, whether future DDTLs keep the lower rating tier, and how Q2 interest expense of $640M (claim #18919) trends against revenue. The fourth is whether backlog anchors such as Meta's $21B commitment (claim #2124) and OpenAI's $6.5B (claim #4297) turn into cash on the timelines the lenders assume. No insider sales appear in this record.

CoreWeave's record shows a company that went public mainly to open doors to debt markets, not to fund itself through equity. The IPO netted roughly $1.4B (claim #4967), a small amount next to the aggregate totals the company later cited: $17.2B raised by May 2025 (claim #4968), $28B in the twelve months to March 2026 (claim #11221), and more than $30B year-to-date by August 2026 (claim #8735). Direct equity dilution has been modest and selective. It came through placements to strategic holders, namely NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305), rather than broad follow-ons. The main source of potential dilution is instead the two convertible issues, the December 2025 notes at 1.75% (claim #18736) and the $3.0B notes due 2032 (claim #2126), with capped calls partly offsetting the first (claim #3986). No ATM program and no insider transactions appear in this dataset, so no conclusion can be drawn on either.

The financing mix follows a clear logic. Asset-backed delayed-draw term loans are sized against customer contracts, running from DDTL 3.0 through 5.5. High-coupon unsecured notes, such as the 9.250% 2030s (claim #4824), and convertibles fill the corporate layer above them. The quality of that collateral is uneven. DDTL 4.0 carried A3 investment-grade ratings (claim #11219). DDTL 5.5, by contrast, priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732) and rests on contracts averaging about three years (claim #8731) against a roughly five-year maturity (claim #8730). That gap in tenor requires explanation.

The stated use of proceeds, customer-contract capex and GPU servers (claim #2138), is consistent with the Q2 2026 capex figure (claim #18918). The dataset does not allow a line-by-line reconciliation. Several internal figures also need reconciling. The December convertible appears at $2.0B, $2.25B, and $2.5875B (claims #18739, #18733, #3980), with net proceeds recorded as both $2.21B and $2.54B (claims #18737, #3981). The April 2026 convertible appears as both $2.6B and $3.0B (claims #10327, #2126).

The measures most worth watching come from Q2 2026. Interest expense was $640M for the quarter (claim #18919), the debt-to-equity figure was recorded at 6 (claim #3058), and operating losses widened (claim #1495). The earlier postponement of debt-service-coverage and contract-realization testing (claims #3912, #3913) means future covenant tests and the renewal of roughly three-year customer contracts are the next evidence to look for. Conversion or repricing of the backlog (claim #1493) into cash flow is the variable the whole structure depends on. The record supplied here is truncated, so these observations may be incomplete.

CoreWeave's record since the March 2025 listing describes a company financed mostly by debt, with equity arriving in a few large, counterparty-specific placements rather than through recurring public issuance. The IPO itself was modest, $1.5B at $40 per claim #19821 and about $1.4B net per claim #4967. Direct dilution since then has come from the $350M strategic customer issuance (claim #4514), NVIDIA's $2B Class A purchase (claim #17594) and Jane Street's $1B investment (claim #10305). The larger potential dilution sits in the convertibles. The December 2025 deal grew to $2.5875B (claim #3980) and was paired with $340M of capped calls (claim #3986). The April 2026 proposal adds $3.0B due 2032 (claim #10325), with a $450M option (claim #10326). Dilution is therefore stepping up, but mostly in contingent form. The mix looks more deliberate than opportunistic. Asset-level delayed-draw facilities carry the GPU build-out: DDTL 3.0, the $8.5B DDTL 4.0 (claim #11218), DDTL 5.0 at $3.1B (claim #9116) and DDTL 5.5 at $2.6B (claim #8729). Unsecured notes and converts sit above them at the corporate level. Two points in that structure require explanation. The January 2026 amendments postponed debt service coverage and contract realization testing and altered the minimum liquidity covenant (claims #3912, #3913, #3911), only months after these facilities were put in place. Credit quality has also varied across the facilities. DDTL 4.0 carried A3/A(low) ratings (claim #11219), while DDTL 5.5 is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). On use of proceeds, the stated purposes are customer-contract capex and GPU servers (claim #2138) and capped call funding (claim #18743). The record contains no deployment data against which to test them, so that question stays open. The record also contains no insider transactions. A reader should watch several things next. First, whether covenant testing resumes on schedule. Second, how DDTL 5.5's roughly five-year maturity (claim #8730) compares with its underlying contracts of about three years (claim #8731). Third, how Q2 2026 interest expense of $640M (claim #7341) develops against a backlog stated near $100B (claim #647). Finally, cumulative capital raised is climbing quickly, from more than $20B year-to-date in May (claim #9123) to more than $30B by August (claim #8735). The pace of new financings deserves the same attention as their terms. The claim list supplied here is truncated after the Q2 2026 results, so later events are not reflected.

CoreWeave's record describes a company that has used public equity sparingly and debt aggressively. The IPO priced at $40 (claim 11616) and raised $1.5B gross (claim 19821), about $1.4B net (claim 4967). That is small against the cumulative totals the company itself reports: $17.2B by May 2025 (claim 4968), $28B over the trailing twelve months by March 2026 (claim 11221), and more than $30B year-to-date by August 2026 (claim 8735). Direct common-stock dilution has come mainly through strategic placements rather than marketed follow-ons, including the $350M customer issuance (claim 4514), NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 2080). The larger latent dilution sits in the two convertible tranches: $2.5875B including the option (claim 3980) and a proposed $3.0B due 2032 with a $450M option (claims 2126, 10326). The record contains no insider sale disclosures, so that part of the mix cannot be assessed.

The structure looks deliberate rather than patched together. Contract-backed delayed-draw facilities fund GPU capex tied to specific customers (claim 2138). The unsecured market, at 9.25% and 9% coupons (claims 4824, 4697), supplies corporate liquidity. The rating path runs from A3 on the $8.5B DDTL 4.0 (claim 11219) to Ba2/BB+ at SOFR+5.50% on the later facility (claims 8732, 8733). That spread suggests the company is tiering collateral quality instead of reaching for one type of capital.

The announced use of proceeds cannot be checked against actual deployment from this record. Commitments such as OpenAI's $6.5B (claim 4297) and Meta's $21B (claim 2124) are stated, but the claims do not show drawn balances by facility. Several figures require explanation. The first convertible shows net proceeds of $2.5422B and a capped call of $340M (claims 3981, 3986) alongside $2.2106B and $295.7M (claims 18737, 18738), which is probably a pre-option versus post-option difference. It is also unclear whether the SOFR+5.50% pricing belongs to DDTL 5.0 or DDTL 5.5 (claims 8733, 8734).

A careful reader should watch four things. First, the January 2026 amendments postponed debt-service-coverage and contract-realization testing and revised the minimum liquidity covenant (claims 3911–3913); the date testing resumes, and the headroom at that point, matters. Second, the underlying contracts average about three years (claim 8731) while the facility runs about five (claim 8730), which raises renewal risk. Third, quarterly interest expense of $640M (claim 7341) should be tracked against revenue. Fourth, conversion-price proximity will determine whether the convertibles settle in shares.

CoreWeave's post-listing capital history is predominantly a debt story, with equity used as a periodic backstop. The IPO at $40 raised $1.5B gross (claim #19821) and roughly $1.4B net (claim #4967). Within weeks the company began a rolling sequence of high-coupon unsecured paper, including the $2.0B 9.250% notes due 2030 (claim #11917) and the $1.75B 9% notes, which were upsized by $250M (claims #4658, #4698). Alongside these came asset-backed delayed-draw facilities: DDTL 3.0 at $2.6B (claim #11922), a $3.0B facility (claim #4277), DDTL 4.0 at $8.5B (claim #11218), DDTL 5.0 at $3.1B (claim #1319), and DDTL 5.5 at $2.6B (claim #8729). Management's own tally exceeds $30B year-to-date (claim #8735).

The dilution vector sits in the convertibles and the strategic placements. These include the $2.5875B December convertible (claim #3980), the $3.0B 2032 convertible with a $450M option (claims #2126, #10326), NVIDIA's $2B (claim #3891), Jane Street's $1B (claim #10305), and an earlier $350M strategic customer issuance (claim #4514). Capped calls blunt part of the convert dilution. The recorded cost requires reconciliation, however: $340M appears in claim #3986 and $295.7M in claim #18738. Claims #11765 and #10327 also cite a $2.6B convertible that does not obviously match the $3.0B figure.

The mix reads as a coherent architecture rather than patching. Contract-secured DDTLs fund GPU capex (claim #2138, claim #1320), while unsecured and convertible layers fill the gaps. Two data points complicate that reading. The January 2026 amendments postponed debt service coverage and contract realization testing and revised the liquidity covenant (claims #3912, #3913, #3911). And pricing has diverged across the stack. DDTL 4.0 carried A3/A(low) ratings (claim #11219), while DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732).

Use-of-proceeds language is facility-specific for the DDTLs. For the notes it is general (claim #11920), and the record does not permit verifying actual deployment against it. No insider transactions appear in the data provided, and the record itself is truncated.

A sophisticated reader should watch several items. The first is when the postponed covenant tests resume. The second is the stated mismatch between DDTL 5.5's roughly five-year maturity and underlying contracts averaging roughly three years (claims #8730, #8731). The third is Q2 interest expense of $640M (claim #7341) against revenue conversion of the $104B backlog (claim #1493). The fourth is the cumulative conversion overhang as the 2030–2032 maturities approach.

CoreWeave's record describes a company financed mainly with debt, with equity used sparingly. Equity came from the $40 IPO (claim 11616), which raised $1.5 billion gross (claim 19821) and roughly $1.4 billion net (claim 4967). Later equity arrived in discrete strategic blocks: a $350 million customer share issuance (claim 4514), NVIDIA's $2 billion Class A investment (claim 3891) and Jane Street's $1 billion (claim 10305). Most of the capital was debt. Stated cumulative raises went from $17.2 billion (claim 4968) to $28 billion over twelve months (claim 11221) and then to more than $30 billion year-to-date by August 2026 (claim 8735). Direct dilution has therefore been modest, but conversion risk has grown. There was a December 2025 convertible of up to $2.59 billion (claim 3980) and an April 2026 $3.0 billion convertible due 2032 (claim 10325). Both carry capped calls, which dampen that exposure without removing it. The record also contains inconsistent sizing figures for these deals, such as $2.6 billion in claim 11765 against $3.0 billion in claim 10325, and $295.7 million against $340 million for the capped calls (claims 18738, 3986). These discrepancies require reconciliation. The mix follows a recognisable strategy. Delayed-draw term loans are secured against specific customer contracts, and claim 2138 ties proceeds to contract capex and GPU servers. Unsecured notes and converts add corporate-level liquidity. Deal terms vary widely, however. DDTL 4.0 carried A3 ratings (claim 11219), while DDTL 5.5 priced at SOFR plus 5.50% with Ba2/BB+ ratings (claims 8733, 8732), and the unsecured notes carry 9% to 9.25% coupons (claims 4697, 4824). Whether proceeds were deployed as announced cannot be checked from this record. Q2 2026 interest expense of $640 million (claim 7341) and a $104 billion backlog (claim 1493) are consistent with capacity being built against contracts, but no disclosed figure reconciles draws to capex. Three items require explanation. In January 2026 the covenants were amended: the minimum liquidity covenant was revised, and debt service coverage and contract realization testing were postponed (claims 3911, 3912, 3913). DDTL 5.5 has a maturity of roughly five years, while the contracts underneath it average about three (claims 8730, 8731). Third, the record contains the $270 million termination fee (claim 4739) and a Series C put repurchase obligation (claim 4276) but does not discuss either further. The things to watch are when the postponed covenants return to testing, how interest expense grows relative to revenue, how the convertibles trade against their conversion terms, and whether renewals keep pace with the average three-year term of the underlying customer contracts. No insider or ATM activity appears in the record.

CoreWeave's record is dominated by debt, not equity. After the $1.5B IPO at $40 (claims 19821, 11616), which netted $1.4B (claim 4967), the company returned to public equity markets only through strategic placements: NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 10305), and a $350M strategic customer issuance (claim 4514). Direct dilution has therefore been episodic and investor-selected. The larger latent dilution sits in two convertible programs, the December 2025 notes upsized from $2.25B to $2.5875B (claims 18733, 3980) and the $3.0B 2032 notes with a $450M option (claims 2126, 10326). These were partly hedged by capped calls of $295.7M to $340M (claims 18738, 3986); the gap between those two figures requires explanation, as does the $2.6B size recorded for the 2026 convertible (claims 11765, 10327).

The mix shows a coherent direction. Early high-coupon unsecured paper, 9.25% (claim 4824) and 9% (claim 4697), gave way to contract-backed delayed-draw facilities. That shift peaked with the $8.5B DDTL 4.0 rated A3/A(low) (claims 11218, 11219). The trajectory is not uniformly improving, however. The later $2.6B DDTL 5.5 priced at SOFR plus 5.50% and was rated Ba2/BB+ (claims 8729, 8733, 8732). Its roughly five-year maturity also runs against underlying customer contracts that average about three years (claims 8730, 8731). Cumulative raising is steep: over $20B year-to-date by May (claim 9123) and over $30B by August (claim 8735).

On use of proceeds, the stated purposes, GPU servers and contract-linked capex (claims 2138, 1320), fit the facility structures. The record contains no deployment data, so actual spending against those stated purposes cannot be verified here. The January 2026 amendments to the minimum liquidity covenant and the postponed debt service coverage and contract realization testing (claims 3911, 3912, 3913) are the clearest signs that operations lagged the financing plan. The record contains no insider transactions.

A careful reader should watch four things. The first is when the postponed covenant tests resume and whether they are met. The second is whether future DDTLs price nearer the investment-grade DDTL 4.0 or the sub-investment-grade 5.5. The third is how Q2 interest expense, recorded at $640 (claim 7341, units as disclosed), develops against revenue as the backlog figure of $104 (claim 18917) converts to revenue. The fourth is conversion overhang as the share price moves relative to the convertible strikes.

CoreWeave's post-IPO capital history is overwhelmingly a debt story, and the pace has accelerated rather than tapered. The company listed at $40 per share (claim 11601) and netted roughly $1.4B (claim 4967). Cumulative raise figures then climb quickly: $17.2B by May 2025 (claim 4968), $28B in trailing commitments by March 2026 (claim 11221), and more than $30B year-to-date by August 2026 (claim 8735). Direct common dilution has been modest by comparison. It comes mainly through strategic placements, including NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 10305) and a $350M customer issuance (claim 4514). Two convertible programs carry the larger latent dilution: the December 2025 notes, which grew to $2.59B with the option (claim 3980), and the proposed $3.0B 2032 converts plus a $450M option (claims 10325, 10326). The capped call allocation rose from $295.7M (claim 18738) to $340M (claim 3986) as the December deal was upsized, which suggests management is actively managing conversion dilution.

The mix reads as a coherent structure rather than patching. Delayed-draw facilities secured against customer contracts fund GPU capex (claims 2138, 1320). Unsecured high-yield notes and converts sit above them at the corporate level. Within the DDTL series, pricing and credit quality vary widely. DDTL 4.0 carried A3/A(low) ratings (claim 11219). DDTL 5.5 is rated Ba2/BB+ at SOFR + 5.50% (claims 8732, 8733). That spread requires explanation, as does the gap between the roughly three-year average tenor of the underlying contracts (claim 8731) and the facility's roughly five-year maturity (claim 8730).

The January 2026 amendments are the least conventional items on record. They amended a minimum liquidity covenant (claim 3911) and postponed both debt service coverage and contract realization ratio testing (claims 3912, 3913). Later facilities reintroduce DSCR covenants (claims 2141, 1324).

Use of proceeds is stated at issuance (claims 11920, 18743), but the record contains no deployment reconciliation, so whether proceeds tracked their announced purpose cannot be assessed here. The record also contains no insider transactions.

Several things are worth watching. The first is when the postponed covenant tests resume and how they read. The second is Q2 2026 interest expense of $640M (claim 7341) set against a reported $104B backlog (claim 1493). The third is whether future contract-backed facilities price closer to the investment-grade DDTL 4.0 or the sub-investment-grade DDTL 5.5.

CoreWeave's record shows a company that has used public equity sparingly and debt heavily. After the $40 IPO (claim #11616) and roughly $1.4B in net proceeds (claim #4967), new equity came mainly from strategic holders: NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). Dilution to date is therefore modest and negotiated. The more meaningful future dilution sits in the convertibles: roughly $2.59B issued in December 2025 (claim #3980), partly offset by capped calls (claim #3986), followed four months later by a $3.0B tranche due 2032 (claim #2126). The record contains no insider transactions, so no view on insider selling can be drawn.

The mix looks structured rather than improvised. Asset-level delayed-draw term loans grew in size and rating quality, from $2.6B (claim #11922) to the $8.5B DDTL 4.0 rated A3/A(low) (claims #11218, #11219) and the syndicated $3.1B DDTL 5.0 (claim #9117). High-yield senior notes sit alongside them: 9.25% notes due 2030 (claim #4824), a 9% issue (claim #4697), and later $1.25B, $1B and $3.5B offerings (claims #2125, #2056, #8927). The trajectory is not uniformly upward, however. The August DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732), well below the DDTL 4.0 rating. The January 2026 covenant amendments, which altered the minimum liquidity requirement and postponed debt-service-coverage and contract-realization testing (claims #3911–#3913), require explanation because they arrived between two large raises.

On use of proceeds, the DDTLs are explicitly tied to customer-contract capex (claim #2138), anchored by commitments such as OpenAI's $6.5B (claim #4297) and Meta's $21B (claim #2124). Q2 2026 capex (claim #18918) and a backlog near $100B (claims #647, #18917) are directionally consistent with that framing. The record does not reconcile specific raises to specific deployments, and the stated use for the May 2025 notes (claim #11920) is not itemized.

A careful reader should watch three things. First, Q2 interest expense of $640M (claim #18919) against a debt-to-equity ratio near 6 (claim #3162). Second, the mismatch between roughly 3-year underlying contracts and a roughly 5-year facility term on DDTL 5.5 (claims #8731, #8730), which leaves renewal risk inside the loan's life. Third, whether the postponed covenant tests resume on schedule. Cumulative raises of more than $30B year-to-date (claim #8735) mean refinancing cadence is now itself a core operating variable. Note also that the supplied record is truncated partway through Q2 2026 results, so later events are not assessed here.

CoreWeave's capital history after the IPO is primarily a debt story. The March 2025 listing at $40 (claim #11601) raised about $1.4B net (claim #4967). Since then, direct equity has come mainly through two strategic placements: NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305). Dilution so far has been episodic rather than continuous. Most of the forward dilution risk sits in the convertibles: the December 2025 issue, sized at $2.59B with the option (claim #3980), and the proposed $3.0B 2032 convert (claim #10325). The record also carries a $2.6B figure for that 2026 convert (claim #10327), and the difference requires reconciliation.

The mix looks more deliberate than patched. Asset-level delayed-draw facilities fund contracted GPU buildouts. DDTL 4.0 at $8.5B carried an A3 rating (claims #11218, #11219), while DDTL 5.5 at $2.6B priced at SOFR + 5.50% with a Ba2 rating (claims #8729, #8733, #8732). The spread between those two facilities suggests a tiered structure in which weaker collateral pays materially more. Unsecured notes and converts fill the corporate layer.

The pace has accelerated. Management cited $28B committed over twelve months (claim #11221), more than $20B year-to-date by May (claim #9123), and more than $30B by August (claim #8735). Several offerings were upsized, including the 2030 notes, which went from $1.5B to $2.0B (claims #11921, #11917).

On use of proceeds, the record states intended uses (claims #1320, #2138) but contains no deployment data, so whether proceeds tracked their stated purpose cannot be assessed. One diversion is explicit: capped calls absorbed roughly $296M to $340M of convertible proceeds (claims #18738, #3986). The record contains no insider transactions.

Three items warrant attention. First, the January 2026 amendments revised the minimum liquidity covenant and postponed contract-realization and debt-service-coverage testing (claims #3911, #3912, #3913). Whether and when those tests resume in their original form is the most direct indicator of covenant headroom. Second, DDTL 5.5 rests on customer contracts averaging about three years (claim #8731) but carries roughly a five-year maturity (claim #8730). That leaves renewal risk inside the loan term. Third, Q2 2026 interest expense of $640M (claim #7341) against a −25% net margin (claim #7343) means servicing costs will increasingly test how quickly the $104B backlog (claim #1493) converts to cash.

CoreWeave's record describes a company that has financed itself overwhelmingly through debt, with equity dilution arriving mainly through strategic placements and convertibles rather than follow-on offerings. After the $40 IPO (claim #11601) and roughly $1.4B of net proceeds (claim #4967), the direct equity events are strategic: a $350M customer share issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). The larger dilution vector is convertible paper. The December 2025 convertible sized at $2.5875B (claim #3980) carries roughly 30.0M conversion shares (claim #3984). The proposed $3.0B convertible due 2032, with a $450M option (claims #2126, #2127), would add to that overhang. Several claims also reference a $2.6B convertible (claims #10327, #11765) that does not reconcile cleanly with either deal, and this requires explanation. Capped call figures are also reported at both $295.7M and $340M (claims #18738, #3986).

The debt mix reads as a deliberate, repeating template rather than patching. Asset-backed DDTL facilities are sized to customer contracts and financed senior notes stacked beside them: the 9.25% notes (claim #4824), the 9% notes (claim #4697), and $3.5B due 2032 (claim #8927). The trajectory in pricing is less uniform. DDTL 4.0 carried A3 ratings (claim #11219), while DDTL 5.5 priced at SOFR plus 5.50% with Ba2/BB+ ratings (claims #8733, #8732). Its underlying contracts average about three years (claim #8731) against a roughly five-year maturity (claim #8730), a tenor gap that deserves attention.

Stated use of proceeds, namely customer-contract capex and GPU servers (claims #2138, #1320), is consistent across filings. However, the record contains no deployment figures that would allow a reader to confirm it. The record also contains no insider transactions, so that dimension cannot be assessed.

The most material items to watch are the January 2026 amendments. These revised the minimum liquidity covenant and postponed debt service coverage and contract realization testing (claims #3911, #3912, #3913), so the terms of resumed testing matter. Q2 interest expense of $640M (claim #7341) against a -25% net margin (claim #7343) sets the scale of the servicing burden. Aggregate raises of more than $30B year-to-date (claim #8735) make the pace of future issuance the variable to track, alongside how much of the $104B backlog (claim #1493) converts to cash.

CoreWeave's post-listing capital history is predominantly a debt story, with common-equity dilution arriving in measured, negotiated blocks rather than broad follow-ons. After the $40 IPO (claim #11616) produced roughly $1.4B in net proceeds (claim #4967), new equity came through strategic placements, the NVIDIA $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305), and through two convertible offerings: the 1.75% notes of December 2025 (claims #18733, #18736) and the proposed $3.0B notes due 2032 (claim #10325). The December convertible's figures require reconciliation, with a $2.25B base (claim #18733) against a $2.59B total (claim #3980) and capped-call costs of $295.7M (claim #18738) versus $340M (claim #3986), most plausibly the difference between pre- and post-option sizing. The capped calls indicate an effort to contain conversion dilution. The record contains no insider sale transactions, so that dimension cannot be assessed.

The debt mix reads as a coherent strategy that has been refined over time. Early unsecured issuance was expensive, with the 9.250% notes due 2030 upsized to $2B (claims #11921, #11917) and the 9% notes to $1.75B (claims #4697, #4698). The company then moved toward contract-backed delayed-draw facilities, which culminated in the $8.5B DDTL 4.0 rated A3/A(low) (claims #11218, #11219). That improvement was not linear. DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732), and unsecured notes continued alongside, including $3.5B due 2032 (claim #8927). Cumulative raises escalated from more than $20B year-to-date in May (claim #9123) to more than $30B by August (claim #8735).

Stated use of proceeds, such as financing customer-contract capex and GPU servers (claim #2138), is disclosed at a general level (claims #11920, #1320). The record offers no deployment data against which to test it.

A careful reader should track several things next. The January 2026 amendments to the minimum-liquidity covenant and the postponed debt-service-coverage and contract-realization testing (claims #3911–#3913) need to show a resumption date. DDTL 5.5's roughly five-year tenor sits against underlying contracts averaging about three years (claims #8730, #8731), which leaves renewal exposure inside the loan term. Q2 2026 interest expense of $640M (claim #7341) at a -25% net margin (claim #7343) places heavy weight on converting the reported $104B backlog (claim #1493) into revenue on schedule.

CoreWeave's record since its $40 IPO (claim #11601) is mostly debt, and it has grown heavier over time. Direct equity dilution has been modest and episodic: IPO net proceeds of $1.4B (claim #4967), then strategic placements from NVIDIA at $2B (claim #3891) and Jane Street at $1B (claim #10305). The more consequential dilution sits in convertible paper. The December 2025 convertible upsized from a $2.0B base (claim #18739) to $2.5875B (claim #3980) and references roughly 30.0M underlying shares (claim #3984). A further $3.0B convertible due 2032 was proposed in April 2026 (claim #10325). Capped calls (claim #3986) cushion part of that overhang, but they do not remove it.

The mix reads as a deliberate structure rather than patching, though that structure is being stretched. Asset-backed delayed-draw facilities anchor the stack, and DDTL 4.0 carried A3/A(low) ratings (claim #11219). Around them sit unsecured senior notes at 9.25% (claim #4824) and 9% (claim #4697), with converts filling the gaps. Later layers look different. DDTL 5.5 is rated Ba2/BB+ at SOFR + 5.50% (claims #8732, #8733), which suggests the investment-grade template did not carry forward. The January 2026 postponement of debt-service-coverage and contract-realization testing (claims #3912, #3913) and the amended liquidity covenant (claim #3911) also require explanation alongside the pace of new borrowing.

On use of proceeds, the record states intent, for example GPU and customer-contract capex (claim #2138) and DDTL 5.0 purposes (claim #1320). It contains no deployment data, so whether actual spending tracks those stated uses cannot be tested here. The data show no insider sales, so no view on insider selling is possible.

Capital raised grew from more than $20B year-to-date in May (claim #9123) to more than $30B by August (claim #8735). Over the same period, quarterly interest expense reached $640M (claim #7341) and debt-to-equity sits near 6x (claim #3058). A sophisticated reader should watch three things. The first is whether the postponed covenant tests resume on schedule. The second is the gap between roughly 5-year facility maturities (claim #8730) and underlying customer contracts that average about 3 years (claim #8731). The third is whether the $104B backlog (claim #962) converts to cash quickly enough to service a capital stack now priced at sub-investment-grade spreads.

CoreWeave's record describes a company that has become a near-continuous issuer. After the $40 IPO (claim #11616) produced roughly $1.4B in net proceeds (claim #4967), the capital that followed was overwhelmingly debt. The company cites $28B of commitments over twelve months (claim #11221) and more than $30B year-to-date by August 2026 (claim #8735). Direct equity dilution has come from a few discrete events: the NVIDIA $2B purchase (claim #3891) and the Jane Street $1B investment (claim #10305). The more consequential dilution risk sits in the convertibles, namely the December 2025 issue sized at $2.5875B including the option (claim #3980), with conversion terms referencing 30,003,356 shares (claim #3984), and the proposed $3.0B 2032 convertible (claim #10325). The record contains no insider transactions, so no read on insider selling is possible.

The mix looks structured rather than improvised. Asset-level delayed-draw facilities fund contracted GPU buildouts, and unsecured notes and convertibles fund the corporate layer. Several deals were upsized in market, including the 2030 notes moving from $1.5B (claim #11921) to $2B (claim #11917) and the July notes rising to $1.75B (claim #4697), which indicates demand rather than strain. The terms have drifted, however. DDTL 4.0 carried A3/A(low) ratings (claim #11219). DDTL 5.5 is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). Its roughly 5-year maturity sits against underlying contracts averaging about 3 years (claims #8730, #8731), which requires explanation.

The stated use of proceeds for the DDTL facilities is financing for customer-contract capex and GPU servers (claim #2138), and it is broadly consistent with a $104B backlog (claim #962). The record does not show actual deployment against those facilities, so tracking cannot be confirmed. On the December convertible, the capped-call allocation appears as $295.7M (claim #18738) and as $340M (claim #3986), likely reflecting the option exercise, but the gap is not reconciled in the record. The $2.6B convertible and $2.5B revolver claims appear under both February and April 2026 dates (claims #11765, #10327), which suggests the terms were revised before launch.

Several items deserve attention next. Covenant testing was postponed and a minimum liquidity covenant was amended in January 2026 (claims #3912, #3913, #3911), so watch whether debt-service-coverage tests (claims #2141, #1324) resume on schedule. Q2 2026 interest expense of $640M (claim #7341) set against a reported debt-to-equity ratio of about 6 (claim #3058) makes the coverage trajectory the central variable. The final question is whether future asset-backed facilities regain investment-grade ratings or continue at sub-investment-grade pricing.

CoreWeave's post-IPO capital history is overwhelmingly a debt story. Direct equity dilution has been comparatively contained: the $40 IPO (claim #11616) raised $1.4 billion net (claim #4967), and the later equity came in discrete, negotiated blocks from strategic counterparties, including NVIDIA's $2 billion (claim #3891), Jane Street's $1 billion (claim #10305), and a $350 million strategic customer issuance (claim #4514). The record shows no at-the-market program, and it contains no insider transaction entries, so nothing can be said about insider selling. Latent dilution sits mainly in the convertibles. The December 2025 issue reached $2.59 billion including the option (claim #3980), and a proposed $3.0 billion convertible due 2032 followed in April 2026 (claim #10325). The capped calls funded from the first issue, $340 million per claim #3986, soften but do not remove that exposure.

The mix looks structured rather than improvised. Unsecured high-yield notes at 9.25% (claim #4824) and 9% (claim #4697) funded the corporate layer. Asset-level delayed draw term loans were tied to customer contracts, which is the stated purpose in claim #2138. The DDTL 4.0 facility carrying A3/A(low) ratings (claim #11219) suggests the contract-backed template reached investment-grade treatment. The sequence was not linear, however. The January 2026 postponement of debt service coverage and contract realization testing (claims #3912 and #3913), together with an amended minimum liquidity covenant (claim #3911), requires explanation, particularly because NVIDIA's equity arrived weeks later. Several note closings are also tagged as delayed (claims #11919, #4699 and #9699).

On use of proceeds, the record states intentions, such as capped calls (claim #18743) and GPU capex (claim #2138), but contains no deployment data against which to test them. Tracking therefore cannot be assessed.

The forward questions are about carrying cost and duration. Q2 2026 interest expense of $640 million (claim #7341) sits against revenue of $2.58 billion (claim #8177) and a net margin of -25% (claim #7343). The newest DDTL 5.5 is rated Ba2/BB+ and priced at SOFR plus 5.50% (claims #8732 and #8733), well below the DDTL 4.0 rating. It also carries roughly a five-year maturity against customer contracts averaging about three years (claims #8730 and #8731). A careful reader should watch whether contract renewals close that gap and whether the postponed covenant tests resume on schedule. Another open question is whether the $104 billion backlog (claim #1493) converts into cash quickly enough to service a 2030–2032 maturity stack that is growing with each quarter's issuance.

CoreWeave's record describes a company financing itself almost continuously. Since the March 2025 IPO at $40 (claim 11616), which netted $1.4B (claim 4967), the company has come to market roughly every four to eight weeks. Management's own running totals rose from more than $20B raised year-to-date by May 2026 (claim 9123) to more than $30B by August (claim 8735). The record contains no share counts, ATM program or insider sales, so dilution cannot be sized directly. Its sources, however, can be traced. Primary equity came from strategic placements rather than public follow-ons: NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 10305). Contingent dilution sits in two convertible tranches, the $2.5875B issue (claim 3980), partly offset by a $340M capped call (claim 3986), and the $3.0B notes due 2032 (claim 2126), which carry a further $450M option (claim 10326). The mix looks structured rather than improvised. Asset-level delayed-draw facilities fund contracted GPU buildouts (claim 2138). Unsecured notes and converts sit at the corporate level. DDTL 4.0 carried A3/A(low) ratings (claim 11219). That coherence has been tested, though. The January 2026 amendments postponed debt service coverage and contract realization testing and revised the minimum liquidity covenant (claims 3912, 3913, 3911), and those changes require explanation alongside a financing program described as disciplined. The most recent facility, DDTL 5.5 at $2.6B (claim 8729), priced at SOFR + 5.50% (claim 8733) with Ba2/BB+ ratings (claim 8732). That is a clear step down in credit quality from DDTL 4.0. It also runs roughly five years (claim 8730) against underlying contracts averaging about three (claim 8731). On use of proceeds, the record states intent (claims 11920, 1320, 2138) but contains no disclosures of actual deployment, so whether the two track cannot be assessed from this data. Several announced closings were tagged as delayed (claims 11919, 4699, 9699) without further detail. For what comes next, the Q2 2026 figures frame the pressure: $640M of interest expense (claim 7341) against revenue of $2.58B (claim 8177), a -25% net margin (claim 7343) and a debt-to-equity ratio near 6 (claim 3058). A careful reader should watch whether the postponed covenant tests resume on schedule. It is also worth watching whether the $104B backlog (claim 1493) turns into cash fast enough to cover the refinancing gap that the shorter contract tenors create, and whether the pricing slide from DDTL 4.0 to DDTL 5.5 continues.

CoreWeave's record since its $40 IPO (claim 11616) describes a company that finances almost continuously, with a new instrument roughly every four to eight weeks. The trajectory is steep. The company cited $28B of commitments over twelve months (claim 11221), more than $20B year-to-date by May (claim 9123), and more than $30B by August (claim 8735). Direct equity dilution has been modest relative to that total: $1.4B in IPO net proceeds (claim 4967), a $350M strategic customer issuance (claim 4514), NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 10305). The larger dilution question is deferred into the convertibles, the $2.59B December issue (claim 3980) and the $3.0B 2032 issue with a $450M option (claims 10325, 10326). The $340M capped call (claim 3986) indicates management is actively managing that overhang. The record contains no insider sale disclosures, so no inference is drawn there.

The mix reads as a deliberate layered structure more than improvisation. Contract-backed delayed-draw facilities fund GPU deployment (claim 2138). Unsecured notes, several upsized from their initial size (claims 11921 against 11917, and 4700 against 4697), sit above them. Converts and strategic equity backfill the base. The quality of the stack, however, is not uniform. DDTL 4.0 carried A3 investment-grade ratings (claim 11219), while the August DDTL 5.5 priced at SOFR plus 5.50% with Ba2/BB+ ratings (claims 8733, 8732) against contracts averaging about three years (claim 8731). That divergence requires explanation.

The use of proceeds is disclosed in general terms (claims 11920, 1320), and the dataset offers no way to reconcile specific draws against specific deployments. The January 2026 postponement of debt service coverage and contract realization ratio testing, together with a minimum liquidity amendment (claims 3912, 3913, 3911), shows that covenant headroom was revised rather than met as originally drafted.

A sophisticated reader should watch four things. First, whether the postponed ratio tests resume on schedule. Second, how Q2's $640M interest expense (claim 7341) trends against $2.58B of revenue (claim 8177) as the June $3.5B notes (claim 8927) come on. Third, whether conversion of the 2025 and 2026 converts brings the deferred dilution forward. Fourth, whether the $104B backlog (claim 1493) converts to cash at a pace that supports the six-times debt-to-equity ratio reported for Q2 (claim 3058).

CoreWeave's record describes a company financing itself almost entirely through debt since its $40 IPO (claim #11616), which produced about $1.4B in net proceeds (claim #4967). Equity dilution has come mostly from strategic placements: the $350M customer share issuance (claim #4514), NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). Convertibles are the growing source of potential dilution. The $2.5875B convertible (claim #3980) was partly offset by $340M of capped calls (claim #3986), and it was followed by a proposed $3.0B 2032 convertible with a $450M option (claims #10325, #10326). The record contains no ATM activity and no insider transactions, so neither can be assessed. The mix follows a recognizable structure. Contract-backed delayed-draw term loans fund GPU capex, and unsecured notes and converts sit at the corporate level. The DDTL sequence runs from $2.6B (claim #11922) to $8.5B at A3/A(low) (claims #11218, #11219), then $3.1B (claim #9116), then $2.6B at Ba2/BB+ priced at SOFR+5.50% (claims #8729, #8732, #8733). That the newest facility is rated sub-investment-grade after the investment-grade DDTL 4.0 requires explanation. The corporate notes carry 9% to 9.25% coupons (claims #4824, #4697). The January 2026 postponement of the debt service coverage and contract realization tests, along with the minimum liquidity amendment (claims #3912, #3913, #3911), was not followed by any disclosed restoration. It reads more like accommodation than design. On use of proceeds, the stated purpose is customer-contract capex and GPU servers (claim #2138). The record contains no deployment data, so whether funds tracked that purpose cannot be verified. The scale is visible in the cumulative totals: $17.2B (claim #4968), $28B over twelve months (claim #11221) and more than $30B year-to-date by August 2026 (claim #8735). It also shows in Q2 2026 interest expense of $640M against $2.58B of revenue (claims #7341, #8177). The DDTL 5.5 facility has a maturity of about five years (claim #8730), while its underlying contracts average about three years (claim #8731). That gap requires explanation, because it places renewal risk on the lender's collateral. A careful reader should watch four things: whether the postponed covenants are tested on schedule, how the $104B backlog (claim #1493) converts into cash relative to debt service, whether future DDTLs price closer to DDTL 4.0 or DDTL 5.5, and how much of the 2032 convertible settles in shares.

CoreWeave's capital record since its $40 IPO (claim 11616) is overwhelmingly a debt story, and the cadence has accelerated rather than settled. The IPO's $1.4B of net proceeds (claim 4967) is now small next to the $28B of equity and debt commitments the company cited over a trailing twelve months (claim 11221) and the more than $30B it cited as raised year-to-date by August 2026 (claim 8735). Common-equity dilution has come mainly through strategic placements, not market offerings: NVIDIA's $2B (claim 17594) and Jane Street's $1B (claim 10305). The larger latent dilution sits in the convertibles, the $2.59B issue (claim 3980) with its $340M capped call (claim 3986) and the proposed $3.0B due 2032 (claim 10325), which carries a further $450M option (claim 10326). The record contains no insider transactions, so no view on insider selling is possible here.

The mix reads as a deliberate architecture more than patching. Asset-level delayed-draw facilities fund contracted capex and GPU servers (claim 2138). Unsecured notes at 9.25% (claim 4824) and 9% (claim 4697) sit alongside them, as do a revolver expanded twice (claims 11923, 11766) and equity from counterparties with commercial ties. The DDTL 4.0 facility's A3/A(low) rating (claim 11219) is the most notable structural achievement. The later DDTL 5.5 facility was rated Ba2/BB+ and priced at SOFR + 5.50% (claims 8732, 8733). Why the credit quality of successive facilities has diverged requires explanation.

The data states use of proceeds at a general level only (claims 11920, 1320, 2138). It does not allow deployment to be traced against those statements. Several closing dates are tagged delayed (claims 11919, 4699, 9699). One convertible appears at $2.6B in some filings and $3.0B in others (claims 10327, 10325), and the record does not resolve which figure applies.

A careful reader should watch four items next. The first is whether the debt service coverage and contract realization tests postponed in January 2026 (claims 3912, 3913) resume on schedule. The second is the gap between contracts averaging about three years (claim 8731) and a roughly five-year facility maturity (claim 8730). The third is Q2 2026 interest expense of $640M (claim 7341) against revenue of $2.58B (claim 8177) and a net margin of about -25% (claim 7343). The fourth is how the roughly $100B backlog (claim 647) turns into cash quickly enough to service the stack. The claim list supplied was truncated, so later events may be missing.

CoreWeave's post-IPO capital history is dominated by debt. The $40 IPO (claim 11616) produced roughly $1.4B in net proceeds (claim 4967), a small figure against the more than $30B of debt and equity the company says it secured in 2026 through August (claim 8735). Most of that came from secured delayed-draw facilities and high-yield notes. Common-equity dilution has come in discrete private placements rather than marketed follow-ons: a $350M strategic customer issuance (claim 4514), NVIDIA's $2B Class A purchase (claim 3891), and Jane Street's $1B investment (claim 10305). The more material share-count exposure is the convertible stack. That stack consists of $2.59B of notes carrying a $340M capped call (claims 3980, 3986), followed by a proposed $3.0B due 2032 with a $450M option (claims 10325, 10326). Conversion outcomes will set the actual dilution trajectory more than any single equity event.

The mix reads as a deliberate structure rather than patching. Contract-backed DDTL facilities finance GPU capex (claim 2138), unsecured notes and converts sit above them, and strategic equity arrives from counterparties with commercial ties. The DDTL series did reach A3/A(low) ratings on the $8.5B DDTL 4.0 (claims 11218, 11219). However, the later DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims 8733, 8732). That suggests the investment-grade template has not carried over uniformly, and the reason requires explanation. The underlying contracts on that facility average about three years (claim 8731) against a roughly five-year maturity (claim 8730). That tenor gap depends on renewal assumptions the record does not describe.

Use-of-proceeds language is stated for several offerings (claims 11920, 1320), but the record contains no deployment data, so actual allocation cannot be tested here. The January 2026 amendments are also relevant. They postponed debt service coverage and contract realization ratio testing and revised the minimum liquidity covenant (claims 3912, 3913, 3911). They indicate that the original covenant timeline was revised within months of the facilities closing.

The record shows no insider transactions, so no read on insider selling is possible. A sophisticated reader should watch four things. The first is when the postponed ratio tests resume and on what terms. The second is Q2 2026 interest expense of about $640M (claim 7341) relative to operating income (claim 9281). The third is debt-to-equity near 6 (claim 3058). The fourth is whether new facilities price closer to the DDTL 4.0 or DDTL 5.5 outcome. Conversion activity on the notes due 2032 is the principal dilution variable.

CoreWeave's post-IPO capital history is overwhelmingly a debt story, and equity dilution has been comparatively measured. After the $40 IPO (claim #11601) netted roughly $1.4B (claim #4967), new share issuance came mainly through negotiated strategic placements: NVIDIA's $2B Class A investment (claim #3891) and Jane Street's $1B (claim #10305). The two convertibles add contingent dilution. The December 2025 convertible of about $2.59B carried conversion terms referencing roughly 30M shares (claims #3980, #3984) and was partly offset by a $340M capped call (claim #3986). The April 2026 $3.0B convertible due 2032 added a $450M upsize option (claims #2126, #10326). The record contains no insider sale disclosures, so that dimension cannot be assessed.

The mix reads as a deliberate two-layer structure rather than opportunistic patching. Asset-level delayed-draw facilities fund contracted build-outs, and the stated use of proceeds for these is customer-contract capex and GPU servers (claims #2138, #1320). Holding-company unsecured notes and convertibles sit above that layer. Several notes offerings were upsized from their initial size, the July 2025 9% notes by $250M (claim #4698). That suggests demand was not the constraint.

The cost and quality of that capital are uneven. The $8.5B DDTL 4.0 carried A3/A(low) ratings (claim #11219). The later $2.6B DDTL 5.5 was rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). It also financed contracts averaging about three years against a roughly five-year loan (claims #8731, #8730). That mismatch requires explanation about re-contracting assumptions.

The January 2026 amendments, which postponed debt service coverage and contract realization ratio testing (claims #3912, #3913), were not revisited in later disclosures. Deployment against announced use of proceeds cannot be verified from this record, because no capex or draw-down figures appear alongside the facility announcements.

The Q2 2026 figures frame what comes next. Revenue was $2.58B (claim #8183), interest expense was $640M (claim #7374), and debt-to-equity sat near 6 (claim #7378). Against that, backlog was reported at $104B (claim #1493). A careful reader should watch four things. The first is when the postponed covenant tests resume and on what terms. The second is whether future DDTL tranches return to investment-grade ratings or continue at sub-investment-grade pricing. The third is how quickly the facilities are actually drawn relative to the more than $30B raised year-to-date (claim #8735). The fourth is whether share price levels bring the convertibles toward conversion, which would shift the balance from leverage to dilution.

CoreWeave's record since its $40 IPO (claim #11616) shows a company funding itself mostly through debt, with equity used sparingly and usually placed with strategic holders. Primary share issuance is limited to the IPO's roughly $1.4B net (claim #4967), a $350M strategic customer issuance (claim #4514), NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #10305). Dilution risk sits mainly in the convertibles. The $2.5875B December 2025 issue converts into roughly 30M shares (claim #3984) and carries a $340M capped call (claim #3986). A $3.0B 2032 convertible with a $450M upsize option followed (claims #10325, #10326). The $2.6B convertible cited in February and April 2026 (claims #11765, #10327) appears to refer back to the December deal rather than a separate issue, but the record does not settle this. No insider transactions appear in the data, so that dimension cannot be assessed.

The debt stack follows a coherent template. Asset-level delayed-draw facilities are tied to specific customer contracts, with DDTL 4.0 financing contract capex and GPU servers (claim #2138). These are layered over unsecured notes and a revolver that grew from $1.5B (claim #11923) to $2.5B (claim #11766). The pace has quickened, from more than $20B year-to-date in May (claim #9123) to more than $30B by August (claim #8735). The terms have also moved. DDTL 4.0 carried A3 investment-grade ratings (claim #11219), while the later DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims #8733, #8732). Unsecured coupons of 9.25% and 9% (claims #4824, #4697) set the baseline cost.

On use of proceeds, the DDTL disclosures are specific about what they fund (claims #2138, #1320). The unsecured notes' stated use (claim #11920) is less traceable to actual deployment. The January 2026 amendment, which postponed debt service coverage and contract realization ratio testing and revised the minimum liquidity covenant (claims #3911–#3913), requires explanation. It came in the same period the company was describing strong contract coverage.

Three things are worth watching. First, whether postponed covenant tests resume on schedule. Second, the carrying cost: Q2 2026 interest expense of $640M (claim #7341) against $2.58B of revenue (claim #8177) and a negative 25% net margin (claim #7343). Third, the tenor gap between customer contracts averaging about three years (claim #8731) and facilities running about five years (claim #8730). The $104B backlog (claim #1493) and Meta's $21B commitment (claim #2124) are the offsetting evidence. The Q2 claim set is truncated in the record provided.

CoreWeave's record shows a company that has financed its growth mainly through debt, with equity used as an occasional supplement. Since the $40 IPO (claim 11616), which netted roughly $1.4B (claim 4967), the cadence has accelerated. Management cited $28B in commitments over the trailing twelve months by March 2026 (claim 11221), more than $20B year-to-date by May (claim 9123) and more than $30B by August (claim 8735). Direct equity dilution has come in discrete strategic blocks: the NVIDIA $2B investment (claim 3891) and the Jane Street $1B investment (claim 10305). The larger contingent share overhang sits in the convertibles. The December 2025 $2.59B offering (claim 3980) carries roughly 30M conversion shares (claim 3984), partly offset by a $340M capped call (claim 3986). A further $3.0B due 2032, with a $450M option (claims 10325, 10326), will add to that overhang. The record contains no ATM program and no insider transactions, so neither can be assessed.

The mix looks coherent rather than improvised. Delayed-draw term loans secured against customer contracts fund contract-specific capex (claim 2138), while unsecured high-yield notes and converts fund the corporate layer. Two notes offerings were upsized: the 2030 notes from $1.5B to $2.0B (claims 11921, 11917) and the July 2025 notes by $250M (claim 4698). That points to demand-led sizing rather than distress. The January 2026 amendments postponing debt-service-coverage and contract-realization testing and revising the minimum liquidity covenant (claims 3912, 3913, 3911) do not fit that picture and require explanation. So does the spread between tranches. DDTL 4.0 carried A3/A(low) ratings (claim 11219), while DDTL 5.5 was rated Ba2/BB+ and priced at SOFR + 5.50% (claims 8732, 8733). The record does not say what drove that difference.

The stated uses of proceeds are specific, but the record contains no capex or deployment figures to check them against. Revenue of $2.58B at 112% growth (claims 8183, 8185) and a backlog near $104B (claim 962) are consistent with capacity coming online, but they do not verify it.

A sophisticated reader should watch four things. First, interest expense of $640M in the quarter (claim 7341) against a -25% net margin (claim 7343). Second, the gap between customer contracts averaging about three years (claim 8731) and facility maturities of about five years (claim 8730). Third, whether the postponed covenant tests resume on schedule. Fourth, the credit quality behind future DDTL tranches.

CoreWeave's record describes a company financing growth mainly through leverage. Equity is used sparingly and tends to come from strategic partners rather than the public market. After the $40 IPO (claim 11616), direct equity came mostly from counterparties with commercial ties: NVIDIA's $2B (claim 17594) and Jane Street's $1B (claim 10305). Dilution so far is therefore modest in cadence but deferred in form. The December convertible of roughly $2.59B (claim 3980) references about 30.0M underlying shares (claim 3984), partly offset by a $340M capped call (claim 3986). A proposed $3.0B convertible due 2032, with a $450M option (claims 10325, 10326), would add a second layer of contingent share issuance.

The debt mix is more coherent than opportunistic. Delayed-draw facilities are secured against customer contracts and scale in size: $2.6B (claim 11922), $3.0B (claim 4277), $8.5B (claim 11218), $3.1B (claim 9116) and $2.6B (claim 8729). Unsecured notes at 9.25% (claim 4824) and 9% (claim 4697) sit alongside them, and the revolver was expanded from $1.5B (claim 11923) to $2.5B (claim 10328). Aggregate commitments rose from $28B over twelve months (claim 11221) to more than $30B year-to-date by August (claim 8735).

Two trajectories require explanation. First, DDTL 4.0 carried A3 investment-grade ratings (claim 11219). The later DDTL 5.5 is rated Ba2/BB+ and priced at SOFR plus 5.50% (claims 8732, 8733). Second, that facility runs about five years (claim 8730) while its underlying contracts average about three (claim 8731), which leaves a renewal dependency.

The January amendments postponed debt-service-coverage and contract-realization testing and revised the minimum liquidity covenant (claims 3912, 3913, 3911). These fit a business drawing capital faster than contracted cash flows mature.

Use of proceeds is stated (claims 11920, 1320), but the record contains no deployment reconciliation, so it cannot be tested here.

Q2 results show the carrying cost. Interest expense was $640M (claim 7374) against $2.58B of revenue (claim 8183). The net margin was -25% (claim 7376) and debt-to-equity about 6 (claim 7378). Backlog of $104B (claim 1493) is the asset the structure depends on.

The record contains no insider transactions, and the claim list appears truncated. Conclusions about insider behavior are therefore not possible.

A careful reader should watch four things: when covenant testing resumes, whether future DDTLs regain investment-grade treatment, how contract renewals track facility tenors, and whether convertible share-price thresholds are approached.

The cadence is the story. Roughly eighteen months separate a $40 IPO producing $1.4B net (claims 11616, 4967) from a disclosed $30B of debt and equity secured year-to-date (claim 8735), with an intermediate marker of $28B in commitments over a trailing twelve months (claim 11221). Equity was a one-time event; everything since has been borrowed, converted, or sold to counterparties with a commercial interest. The escalation is not linear but compounding: $2.0B of 9.250% notes in May 2025 (claim 4824), $1.75B at 9% two months later after a $250M upsize (claims 4698, 4697), then $2.59B of converts in December (claim 3980), then an $8.5B DDTL in March 2026 (claim 11218) and a further $3.1B and $2.6B in May and August (claims 9116, 8729).

The mix reads as deliberate rather than opportunistic, and the pricing evidence supports that. Moving from high-coupon unsecured paper to contract-backed infrastructure debt rated A3/A(low) (claim 11219), and then to the first publicly syndicated HPC infrastructure financing (claim 9117), is a financing engineer's progression, not patching. But the structure has bifurcated: DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims 8733, 8732), well outside the investment-grade tranche. The same disclosure notes a roughly five-year facility maturity against customer contracts averaging approximately three years (claims 8730, 8731). That mismatch requires explanation and is the single most load-bearing fact in the record.

Use of proceeds has been stated consistently as customer-contract capex and GPU servers (claim 2138), and the demand-side anchors are large and specific — $6.5B committed from OpenAI (claim 4297), $21B from Meta (claim 2124), backlog near $104B (claim 962). The record does not, however, permit independent verification that drawn capital matched stated purpose; that gap should be treated as unverified rather than contradicted. Separately, January 2026 brought postponement of debt service coverage and contract realization ratio testing and an amendment to the minimum liquidity covenant (claims 3912, 3913, 3911) — revised terms, disclosed, worth tracking forward.

Watch the covenant amendments for recurrence, the ratings spread between infrastructure and corporate tranches, and whether Q2's $640M interest expense against a -25% net margin and ~6x debt-to-equity (claims 7341, 10221, 3058) begins converting backlog into cash. No insider sales appear on this record.

The cadence here is not periodic but continuous. From the $40 IPO in March 2025 (claim 11616) and its modest $1.4B of net proceeds (claim 4967), the company moved within eight weeks to a $1.5B revolver expansion (claim 11923), then to $2.0B of 9.250% senior notes (claim 4824), and has not paused since — roughly $28B of commitments in the twelve months to March 2026 (claim 11221) and more than $30B year-to-date by August 2026 (claim 8735). Straight equity is a small share of that: the IPO, $350M of strategic customer shares (claim 4514), NVIDIA's $2B (claim 3891), and Jane Street's $1B (claim 2080). The dilution that matters is deferred, sitting in roughly $8.2B of converts — $2.5875B in December 2025 (claim 3980), $2.6B in February 2026 (claim 11765), $3.0B due 2032 in April 2026 (claim 2126) — partially offset by $340M of capped calls (claim 3986).

The mix reads as strategy rather than patching. Stated use of proceeds is consistently customer-contract capex and GPU servers (claims 2138, 1320), and the contracted demand behind it is disclosed: OpenAI at $6.5B (claim 4297), Meta at $21B (claim 2124), backlog near $104B (claim 962). Structure has tracked that logic, culminating in the $8.5B DDTL 4.0 carrying A3/A(low) ratings (claims 11218, 11219).

Two items require explanation. The pricing path is not monotonic: DDTL 5.5 in August 2026 closed at Ba2/BB+ and SOFR+5.50% (claims 8732, 8733), a materially wider marginal cost than the investment-grade tranche five months earlier. And the same disclosure pairs a roughly five-year facility maturity against customer contracts averaging about three years (claims 8730, 8731) — a tail the contracts do not currently cover. Separately, DSCR and contract-realization testing were postponed and the minimum liquidity covenant amended in January 2026 (claims 3912, 3913, 3911), before DSCR covenants reappeared in DDTL 4.0 and 5.0 (claims 2141, 1324).

Watch whether the IG execution repeats or the high-yield spread becomes the run rate; whether postponed tests resume and are met; and whether backlog converts to cash faster than $640M of quarterly interest (claim 7341) against debt-to-equity near six (claim 3058). No insider transaction data appears on this record, so no read on insider behavior is available.

The cadence here is not episodic; it is industrialized. From the $40 IPO in March 2025 (claim #11616), which netted only about $1.4B (claim #4967) against $17.2B already raised across debt and equity (claim #4968), CoreWeave has closed a financing roughly every four to six weeks, with ticket sizes ratcheting from the $1.5–2.0B range in mid-2025 (claims #11917, #4658) to the $8.5B DDTL 4.0 in March 2026 (claim #2137), $28B of commitments in twelve months (claim #11221), and more than $30B year-to-date by August (claim #8735). Equity is by far the smallest component and has been strategic rather than market-facing — NVIDIA's $2B (claim #3891) and Jane Street's $1B (claim #2080). There is no ATM program and no insider transaction on this record, so no read on insider behavior is available; that absence should be treated as missing data, not as evidence.

The mix reads as architecture rather than opportunistic patching. Contract-backed delayed-draw term loans fund GPU and customer-contract capex (claim #2138), unsecured notes carry corporate risk (claims #4824, #8927), and convertibles paired with capped calls — the $2.5875B December 2025 issue alongside $340M of capped calls (claims #3980, #3986), then $3.0B due 2032 (claim #2126) — buy lower coupons while deferring dilution. Pricing improved from 9.250% (claim #4824) to investment-grade A3/A(low) on DDTL 4.0 (claim #11219), then reverted to Ba2/BB+ at SOFR+5.50% on the August DDTL 5.5 (claims #8732, #8733). That tiering is the structure working as designed, but it also marks where collateral quality thins.

Use of proceeds tracks the stated purpose consistently across facilities (claims #2138, #1320), with $104B of backlog (claim #1493), Meta's $21B commitment (claim #2124) and OpenAI's $6.5B (claim #4297) as the underlying collateral narrative. The item requiring explanation is January 2026: a minimum liquidity covenant amendment plus postponement of both debt-service-coverage and contract-realization testing (claims #3911–#3913), taken immediately before the largest facilities closed and not revisited on this record since.

Watch three things: whether roughly five-year facility maturities against roughly three-year customer contracts (claims #8730, #8731) force re-contracting or refinancing; whether $640M of quarterly interest (claim #7341) against $2.58B of revenue (claim #8177) and 6x debt-to-equity (claim #7345) compresses further; and whether the next equity tranche comes from strategics or from the open market.

The cadence here is not periodic, it is continuous. From a $1.4B IPO net raise in March 2025 (claim 4967) the company reached a self-reported $17.2B of cumulative debt and equity by May 2025 (claim 4968), $28B of commitments in a trailing twelve months by March 2026 (claim 11221), and more than $30B year-to-date by August 2026 (claim 8735). Intervals between capital events compress rather than lengthen, and the absolute size of each event grows: $2B of 9.250% notes in May 2025 (claim 11917) against an $8.5B delayed-draw facility in March 2026 (claim 11218) and a $3.5B unsecured offering in June 2026 (claim 8927).

The mix reads as designed rather than patched. Unsecured high-coupon paper front-loaded the build, convertibles absorbed the middle (the $2.5875B issue with $340M of capped calls, claims 3980 and 3986, then $3.0B due 2032 with a $450M option, claims 2126 and 2127), and contract-backed delayed-draw facilities carried the later, heavier capex — DDTL 3.0 through 5.5, culminating in the first publicly syndicated HPC infrastructure-backed financing (claim 9117). Strategic equity from NVIDIA (claim 3891) and Jane Street (claim 2080) arrived without an ATM. Pricing did not improve monotonically: DDTL 4.0 carried A3/A(low) ratings (claim 11219) while DDTL 5.5 four months later priced at SOFR+5.50% with Ba2/BB+ (claims 8733, 8732), which indicates collateral-tiering, not a uniform credit re-rating.

Stated use of proceeds has tracked deployment consistently — customer-contract capex and GPU servers (claims 2138, 1320) — and the $6.5B OpenAI and $21B Meta commitments (claims 4297, 2124) plus a $104B backlog (claim 1493) supply the contractual counterpart.

Three items warrant attention. The January 2026 amendments postponing debt service coverage and contract realization ratio testing and revising minimum liquidity (claims 3911–3913) were followed by DSCR covenants reappearing in DDTL 4.0 and 5.0 (claims 2141, 1324); whether those now bind is the question. Second, underlying customer contracts average roughly three years against a five-year facility maturity (claims 8731, 8730) — a tenor gap requiring explanation. Third, $640M of quarterly interest expense against $2.58B of revenue and a 6x debt-to-equity ratio (claims 7341, 8177, 3058), with current and quick ratios below one (claims 7379, 7380). No insider transactions appear on this record, so nothing can be said there.

The defining feature here is not dilution but its substitution. Common-equity issuance is comparatively restrained across eighteen months — $1.4B of IPO net proceeds (claim 4967), $350M in strategic customer shares (claim 4514), NVIDIA's $2B Class A investment (claim 3891) and Jane Street's $1B (claim 10305) — against a debt and hybrid stack that the company itself frames as $28B of commitments in twelve months (claim 11221) and more than $30B year-to-date by August (claim 8735). The counting basis shifts between those disclosures and the >$20B figure of May (claim 9123), so the totals are not additive and should not be read as such.

The mix reads as sequenced rather than opportunistic. Unsecured coupons of 9.250% (claim 4824) and 9% (claim 4697) in mid-2025 gave way to converts once the equity had a print: $2.5875B in December (claim 3980), $2.6B in February (claim 11765), $3.0B due 2032 in April (claim 2126). That is a deliberate migration down the cost curve, and the contract-then-finance rhythm — OpenAI's $6.5B (claim 4297), Meta's $21B (claim 2124), each trailed by a delayed-draw facility sized to it (claims 11218, 1319, 8729) — is coherent asset-liability matching. Stated use of proceeds (claims 2138, 1320) is consistently capex and GPU servers, and the $104B backlog (claim 1493) is directionally consistent with deployment, though the record here does not permit tracing dollars to assets.

Three items require explanation rather than alarm. The $340M capped call accompanying the December converts (claim 3986) has no disclosed equivalent for the February or April issues. The A3/A(low) rating on DDTL 4.0 (claim 11219) did not carry to DDTL 5.5, priced at SOFR+5.50% and rated Ba2/BB+ (claims 8733, 8732); the marginal facility is now worse paper. And the underlying contracts average roughly three years against a five-year maturity (claims 8731, 8730) — a refinancing tail that must be re-contracted.

Watch the January covenant actions: minimum liquidity amended (claim 3911), debt service coverage and contract realization testing both postponed (claims 3912, 3913), against 6x debt-to-equity (claim 3058) and $640M of quarterly interest expense (claim 7341). No insider transaction data appears in this record, so nothing can be said about it.

The cadence is continuous rather than episodic, and it steepens. From a $40 IPO (claim 11601) yielding $1.4B net (claim 4967) against $17.2B cumulative capital by May 2025 (claim 4968), the company reached $28B of commitments in a trailing twelve months by March 2026 (claim 11221), more than $20B year-to-date by May (claim 9123) and more than $30B by August (claim 8735). Financing events now cluster at roughly monthly intervals, which is a structural feature of the model, not a sign of distress by itself.

The mix reads as deliberate, with one seam worth watching. The early post-IPO reliance on expensive unsecured paper — the 9.250% notes due 2030 at $2.0B (claim 11917) and the 9% $1.75B issue (claim 4697) — has been progressively displaced by contract-backed delayed draw facilities: DDTL 3.0 at $2.6B (claim 11922), DDTL 4.0 at $8.5B carrying A3/A(low) ratings (claims 11218, 11219), DDTL 5.0 at $3.1B as the first publicly syndicated HPC infrastructure financing (claims 9116, 9117). That is a genuine cost-of-capital migration. The seam is DDTL 5.5, $2.6B rated Ba2/BB+ at SOFR+5.50% (claims 8729, 8732, 8733) against roughly three-year underlying contracts (claim 8731) — a materially weaker tranche than DDTL 4.0, which requires explanation as to collateral and contract quality rather than assumption of uniform improvement. Equity has been used sparingly and strategically: NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 2080), and converts paired with $340M of capped calls (claim 3986), indicating dilution is being actively managed rather than absorbed. No insider transactions and no ATM appear on this record; that absence is itself informative but should not be over-read.

Use of proceeds tracks stated purpose reasonably well — customer contract capex and GPU servers (claims 2138, 1320) against a $104B backlog (claim 962) and named commitments from OpenAI at $6.5B (claim 4297) and Meta at $21B (claim 2124).

What merits attention next: the postponed contract realization ratio and debt service coverage testing (claims 3913, 3912) and the amended minimum liquidity covenant (claim 3911), which are delayed rather than resolved; the DDTL 5.0 draw deadline (claim 1321); and $640M of quarterly interest expense (claim 7341) against 6x debt-to-equity (claim 3058) and sub-1 current and quick ratios (claims 7379, 7380). The ratings trajectory across successive DDTLs is the cleanest single tell.

The defining feature of this record is not any single raise but the absence of pauses between them. From the March 2025 IPO at $40 (claim 11616) yielding roughly $1.4B net (claim 4967), the company reached $17.2B of cumulative debt and equity by mid-May 2025 (claim 4968), then $28B of commitments in the trailing twelve months by March 2026 (claim 11221), then more than $30B year-to-date by August 2026 (claim 8735). Financing is functionally a continuous operating process, not a periodic event.

The mix is more coherent than opportunistic, and the trajectory has direction. Early 2025 relied on expensive unsecured corporate credit — the $2.0B 9.250% notes due 2030 (claim 11917) and the $1.75B 9% notes (claim 4697), both of which were upsized from announced $1.5B (claims 11921, 4700; increase of $250M per claim 4698), indicating demand rather than distress. From late 2025 the structure migrates toward contract-backed delayed-draw term loans — DDTL 4.0 at $8.5B carrying A3/A(low) ratings (claims 2137, 11219), then DDTL 5.0 at $3.1B described as the first publicly syndicated HPC infrastructure-backed financing (claims 1319, 9117). That migration lowers cost by pledging customer contracts rather than the enterprise. Equity has been used sparingly and strategically — NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 10305), a $350M strategic customer issuance (claim 4514) — with no ATM and no insider sales on record. Convertibles, including $2.5875B in December 2025 with a $340M capped call (claims 3980, 3986), sit between.

Stated use of proceeds is consistently customer-contract capex and GPU servers (claims 2138, 1320) and nothing in the record contradicts that, though the data does not permit tracing actual deployment against it.

Three items require explanation rather than alarm. The January 2026 postponements of debt service coverage and contract realization ratio testing, and the minimum liquidity amendment (claims 3912, 3913, 3911), are covenant relief obtained before the largest facilities closed. The August 2026 facility priced at SOFR+5.50% with Ba2/BB+ ratings (claims 8733, 8732) is materially weaker than March's investment-grade tranche. And underlying contracts average roughly three years (claim 8731) against facility maturities near five to five-and-a-half (claims 8730, 9120) — a tenor gap that must be closed by renewal. Watch DDTL draw pace against the availability deadline (claim 1321), the 1.x DSCR covenant (claim 1324) against $640M of quarterly interest expense (claim 7341), and conversion of the $104B backlog (claim 962).

The shape here is a company that treated its IPO as an opening deposit rather than a funding event. The $40 pricing (claim 15974) netted roughly $1.4B (claim 4967), and within six weeks the balance sheet had moved to a $1.5B revolver expansion and a $350M letters-of-credit facility (claims 11923, 11924). By 2025-05-13 the cumulative debt-and-equity figure was already $17.2B (claim 4968), against $28B of commitments in the trailing twelve months as of 2026-03-30 (claim 11221) and more than $30B year-to-date by 2026-08-09 (claim 8735). Cadence is not episodic; it is continuous, and it accelerates.

The mix reads as coherent rather than opportunistic, but coherent in a specific and narrow way. Common-equity dilution is small and almost entirely strategic: the $350M customer share issuance (claim 4514), NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 2080). The heavy lifting is contract-collateralized term debt — DDTL 3.0 at $2.6B (claim 11922), DDTL 4.0 at $8.5B (claim 11218), DDTL 5.0 at $3.1B (claim 9116), DDTL 5.5 at $2.6B (claim 8729) — each explicitly tied to customer-contract capex and GPU servers (claims 2138, 1320). Stated use of proceeds tracks deployment as far as the record permits, and the counterparty commitments are visible against it: OpenAI at $6.5B (claim 4297), Meta at $21B (claim 2124).

Cost of capital moved in two directions at once. The 9.250% 2030 notes (claim 4824) and 9% 2031 notes (claim 4697) sit alongside an A3/A(low) rating on the DDTL 4.0 structure (claim 11219), while DDTL 5.5 priced at SOFR+5.50% with Ba2/BB+ (claims 8733, 8732). The convertible stack — $2.5875B (claim 3980), $2.6B (claim 11765), $3.0B due 2032 (claim 2126) — carries a capped call only on the first tranche, at $340M (claim 3986).

Three items require explanation rather than alarm: the contract realization ratio testing postponement, the debt service coverage ratio testing postponement, and the minimum liquidity covenant amendment, all dated 2026-01-01 (claims 3913, 3912, 3911). Against $640M of quarterly interest expense (claim 7341) and debt-to-equity near 6 (claim 3058), watch whether those tests resume on schedule and whether the DDTL 5.0 draw deadline (claim 1321) is met.

No insider transactions appear on this record, and the claim list is truncated; that absence should not be read as a finding.

The cadence compresses rather than stabilizes. The March 2025 IPO at $40 (claim 11601) returned roughly $1.4B net (claim 4967) against $17.2B of total debt and equity already raised by May 2025 (claim 4968); that became $28B of commitments in twelve months by March 2026 (claim 11221) and more than $30B year-to-date by August 2026 (claim 8735). Equity is the minority instrument: the $350M strategic customer issuance (claim 4514), NVIDIA's $2B Class A investment (claims 3891, 9316), Jane Street's $1B (claim 2080), and two convertible tranches — $2.5875B in December 2025 alongside $340M of capped calls (claims 3980, 3986) and $2.6B in February 2026 (claim 11765), with a further $3.0B due 2032 plus a $450M purchaser option proposed in April 2026 (claims 10325, 10326). The capped calls defer dilution rather than remove it, so the dilution trajectory is back-loaded toward 2030-2032 maturities.

The mix is coherent at the structural level and opportunistic at the edges. The spine is a repeating delayed-draw term loan program secured by contracted infrastructure — DDTL 3.0 (claim 11922), 4.0 at $8.5B (claim 11218), 5.0 at $3.1B (claim 9116), 5.5 at $2.6B (claim 8729) — and its pricing record is the most informative series here. Unsecured paper at 9.250% in May 2025 (claim 4824) gave way to A3/A(low) ratings on DDTL 4.0 (claim 11219) and a first public syndication (claim 9117), yet DDTL 5.5 priced at SOFR+5.50% rated Ba2/BB+ (claims 8732, 8733). Cost of capital improved and then widened; the collateral and structural differences behind that reversal requires explanation.

Stated use of proceeds tracks deployment closely — customer contract capex and GPU servers (claims 2138, 1320) — and the offtake is real: OpenAI $6.5B (claim 4297), Meta $21B (claim 2124), backlog near $104B (claim 962). The unreconciled item is tenor: DDTL 5.5 runs about five years against contracts averaging about three (claims 8730, 8731).

Watch the January 2026 covenant package, where minimum liquidity was amended (claim 3911) and contract realization and debt service coverage testing were postponed (claims 3913, 3912) — postponed, not resolved — against $640M quarterly interest expense (claim 7341), roughly 6x debt-to-equity (claim 8202), and a sub-1 current ratio (claim 8203). Also note the April 8 record carrying both a $2.6B and a $3.0B convertible size (claims 10327, 10325); one was revised. No insider transactions appear on record, so that dimension cannot be assessed.

The cadence here is not episodic but continuous. From the $40 IPO in March 2025 (claim 15974), which netted only $1.4B (claim 4967), the company has gone to market roughly every six to eight weeks without a meaningful pause, and its own running totals describe the trajectory better than any single raise: $17.2B cumulative by May 2025 (claim 4968), $28B of commitments in the trailing twelve months by March 2026 (claim 11221), and more than $30B year-to-date by August 2026 (claim 8735).

The mix reads as deliberate rather than opportunistic, with one qualification. Common-equity dilution has been narrow and strategic — NVIDIA's $2B (claim 3891), Jane Street's $1B (claim 10305), $350M in customer shares (claim 4514) — while the actual dilution vector is convertible: $2.5875B in December 2025 with $340M of capped calls (claims 3980, 3986), $2.6B in February 2026 (claim 11765), $3.0B plus a $450M option in April 2026 (claims 2126, 2127). Three converts in five months, each larger than the last. The secured side shows genuine engineering: unsecured coupons of 9.25% and 9% in mid-2025 (claims 4824, 4697) gave way to the A3/A(low) contract-backed DDTL 4.0 at $8.5B (claims 11218, 11219) and the first publicly syndicated HPC-backed financing (claim 9117). But DDTL 5.5, at SOFR+5.50% and rated Ba2/BB+ (claims 8733, 8732), sits well below that — the stack is bifurcating, and marginal cost is not monotonically improving.

Stated use of proceeds — customer contract capex and GPU servers (claims 2138, 1320) — tracks the disclosed backlog of $104B (claim 962) and named commitments from OpenAI and Meta (claims 4297, 2124). The recurring "expected closing" tags are settlement mechanics, not slippage. What requires explanation is January 2026: minimum liquidity amended (claim 3911), DSCR testing postponed (claim 3912), contract realization ratio testing postponed (claim 3913) — covenant relief obtained two months before the investment-grade facility priced.

Watch the tenor mismatch: contracts average roughly three years (claim 8731) against maturities of five to 5.5 years and a 2032 DDTL 4.0 (claims 8730, 9120, 11222). Watch whether the postponed tests resume on schedule, and watch $640M of quarterly interest against $2.58B of revenue at 6x debt-to-equity (claims 7341, 8177, 3058). No insider sales appear on this record at all, which is itself worth noting. The Core Scientific $270M termination fee (claim 4739) is no longer discussed.

The defining feature here is not dilution but leverage. Equity issuance since the March 2025 IPO at $40 (claim 11601) has been episodic and strategic rather than continuous — $1.4B of IPO net proceeds (claim 4967), a $350M strategic customer share issuance (claim 4514), NVIDIA's $2B Class A investment (claim 3891), and Jane Street's $1B (claim 2080). There is no ATM program on record and no insider transaction data at all, which is a real gap in this record and limits any read on management conviction. What has scaled is debt, and the cadence is close to monthly: from $17.2B total capital by May 2025 (claim 4968) to $28B of commitments in twelve months (claim 11221) to more than $30B year-to-date by August 2026 (claim 8735).

The mix reads as coherent rather than opportunistic, and the direction of travel is the argument. Early financings were expensive unsecured paper — 9.250% notes due 2030 (claim 11917) and 9% notes at $1.75B (claim 4658). The structure then migrated toward contract-backed infrastructure facilities, culminating in the $8.5B DDTL 4.0 carrying A3/A(low) ratings (claims 2137, 11219) and the first publicly syndicated HPC infrastructure-backed financing at $3.1B (claims 9116, 9117). Convertibles were paired with capped calls — $340M against the $2.5875B issue (claims 3986, 3980) — which is the behavior of an issuer managing dilution deliberately, not patching.

Two things require explanation. The August 2026 DDTL 5.5 closed at SOFR+5.50% and rated Ba2/BB+ (claims 8729, 8733, 8732), materially wider and several notches below the March facility, suggesting the investment-grade execution is not generalizable across the collateral pool. Separately, underlying customer contracts average roughly three years (claim 8731) against facility maturities near five (claim 8730) and DDTL 4.0 running to March 2032 (claim 11222).

Use of proceeds tracks stated purpose — customer contract capex and GPU servers (claims 2138, 1320) — and is corroborated by the $104B backlog (claim 1493) and named commitments from OpenAI and Meta (claims 4297, 2124). The postponement of debt service coverage and contract realization ratio testing, plus the minimum liquidity amendment (claims 3912, 3913, 3911), is the item to watch: whether testing resumes on schedule against covenants now embedded in DDTL 4.0 and 5.0 (claims 2141, 1324), alongside $640M of quarterly interest expense (claim 7341) at a 6x debt-to-equity ratio (claim 3058).

CoreWeave's record, as provided, is mostly debt, and the dilution story is secondary to the leverage story. The list is also truncated mid-entry and contains no insider transactions, so this read is limited to what appears. Equity issuance is episodic rather than continuous. After the $40 IPO (claim 11601) and roughly $1.4B in net proceeds (claim 4967), new shares arrived mainly through strategic placements: a $350M customer issuance (claim 4514), NVIDIA's $2B investment (claim 3891) and Jane Street's $1B (claim 10305). The more material dilution path runs through convertibles. The $2.5875B offering (claim 3980) carries conversion terms referencing roughly 30.0M shares (claim 3984), partly offset by a $340M capped call (claim 3986). A further $3.0B convertible due 2032 with a $450M option (claims 10325, 10326) followed. The mix reads as a deliberate structure rather than patching. Contract-backed delayed-draw facilities fund GPU capex, and unsecured notes and converts fill the corporate layer. The progression of DDTL 3.0 at $2.6B (claim 11922), DDTL 4.0 at $8.5B carrying A3/A(low) ratings (claims 11218, 11219) and DDTL 5.0 at $3.1B (claim 9116) is consistent with that design. Two items require explanation. First, the January 2026 postponement of debt service coverage and contract realization testing, alongside a minimum liquidity amendment (claims 3912, 3913, 3911), came between the first convertible and the largest DDTL. Second, the later $2.6B facility is rated Ba2/BB+ and priced at SOFR + 5.50% (claims 8732, 8733). It also pairs contracts averaging about 3 years with an approximately 5-year maturity (claims 8731, 8730), a step away from the investment-grade terms of DDTL 4.0. Whether announced use of proceeds tracks deployment cannot be assessed from this record. Use-of-proceeds claims exist (claims 11920, 1320, 2138), but their substance is not included, and no deployment figures are shown. The operating context is Q2 2026 revenue of $2.58B (claim 8183), interest expense of $640M (claim 7341), a -25% net margin (claim 10221) and debt-to-equity near 6 (claim 7345). A sophisticated reader should watch four things: whether deferred covenant tests resume on schedule, whether new facility pricing continues to drift from the DDTL 4.0 terms, and how the 2032 convertibles are settled. The fourth is reconciliation of backlog, cited as $104 (claim 962) and as $100 billion (claim 647), against contract duration assumptions embedded in the facilities.

CoreWeave's record since the $40 IPO (claim 11601) shows a company that has financed itself mostly with debt, with dilution arriving in structured forms rather than through follow-on common offerings. The IPO's roughly $1.4B in net proceeds (claim 4967) is small next to what came after: more than $20B raised year-to-date by May 2026 (claim 9123) and more than $30B by August (claim 8735). Direct equity has come through negotiated placements: a $350M strategic customer issuance (claim 4514), NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 10305). Convertibles carry the rest of the equity-linked exposure. These are the $2.59B December 2025 notes with a $340M capped call (claims 3980, 3986) and the $3.0B 2032 notes with a $450M option (claims 10325, 10326). The earlier $2.6B convertible figure (claim 11765) appears to have been revised upward before pricing. No insider transactions appear in the record, so no conclusion about insider positioning is possible.

The mix reads as a deliberate architecture. Contract-backed delayed-draw term loans fund specific GPU deployments, while unsecured notes and converts support the corporate layer. The high-yield notes were repeatedly upsized: $1.5B to $2B at 9.25% (claims 11921, 11917) and $1.5B to $1.75B at 9% (claims 4700, 4697). That suggests demand was ample, but the pricing was expensive. Quality across the DDTL series is uneven. DDTL 4.0 earned A3/A(low) ratings (claim 11219), but DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims 8733, 8732). Its roughly five-year maturity sits against underlying contracts averaging about three years (claims 8730, 8731), and that gap requires explanation.

On use of proceeds, facilities are described as funding customer-contract capex and GPU servers (claims 2138, 1320). However, the record contains no draw-level or deployment disclosure, so whether actual spending matches the stated purpose cannot be tested from this data. The January 2026 amendments postponed debt-service-coverage and contract-realization testing and changed the minimum liquidity covenant (claims 3911–3913). They indicate that the original covenant schedule was delayed.

A careful reader should watch four things. The first is Q2 interest expense of $640M (claim 7341) against $2.58B of revenue (claim 8183), with a debt-to-equity ratio near 6 (claim 3058). The second is when the postponed covenant tests resume. The third is how quickly the $104B backlog (claim 962) turns into cash flow capable of servicing a stack that is increasingly priced below investment grade. The fourth is the share count once the convertibles move into the money.

The record describes a company financing itself almost entirely through debt since its IPO. Equity dilution has come in discrete blocks rather than a steady drip. After the $40 IPO (claim 11601) and roughly $1.4B in net proceeds (claim 4967), the equity-linked events on file are a strategic customer share issuance (claim 4514), NVIDIA's $2B investment (claim 3891), Jane Street's $1B investment (claim 10305), and two convertible offerings: $2.5875B in late 2025 (claim 3980), and a proposed $3.0B due 2032 with a $450M option (claims 10325, 10326). Most of the eventual dilution therefore sits in conversion terms rather than in shares already issued. The $340M capped call (claim 3986) shows the company tried to limit that exposure. The record contains no insider transactions, so nothing can be said about insider behavior.

The mix looks deliberate rather than improvised. Asset-backed delayed-draw facilities have grown from $2.6B (claim 11922) to $8.5B for DDTL 4.0 (claim 11218), which carried A3/A(low) ratings (claim 11219). The company says it has raised more than $30B year-to-date (claim 8735). Two things complicate that picture. In January 2026 the company amended its minimum liquidity covenant and postponed debt service coverage and contract realization testing (claims 3911–3913). More recently, DDTL 5.5 priced at SOFR + 5.50% with Ba2/BB+ ratings (claims 8733, 8732), well below DDTL 4.0's grade. The facilities appear to be rated according to which customer contracts back them, and investors should read each one individually rather than as a uniform stack.

The stated use of proceeds, customer-contract capex and GPU servers (claim 2138), is consistent with the facility structure. However, the record contains no deployment data, so actual spending cannot be checked against the announced purpose. Figures for the 2026 convertible also require explanation: $2.6B appears in claims 11765 and 10327, while $3.0B appears in claims 10325 and 2126. Several claims also report amounts in inconsistent units.

Readers should watch four things. The first is when the postponed coverage tests resume and whether they pass. The second is the mismatch between customer contracts averaging about 3 years (claim 8731) and a facility maturity of about 5 years (claim 8730). The third is how Q2 interest expense of $640M (claim 7341) compares with quarterly revenue of $2.58B (claim 8177). The fourth is whether the reported backlog of roughly $104B (claim 962) turns into cash on schedule to service a debt-to-equity ratio near 6 (claim 3058).

CoreWeave's record since the March 2025 IPO at $40 (claim 11616) is mostly a debt story. Equity has been a small share of capital raised. The IPO netted $1.4B (claim 4967), NVIDIA invested $2B (claim 3891), and Jane Street invested $1B (claim 10305). Against that, the company reports more than $30B of debt and equity secured in 2026 alone (claim 8735). Common-share dilution has therefore been modest in cadence. Latent dilution sits mainly in the convertibles: the $2.59B December 2025 issue, partly hedged by a $340M capped call (claims 3980, 3986), and the proposed $3.0B 2032 notes with a $450M upsize option (claims 10325, 10326).

The mix reads as a deliberate two-track structure rather than patching. Contract-backed delayed-draw facilities fund GPU deployment tied to named customer commitments (claims 2138, 1320). Unsecured notes and converts sit at the corporate level. The DDTL sequence grew from $2.6B (claim 11922) to $8.5B at A3/A(low) (claims 11218, 11219). Several notes deals were upsized, including the 2030 notes from $1.5B to $2.0B (claims 11921, 11917), which indicates available demand. Two developments temper that reading. The January 2026 amendments postponed debt-service-coverage and contract-realization testing and revised the minimum liquidity covenant (claims 3912, 3913, 3911). The later DDTL 5.5 also priced at SOFR + 5.50% with Ba2/BB+ ratings (claims 8733, 8732), below the investment-grade tier achieved months earlier.

Stated use of proceeds for the facilities is specific: customer-contract capex and GPU servers. However, the record contains no deployment reconciliation, so whether proceeds tracked announcements cannot be assessed from this data. The operating figures show the load. Q2 2026 revenue was $2.58B (claim 8183), interest expense was about $640M (claim 7341), and debt-to-equity was near 6 (claim 8202). Net margin is reported as -25% (claim 7343), although claim 8200 records it unsigned and requires reconciliation.

A careful reader should watch four things. The first is whether the postponed covenant tests resume on schedule. The second is the mismatch between roughly three-year underlying contracts and roughly five-year facility maturities (claims 8731, 8730). The third is the concentration of maturities across 2030–2032. The fourth is whether the approximately $100B backlog (claim 647) converts fast enough to service the added interest burden. The record includes no insider transactions, and the claim list appears truncated, so this read is incomplete on both points.

CoreWeave's capital record since its IPO at $40 per share (claim 11616) is dominated by debt rather than equity. The IPO itself was modest in net terms at $1.4B (claim 4967). What followed was a steady run of secured and unsecured borrowing, and the pace has increased. The company described $28B of equity and debt commitments over twelve months (claim 11221), then more than $30B secured year-to-date by August 2026 (claim 8735).

Direct dilution has been relatively contained. It came through three channels. The first was strategic equity placements: the $350M customer share issuance (claim 4514), NVIDIA's $2B investment (claim 3891) and Jane Street's $1B (claim 10305). The second was the convertibles. The $2.5875B December 2025 offering (claim 3980) references roughly 30.0M underlying shares (claim 3984) and was partly hedged by $340M of capped calls (claim 3986). The $3.0B 2032 convertible (claim 10325) carries a further $450M option (claim 10326). Taken together, the equity side reads as sponsor-aligned placement rather than open-market issuance. No insider transactions appear in the record, so that dimension cannot be assessed.

The mix follows a coherent template. Contract-backed delayed-draw facilities are layered with high-coupon unsecured notes: 9.250% on the $2B 2030 notes (claims 11917, 4824) and 9% on the upsized $1.75B issue (claims 4697, 4698). Within the DDTL series, the terms have diverged. DDTL 4.0 at $8.5B carried A3/A(low) ratings (claims 11218, 11219). DDTL 5.5 at $2.6B was rated Ba2/BB+ and priced at SOFR+5.50% (claims 8729, 8732, 8733). That later facility also runs about five years (claim 8730) against underlying contracts averaging about three (claim 8731). That tenor mismatch requires explanation.

Use-of-proceeds statements exist for the notes (claim 11920), DDTL 5.0 (claim 1320) and customer-contract capex (claim 2138). However, the record contains no deployment data against which to test them.

The covenant amendments of January 2026 also require explanation. These postponed debt-service-coverage and contract-realization testing and amended minimum liquidity (claims 3911–3913). They sit beside Q2 2026 interest expense of $640M (claim 7341), revenue of about $2.58B (claim 8177) and a negative 25% net margin (claim 7343).

A reader should watch four things next:

- whether postponed covenant tests resume on schedule; - how the $104B backlog (claim 1493) converts to cash relative to amortization of the shorter-contract facilities; - whether future DDTL tranches price closer to DDTL 4.0 or DDTL 5.5; - any conversion activity on the two convertible series.

CoreWeave's capital history since its March 2025 IPO at $40 per share (claim 11616) is dominated by debt. Equity has been a minority of the capital raised. The IPO netted roughly $1.4B (claim 4967). Direct equity since then has come mainly from strategic placements: a $350M customer share issuance (claim 4514), NVIDIA's $2B investment (claim 3891) and Jane Street's $1B (claim 10305). The more meaningful source of potential dilution is convertibles. The December 2025 $2.5875B convertible (claim 3980) references roughly 30.0M conversion shares (claim 3984) and is partly offset by a $340M capped call (claim 3986). A proposed $3.0B convertible due 2032 followed in April 2026 (claim 10325). Its final size is not settled on the record, since a $2.6B figure appears alongside it (claim 10327). The record contains no insider transaction claims, so that part of the question cannot be assessed.

The debt mix reads as a deliberate structure rather than ad hoc patching. The company layers delayed-draw facilities secured against customer contracts, which the filings describe as financing contract capex and GPU servers (claim 2138), on top of unsecured high-yield notes. Those notes include the $2B 9.25% notes due 2030 (claim 11917), the upsized $1.75B 9% notes (claim 4697) and $3.5B due 2032 (claim 8927). The trajectory is not uniformly improving, though. DDTL 4.0 carried A3/A(low) ratings (claim 11219). The later $2.6B DDTL 5.5 was rated Ba2/BB+ (claim 8732) and priced at SOFR + 5.50% (claim 8733). It also pairs an approximately five-year maturity (claim 8730) with underlying contracts averaging about three years (claim 8731), a duration gap that requires explanation.

The record states use of proceeds, but it does not show how the money was deployed. The stated purposes therefore cannot be checked against actual spending. The January 2026 amendments deserve attention. They postponed debt service coverage testing (claim 3912), postponed contract realization testing (claim 3913) and revised the minimum liquidity covenant (claim 3911). Together these suggest that deployment or revenue timing lagged the original facility assumptions.

A careful reader should watch three things. The first is whether the postponed covenant tests are reinstated on schedule. The second is how Q2 2026 interest expense of $640M (claim 7341) moves against quarterly revenue of $2.58B (claim 8177) as the facilities are drawn. The third is whether the roughly $100B backlog (claim 647) converts to revenue fast enough to service the more than $30B raised year-to-date (claim 8735) before the 2030 notes approach refinancing.

CoreWeave's post-IPO capital history is overwhelmingly debt-led rather than equity-led, so share-count dilution has so far been modest while balance-sheet leverage has grown quickly. After an IPO at $40 (claim 11601) netting roughly $1.4B (claim 4967), the company layered secured delayed-draw facilities, high-coupon senior notes, and convertibles at close to monthly intervals. Cumulative figures rise from $17.2B raised (claim 4968) to $28B committed over twelve months (claim 11221) and more than $30B year-to-date by August 2026 (claim 8735). The direct equity that did arrive was strategic and placed privately: NVIDIA's $2B (claim 3891) and Jane Street's $1B (claim 10305). Prospective dilution is concentrated in the convertibles, the $2.5875B issue (claim 3980), partly offset by $340M of capped calls (claim 3986), and the proposed $3.0B 2032 convert (claim 10325).

The mix looks coherent in structure. Contract-backed DDTL facilities fund GPU servers against named customer commitments (claims 2138, 1320), and unsecured notes and converts sit above them. The DDTL 4.0 facility drew A3 investment-grade ratings (claim 11219). The later DDTL 5.5 was rated Ba2/BB+ (claim 8732) and priced at SOFR + 5.50% (claim 8733). That difference in credit quality between the tiers is not explained in the record.

The stated use of proceeds broadly matches the facility design, but the record lacks deployment data to confirm it. The January 2026 covenant changes require explanation. They loosened the minimum liquidity covenant (claim 3911) and postponed testing of both the debt service coverage ratio and the contract realization ratio (claims 3912, 3913). Postponing a test that measures whether contracts are converting to revenue sits awkwardly beside facilities secured by those same contracts. The DDTL 5.5 also pairs a roughly five-year maturity (claim 8730) with underlying customer contracts that average about three years (claim 8731).

By Q2 2026 the servicing burden was substantial. Interest expense of $640M (claim 7341) was set against revenue of $2.58B (claim 8177) and a negative 25% net margin (claim 7343). Backlog of roughly $104B (claim 962) is the stated offset.

The record contains no insider transactions, so no conclusion about insider positioning is possible. Readers should watch four things. First, whether postponed covenant tests resume on schedule. Second, how quickly DDTL capacity is drawn relative to contract delivery. Third, whether interest expense grows faster than revenue. Fourth, how the convertibles trade relative to their conversion terms.

CoreWeave's record shows a company that has used equity sparingly and debt heavily. Common-share issuance clusters around the $40 IPO (claim #11616), which produced roughly $1.4B in net proceeds (claim #4967). Later equity arrived as strategic placements rather than broad offerings: $350M to a strategic customer (claim #4514), $2B from NVIDIA (claim #3891), and $1B from Jane Street (claim #10305). The convertibles carry the larger latent dilution. The $2.59B December 2025 issue (claim #3980) was paired with $340M of capped calls (claim #3986), which suggests the company is managing that exposure deliberately. The proposed $3.0B 2032 convertible, with a $450M upsize option (claims #10325, #10326), adds to it, and the record does not show a matching hedge. By August the company reported a debt-to-equity ratio near 6 (claim #3058).

The debt mix looks more like a strategy than a series of patches. Each delayed-draw facility is tied to contracted capex: DDTL 3.0 at $2.6B (claim #11922), DDTL 4.0 at $8.5B (claim #11218), DDTL 5.0 at $3.1B (claim #9116), and DDTL 5.5 at $2.6B (claim #8729). Unsecured notes run alongside them, beginning at 9.25% (claim #4824) and extending to $3.5B due 2032 (claim #8927). The trajectory within that ladder requires explanation, though. DDTL 4.0 carried A3/A(low) ratings (claim #11219), while DDTL 5.5 is rated Ba2/BB+ and priced at SOFR + 5.50% (claims #8732, #8733). DDTL 5.5 also runs about five years (claim #8730) against customer contracts averaging about three (claim #8731). Separately, the January 2026 amendments postponed debt service coverage and contract realization testing and revised the minimum liquidity covenant (claims #3912, #3913, #3911). These changes came before the largest facilities closed.

Stated use of proceeds is consistent at the facility level: DDTL 4.0 is earmarked for customer-contract capex and GPU servers (claim #2138). The record does not let a reader trace actual deployment against those purposes. Operating results show the carrying cost: Q2 2026 interest expense of $640M (claim #7341) against revenue of $2.58B (claim #8183) and a net margin near -25% (claim #7343). No insider transactions appear in the record as provided, so no conclusion can be drawn there.

Four things are worth watching. The first is whether the postponed covenant tests resume on schedule. The second is how contracts behind DDTL 5.5 are renewed before the loan matures. The third is whether pricing on new facilities keeps widening after the A3 tranche. The fourth is how the roughly $100B backlog (claim #647) converts into revenue that can service capital raised at a pace exceeding $30B year-to-date (claim #8735).

CoreWeave's record describes a company that has financed itself almost entirely through leverage since its IPO, with equity dilution arriving mostly through strategic placements and convertibles rather than follow-on offerings. The IPO itself was modest against what followed, at $1.4B net per claim #4967. The company later cited $28B of commitments over twelve months (claim #11221) and more than $30B year-to-date by August 2026 (claim #8735). Direct dilution came in discrete blocks: the $350M strategic customer issuance (claim #4514), NVIDIA's $2B investment (claim #3891), and Jane Street's $1B (claim #10305). Latent dilution sits in the $2.59B convertible of December 2025 (claim #3980) and the $3.0B convertible due 2032 with a $450M option (claims #10325, #10326). The record shows a $340M capped call only on the earlier convertible (claim #3986). The $2.6B convertible cited in February and again in April (claims #11765, #10327) requires explanation, since the record does not make clear whether it is a separate issue or a rounded restatement of the December deal.

The mix reads as a deliberate architecture more than patching. Asset-backed delayed draw facilities fund contracted capacity, and unsecured notes and convertibles fill the corporate layer, with repeated upsizing suggesting available demand: $1.5B to $2B in May 2025 (claims #11921, #11917) and $1.5B to $1.75B in July (claims #4700, #4698). Two points qualify that reading. The January 2026 amendment postponed debt service coverage and contract realization testing and revised the minimum liquidity covenant (claims #3911 through #3913). Credit quality has also diverged across facilities. DDTL 4.0 carried A3 ratings (claim #11219), while DDTL 5.5 priced at SOFR plus 5.50% with Ba2/BB+ ratings (claims #8733, #8732). DDTL 5.5 has a roughly five-year maturity supported by customer contracts averaging about three years (claims #8730, #8731).

Stated uses of proceeds, including GPU servers tied to customer contracts (claims #2138, #1320), are consistent across filings. However, the record contains no deployment data against which to test them. No insider transactions are on record.

A careful reader should watch four things. The first is whether $640M of quarterly interest expense (claim #7341) against $2.58B of revenue (claim #8183) narrows as backlog, cited at $104B (claim #1493), converts. The second is whether the postponed covenant tests resume on schedule. The third is how the gap between contract duration and loan maturity in the newer facilities is refinanced. The fourth is whether later convertibles carry dilution hedges.

CoreWeave's record describes a company financing almost entirely through debt, with equity used sparingly and mostly placed with strategic holders. After the $40 IPO (claim 11601) and roughly $1.4B of net proceeds (claim 4967), later equity came from NVIDIA's $2B investment (claims 3891, 9316) and Jane Street's $1B (claim 10305), not from broad follow-ons. Direct share dilution therefore looks modest. The convertibles change that picture. The $2.5875B offering (claim 3980) was paired with a $340M capped call (claim 3986), a $2.6B convertible appears in February 2026 (claims 11765, 10327), and a $3.0B issue due 2032 was proposed with a $450M option (claims 10325, 10326). The record does not make clear whether the $2.6B references restate the December deal or describe a new one, and that requires explanation before potential conversion dilution can be sized. The unit behind the conversion figure in claim 3984 is also not stated.

The debt mix follows a repeatable template rather than ad hoc patching. The company raises high-yield senior notes, several of them upsized: $2.0B at 9.25% (claims 11917, 4824) after a $1.5B launch (claim 11921), and $1.75B at 9% (claims 4697, 4698). Around those notes sits a ladder of contract-backed delayed-draw facilities, from DDTL 3.0 at $2.6B (claim 11922) to DDTL 4.0 at $8.5B (claim 11218). The revolver also grew from $1.5B to $2.5B (claims 11923, 11766). Cumulative figures keep being revised upward, from $17.2B (claim 4968) to more than $30B year-to-date (claim 8735).

Use of proceeds is sparsely captured. Claim 2138 ties DDTL 4.0 to customer-contract capex and GPU servers, but the stated uses in claims 11920 and 1320 are not detailed enough to test against deployment. The observable outcome is a debt-to-equity ratio near 6 (claim 3058) and $640M of quarterly interest expense (claim 7341) against $2.58B of revenue (claim 8177). Revenue grew 112% (claim 8185), and backlog stood at $104B (claim 1493).

The record shows no insider transactions, so that dimension cannot be assessed. Several items deserve attention next. The January 2026 amendments postponed debt-service-coverage and contract-realization testing and revised the liquidity covenant (claims 3911 to 3913), so it matters when that testing resumes. DDTL 5.5 carries a roughly five-year maturity (claim 8730) against underlying contracts averaging about three years (claim 8731). It is also rated Ba2/BB+ at SOFR plus 5.50% (claims 8732, 8733), compared with A3 on DDTL 4.0 (claim 11219). Finally, several note closings are marked delayed (claims 4699, 9699).

CoreWeave's capital trajectory through mid-2026 shows a disciplined pivot from conventional post-IPO financing to infrastructure-backed debt securitization. The company raised $1.4 billion net in its March 2025 IPO at $40 per share (claims 11601, 4967), then deployed an escalating mix of convertible debt ($2.6 billion offered December 2025, claim 3980; $2.6 billion offered February 2026, claim 11765) and traditional senior notes ($1.75 billion July 2025, claim 4658; $3.5 billion June 2026, claim 8927). By August 2026, the company had secured over $30 billion in combined debt and equity capital year-to-date (claim 8735).

The shift toward infrastructure financing is strategic rather than opportunistic. In March 2026, CoreWeave closed an $8.5 billion delayed draw term loan facility rated investment-grade A3/A(low) (claims 11218, 11219), explicitly structured to finance customer contract capex and GPU servers (claim 2138). By May 2026, a commercially syndicated successor facility demonstrated that capital markets would fund the model directly—the $3.1 billion DDTL 5.0 was described as the first publicly syndicated HPC infrastructure-backed financing (claim 9117). This represents material proof-of-concept for the thesis that customer contracts, not balance-sheet quality, secure capital.

Strategic equity raises anchor the narrative credibly. NVIDIA's $2 billion investment (claim 3891) and Meta's $21 billion payment commitment (claim 2124) in early 2026 preceded the broadest debt issuances and represent customer-anchored funding. Similarly, the $104 billion sales backlog (claim 1493) and 112 percent year-over-year revenue growth in Q2 (claim 8185) track announced use of proceeds into capacity deployment.

Yet material cautions exist. Debt-to-equity reached 6x (claim 3058) by Q2 2026, while net losses widened despite revenue growth (claim 1495), driven by $640 million in quarterly interest expense (claim 7374)—roughly 25 percent of revenue. More significantly, the company postponed debt service coverage ratio testing in January 2026 (claim 3912), requiring explanation. A sophisticated reader should assess whether the infrastructure-backed facility structure can sustainably support this leverage as interest competes with reinvestment. Customer concentration risk (OpenAI $6.5 billion commitment, claim 4297; Meta $21 billion; NVIDIA) and the pace of backlog realization relative to stated timelines are critical to whether current debt service burdens become tractable or deferred again.

CoreWeave's eighteen-month capital trajectory from IPO to mid-2026 reveals a capital structure oscillating between strategic asset-backed financing and aggressive opportunistic equity raises. The company went public at $40 per share in March 2025, netting $1.4 billion, per claim 4967, then immediately entered a phase of sustained debt issuance: $1.5 billion in revolving credit facilities and $3.25 billion in senior notes between May and July 2025 (claims 5010, 4700, 4697). This early debt concentration suggests the capital markets had differentiated CoreWeave's ability to lever infrastructure commitments—which materialized explicitly after the October 2025 announcement of a $3 billion delayed draw facility (claim 4277) and, by March 2026, an $8.5 billion infrastructure-backed delayed draw closing with investment-grade ratings (claim 2137), structured against customer contracts averaging three years in duration (claim 8731).

The equity pattern complicates this narrative. NVIDIA's $2 billion investment (claim 3891) in January 2026 and Jane Street's $1 billion in April 2026 (claim 2080) occurred after the core infrastructure debt was locked, suggesting these were growth and buffer tranches rather than primary infrastructure funding. Concurrently, CoreWeave layered in convertible securities—$2.587 billion closed in December 2025 (claim 3981) and $3 billion proposed in April 2026 (claim 2126)—a structure that defers equity dilution while preserving optionality. The company disclosed $28 billion in equity and debt raised within twelve months ending March 2026 (claim 11221).

The use of proceeds aligns with stated infrastructure strategy: the $8.5 billion delayed draw was explicitly structured for customer contract capital expenditures and GPU servers (claim 2138), matching the $104 billion sales backlog (claim 1493) and customer commitments from OpenAI ($6.5 billion, claim 4297) and Meta ($21 billion, claim 2124). However, Q2 2026 results carry a discordant signal. Despite $2.58 billion in revenue representing 112 percent year-over-year growth (claims 8183, 8185), CoreWeave reported a negative net margin of negative 25 percent and return on equity of negative 48 percent (claims 7371, 7377). Interest expense of $640 million (claim 7341) consumes material cash flow. The company postponed debt service and contract realization ratio covenant testing in January 2026 (claims 3912, 3913), and the debt-to-equity ratio reached 6 (claim 7378).

A sophisticated reader should monitor whether revenue scaling outpaces interest servicing and whether the customer commitments backing infrastructure debt remain enforceable. The covenant postponements and reported current ratio approaching zero warrant scrutiny into liquidity runway relative to the facilities' $7.5 billion initial borrowing capacity (claim 11220).

CoreWeave has executed one of infrastructure cloud's most aggressive capital raises: $30 billion in debt and equity since the March 2025 IPO, per claim 8735. The financing architecture is layered: bank facilities ($1.5 billion revolving credit, $350 million letters of credit, claim 5010) were followed by unsecured senior debt starting July 2025—$1.5 billion and $1.75 billion tranches maturing 2030–2032, per claims 4700 and 4697. Convertible issuances then accelerated: $2.587 billion in December (claim 3980) and $2.6–$3 billion in April 2026 (claims 10327, 10325). Infrastructure-backed delayed draw facilities form the anchor: $2.6 billion initial closing, expanded to $8.5 billion by March 2026 (claim 11218), then $3.1 billion in second syndication (May, claim 9116). Strategic equity from NVIDIA ($2 billion, January) and Jane Street ($1 billion, April) contributed $3 billion. The mix reflects coherent strategy, not pure opportunism. DDTL facilities are contractually tied to customer commitments (three-year duration, claim 8731) from OpenAI ($6.5 billion, claim 4297) and Meta ($21 billion, claim 2124), while convertible and bank debt fund operations. Use of proceeds aligns with results: Q2 2026 revenue was $2.58 billion against $104 billion sales backlog (claims 8183, 962), confirming backlog-to-revenue conversion is occurring.

However, three material signals require scrutiny. Senior note interest payments were delayed in July 2025 (claim 4659), weeks after issuance. Debt service coverage ratio and contract realization testing were postponed in January 2026 (claims 3912, 3913)—covenant amendments within six months of facility maturity signal default risk or near-miss. Q2 2026 results show negative 25 percent net margin with $640 million quarterly interest expense (claims 7371, 7374), meaning interest consumes 26 percent of revenue. Debt-to-equity stands at 6x (claim 3162); return on equity was negative 48 percent (claim 7372). Readers should monitor whether Q3–Q4 2026 backlog conversion produces positive margins and whether postponed covenants are reinstated; reinstatement failure would signal lender conviction that solvency pressure is material.

CoreWeave has raised more than $30 billion in debt and equity since its May 2025 IPO, per claim 8735, deploying capital across infrastructure, GPU acquisition, and customer contract fulfillment. The trajectory has been relentless: within fifteen months of public markets, the company moved through an IPO ($1.4B net proceeds, claim 4967), two NVIDIA equity investments ($2B and $2B, claims 3891, 9316), a Jane Street $1B commitment (claim 10305), and announced Meta partnerships worth $21B (claim 2124)—alongside parallel debt raises including convertible notes ($2.6B and $3.0B, claims 3980, 10325), senior unsecured debt ($1.75B, $1.5B, $1.0B, $3.5B across tranches, claims 2057, 4700, 2056, 8927), and structured delayed-draw term loans rated investment-grade (DDTL 4.0 of $8.5B, DDTL 5.0 of $3.1B, DDTL 5.5 of $2.6B, claims 11218, 9116, 8729).

The capital strategy is notably coherent rather than opportunistic. Customer commitments appear to anchor the raises: the OpenAI revenue contract of $6.5B (claim 4297), Meta's $21B payment commitment (claim 2124), and a disclosed sales backlog of $104B (claim 962) provide explicit justification for the scale of both equity and debt deployment. The company has positioned debt issuance as infrastructure-specific, with stated facilities funding "customer contract capital expenditures and GPU servers" (claim 2138) and receiving separate investment-grade treatment (A3/A(low) ratings on DDTL 4.0, claim 11219). This layering—equity from marquee strategic investors, debt from structured infrastructure financing, customer commitments providing repayment visibility—reflects strategic intentionality.

However, the operational math warrants scrutiny. Q2 2026 revenue reached $2.58B with 112% year-over-year growth (claim 8177), yet the company posted a $626M net loss (claim 3014) against $640M in interest expense (claim 3015), producing a negative 25% net margin (claim 10221) and a debt-to-equity ratio of 6.0x (claim 3162). Multiple covenant tests have been postponed (claim 3913), and convertible note payment schedules flagged as delayed (claim 3982). The company is borrowing and raising capital to finance growth, not yet generating positive operating cash flow to service debt.

What follows requires close attention: actual execution against the $6.5B OpenAI contract and $21B Meta commitment, covenant compliance in refinancing cycles, and timeline to operating cash-flow positive performance. A sophisticated reader should monitor whether the backlog converts to realized revenue at claimed margins, whether debt service pressures force revised capital discipline, and whether the investment-grade rating on infrastructure debt can sustain if operating losses persist.

CoreWeave has executed a hyperscaled capital raise that tests the limits of a build-to-customer-demand model. From IPO in May 2025 through August 2026, the company secured more than $30 billion in combined debt and equity capital (per claim #8735), deploying nearly all of it into GPU infrastructure and data center capex. The financing mix is heavily debt-weighted: roughly $22 billion in various debt instruments (senior notes, convertible notes, delayed-draw term loans at multiple tranches) against $8.75 billion in equity (IPO proceeds of $1.4 billion per claim #4967, NVIDIA's $4 billion combined Class A investment per claims #3891 and #9316, Jane Street's $1 billion per claim #10305, and strategic customer share issuances). This structure is not coherent strategy but rather escalating financial engineering—each capital raise appears reactive to deployment capacity and customer demand rather than proactive capital planning. The company amended key debt covenants in January 2026 (minimum liquidity and DSCR testing postponements per claims #3911 and #3913), a material signal of early stress just eight months into its public life.

The use of proceeds aligns with stated intent but conceals an uncomfortable dependency: customer prepayments are funding capex in 2026 to the tune of $9 billion (per claim #8575), meaning CoreWeave is essentially running a pass-through model where customer cash initiates build cycles. Announced proceeds for customer contract capital expenditures and GPU servers (claim #2138) track the deployment, yet this arrangement creates undisclosed concentration: the company has $6.5 billion in committed OpenAI revenue and a $21 billion Meta payment commitment (per claims #4297 and #2124), alongside NVIDIA as both customer and investor. With Q2 2026 interest expense at $640 million against $2.58 billion in quarterly revenue and a net loss of $626 million (per claims #7341 and #3014), the company is consuming capital rather than generating it—the debt service burden is material and growing.

The $104 billion revenue backlog (per claim #1493) appears substantial until examined against a debt-to-equity ratio of 6:1 (per claim #7345) and customer contracts averaging three years in duration (per claim #8731), while the debt stack carries maturities of 5 to 5.5 years. The mismatch between contract economics and liability duration requires scrutiny. A sophisticated reader should monitor three dynamics: whether CoreWeave can deploy $30 billion of capex without further capital raises, whether customer concentration dissolves as revenue diversifies or hardens as a structural constraint, and whether negative net margins compress as fixed capex is amortized against growing revenue or remain structurally negative due to interest burden. The covenant amendments and delayed financing disclosures (per claims #4699 and #3982) suggest liquidity management is tighter than the headline backlog implies.

CoreWeave's financing sequence from May 2025 to August 2026 shows a company simultaneously scaling at extreme speed and managing the capital intensity that comes with it. The $30 billion-plus raised in that period (claim 8735) arrived through multiple instruments: straight debt dominates, but the mix includes convertibles and strategic equity—a combination that appears opportunistic rather than orchestrated.

Debt instruments comprise the bulk. An IPO in May 2025 generated $1.4 billion in net proceeds (claim 4967), followed by layered senior note issuances totaling roughly $8.5 billion through various tranches (claims 4700, 4658, 2056, 8927). In December 2025, CoreWeave issued $2.587 billion in convertible notes with a $340 million capped-call overlay (claims 3980, 3986). Delayed-draw facilities expanded repeatedly, reaching over $14 billion across multiple tranches sized to customer contracts and capex deployment (claims 2137, 9115, 8729). Equity raises were more selective: NVIDIA invested $4 billion in two tranches (claims 3891, 9316), Jane Street added $1 billion (claim 2080), and a strategic customer bought $350 million in equity (claim 4514).

The stated use of proceeds partially tracks execution. Senior notes funded general corporate purposes; the delayed-draw facilities explicitly backed customer-contract-driven capex, primarily GPU deployments (claim 2138). This directly reflects two massive consumption commitments: OpenAI's $6.5 billion revenue commitment starting September 2025 (claim 4297) and Meta's $21 billion commitment from April 2026 (claim 2124). These revenue streams justified the capex, and in turn the debt stack needed to fund it.

Yet the operating results create friction. Q2 2026 showed $2.58 billion in revenue growing 112 percent, but net losses reached $626 million with a negative 25 percent net margin (claims 8183, 7369, 3014). Interest expense hit $640 million that quarter alone (claim 7341), approaching quarterly revenue in absolute terms. The debt-to-equity ratio stands at 6x (claim 7350). In January 2026, CoreWeave postponed testing of its contract realization ratio and minimum liquidity covenants (claims 3913, 3911), signaling tighter cash flow than publicly disclosed.

A sophisticated reader should monitor whether the $100-billion-plus backlog converts to profitable scale (claim 647), watch for debt refinancing pressure in a higher-rate environment, and assess whether customer concentration in two names creates unhedged revenue dependency. The capital mix—debt-heavy, minimal IPO-base dilution, strategic checks from adjacent infrastructure players—suggests financing availability took precedence over capital structure coherence. That remains sustainable only if margin velocity arrives soon.

CoreWeave's $30 billion in debt and equity capital raised through August 2026 (claim 8735) reflects a deliberate infrastructure financing strategy increasingly shadowed by financial strain. The capital stack mixes $8.5 billion and $3.1 billion delayed-draw term facilities rated Ba2/BB+ (claims 9122, 9116, 8732), approximately $7 billion in senior notes maturing 2030–2032, $2.588 billion in convertible notes (claim 3980), and equity from NVIDIA ($4 billion total, claims 3891, 9316), Jane Street ($1 billion, claim 2080), plus $1.4 billion IPO proceeds (claim 4967). This allocation is coherent, not reactive: the company structured customer-backed asset financing to fulfill explicit multi-year commitments from OpenAI ($6.5 billion, claim 4297) and Meta ($21 billion, claim 2124), with average contract duration around three years (claim 8731). Customer prepayments of $9 billion funded capex in 2026 (claim 8575), and stated use of proceeds—customer contract capex and GPU servers (claim 2138)—tracked a $104 billion sales backlog (claim 1493).

Yet covenant stress surfaced by January 2026. The company announced postponement of debt service coverage ratio testing (claim 3912), minimum liquidity covenant amendment (claim 3911), and contract realization ratio testing postponement (claim 3913). These are not routine; lenders postpone covenant windows when near-term compliance becomes uncertain. Q2 2026 performance illuminates the pressure: $2.58 billion in quarterly revenue with 112% year-over-year growth (claims 8183, 8185), but a $626 million net loss and $640 million in interest expense (claims 3014, 7374). Interest expense consumed roughly one-quarter of quarterly revenue, leaving minimal margin for operational leverage or refinancing flexibility. Reported debt-to-equity reached 6x with negative 48% return on equity (claims 3162, 3059); current and quick ratios approached zero (claims 7379, 7380).

The structure remains sound if customer commitments hold and growth accelerates. Critical vectors to monitor: whether the Meta and OpenAI contracts—together 26% of stated backlog—maintain schedule or face revision; whether postponed covenant tests reset in late 2026 or require additional amendments; and whether leverage ratios converge toward investment-grade levels or remain elevated into refinancing windows, particularly given convertible notes maturing December 2032 (claim 3983). Customer prepayment dependency creates operational rigidity if contract timing shifts or capex requirements compress.

CoreWeave's capital structure reflects a singular thesis: leverage customer-backed contracts to secure infrastructure financing. Through mid-2026, the company raised over $30 billion (claim 8735), distributed across heavily leveraged debt and strategic equity. The debt stack includes five delayed-draw term loan facilities totaling approximately $28 billion (claims 4277, 4592, 2137, 9122, 1319, 8729), senior notes from multiple offerings exceeding $10 billion (claims 4700, 4697, 2125, 8927), $5.6 billion in convertible notes (claims 3980, 2126), and revolving facilities of $1.85 billion. Against this $45+ billion in debt, equity raised totaled roughly $7.2 billion from the IPO ($1.4 billion net, claim 4967), customer share issuances ($350 million, claim 4514), NVIDIA's staged $4 billion participation (claims 3891, 9316), and Jane Street's $1 billion commitment (claim 2080), yielding a 6x debt-to-equity ratio (claim 3162). This leverage is anchored to $104 billion in contracted backlog (claim 1493), with OpenAI at $6.5 billion (claim 4297) and Meta at $21 billion (claim 2124).

The capital sequence shows deliberate de-risking rather than opportunistic patching. The IPO provided the public platform; the December 2025 convertible ($2.587 billion, claim 3980) arrived after operational traction; infrastructure-backed DDTL syndication—achieving Ba2/BB+ ratings by May 2026 (claim 8732)—followed anchor customer commitments. This progression suggests each capital source entered at progressively de-risked milestones. The newest DDTL's 5.5-year maturity (claim 8730) against three-year average customer contracts (claim 8731) creates a refinancing pressure point.

Deployment tracks stated use. Q2 2026 revenue of $2.58 billion (claim 8183) grew 112% year-over-year (claim 8185), with capex spending of $32.5 billion projected for full-year 2026 (claim 7925) and approximately $9 billion from customer prepayments (claim 8575). Quarterly interest expense of $640 million (claim 7341)—nearly 25% of revenue—combined with negative net margins of -25% (claim 10221) and -48% return on equity (claim 10220) indicates debt service is compressing profitability despite scale. January 2026 covenant amendments and DSCR testing deferrals (claims 3911, 3913) signal active management of near-term liquidity during scaling.

Investors should monitor customer concentration (OpenAI and Meta represent approximately 26% of backlog), refinancing windows as maturities approach 2030–2032, and execution risk for infrastructure-backed syndication should customer contract economics shift.

CoreWeave has pursued an infrastructure-backed financing strategy rather than traditional capital-stacking opportunism. Beginning with a $1.4 billion IPO in May 2025 (per claim 4967), the company quickly layered debt, convertible notes totaling $2.587 billion (per claim 3980), and strategic equity from infrastructure partners, ultimately raising over $30 billion in debt and equity capital in 2026 alone (per claim 8735).

The dilution trajectory reflects this deliberately: minimal equity dilution from convertibles despite $2.587 billion issuance, with only $30 million in actual conversion terms noted (per claim 3984). Instead, the capital mix prioritized asset-backed lending—particularly delayed-draw term loans whose syndication CoreWeave claims as a first for HPC infrastructure (per claim 9117). This financing approach aligns with the announced use of proceeds: infrastructure capex tied to customer contract capital expenditures and GPU servers (per claim 2138), with customer prepayments funding capex directly to the tune of $9 billion in 2026 (per claim 8575).

The strategy shows coherence but reveals underlying stress. Covenant amendments and testing postponements in January 2026 (claims 3911, 3913) indicate initial debt terms were overambitious. By Q2 2026, reported metrics showed a 6:1 debt-to-equity ratio (per claim 8202), negative 25 percent net margin (per claim 10221), and negative 48 percent return on equity (per claim 10220), with current and quick ratios both at zero (claims 7379, 7380). Interest expense reached $640 million quarterly (per claim 7341) against roughly $2.6 billion in revenue (per claim 8183).

This structure only functions if the $104 billion sales backlog (per claim 962) delivers revenues at acceptable margins and if customer prepayment flows sustain capex execution amid capital guidance of $32.5 billion to $37 billion (claims 7925, 7890). Revenue growth of 112 percent year-over-year (per claim 8185) supports visibility, but widening operating and net losses (per claim 1495) require explanation. A sophisticated reader should monitor: whether prepayment revenue recognizes at margins sufficient to service 5-year debt maturities (per claim 9120); covenant compliance trajectories given near-zero working capital; and actual capex deployment relative to announced guidance. The company has wagered on execution risk rather than capital risk, shifting dilution burden to leverage.

CoreWeave has raised over $30 billion in debt and equity since May 2025 (claim 8735), executing a capital structure that is architecturally coherent but operationally strained. The dilution story breaks cleanly into phases: IPO proceeds of $1.4 billion in May (claim 4967), followed by strategic equity raises—$350 million from customers (claim 4514), $2 billion from NVIDIA in January 2026 (claim 3891), and $1 billion from Jane Street in April (claim 2080)—that anchor the shareholder base rather than fragment it. Equity dilution from insiders is absent from the record. The debt mix tells a deliberate story. Early senior notes in 2025 at 9–9.25% give way to convertible securities starting December 2025, with $2.59 billion in net proceeds (claim 3981) and $3 billion more in April 2026 (claim 2126), reducing near-term dilution by pushing conversion decisions to equity holders. Overlaying this are delayed draw term loans—$8.5 billion in March (claim 2137) and $3.1 billion in May (claim 1319)—priced at customer contract cash flows rather than corporate credit. The newest tranche, $2.6 billion in August, carries secured DDTL status and Ba2 / BB+ ratings (claim 8732), marking a shift to asset-backed lending as unsecured capacity compressed. The use of proceeds aligns with backlog. Sales backlog of $104 billion (claim 962) is funded partly by $21 billion Meta commitment and $6.5 billion OpenAI commitment (claims 2124, 4297), and partly by customer prepayments that directly finance capex—$9 billion in 2026 alone (claim 8575). Full-year capex guidance of $32.5 to $37 billion (claims 7890, 7925) is aggressive but proportional to contracted revenue. However, operating reality requires close watching. Q2 net loss of $626 million against $2.58 billion revenue (claims 3014, 8183) and interest expense of $640 million (claim 7374) create a debt-to-equity ratio of 6x (claim 7378) and negative return on equity of 48 percent (claim 7372). January 2026 covenant amendments for debt service coverage and minimum liquidity testing (claims 3912, 3911) signal the capital plan runs close to covenant thresholds, dependent on customer cash realization. The convertible maturity and earliest redemption dates (claim 3985) approach, creating refinancing timing risk. What matters next: whether customer prepayments and contract realization sustain draw-down on the $8.5 billion and $3.1 billion DDTLs, and whether interest expense ($640 million quarterly) compresses as debt cost reduction materializes (claim 8547). The backlog is substantial; execution risk is the constraint.

CoreWeave has orchestrated a capital structure that appears deliberately engineered rather than opportunistically assembled. From IPO in May 2025 through August 2026, the company raised over $30 billion, per claim 8735. The mix reveals coherence: customer prepayments and direct commitments (Meta's $21 billion payment commitment, per claim 2124; OpenAI's $6.5 billion committed revenue, per claim 4297) anchor infrastructure-backed delayed draw term loans ($8.5 billion facility rated Ba2/BB+, per claims 9122 and 8732). Convertible offerings ($2.587 billion in December 2025, per claim 3980; $3 billion in April 2026, per claim 2126) provide capital at lower coupon than senior debt. Senior notes proliferated across 2025-2026 (9.250%, 9%, multiple tranches totaling several billion), with pricing tightening as credit ratings improved. Strategic equity investors—NVIDIA ($2 billion, per claim 3891), Jane Street ($1 billion, per claim 2080)—arrived as both capital partners and implicitly de-risking customer concentration. This is not a company patching funding gaps; it is matching customer contract duration to infrastructure financing tenor.

The stated use of proceeds tracks actual deployment closely enough. Customer contract capital expenditures and GPU servers, per claim 2138, consumed $9 billion in Q2 2026 capex (claim 1524) and are guided to $32.5 billion for full-year 2026 (claim 7925). Customer prepayments of $9 billion are funding capex (claim 8575). A $104 billion sales backlog (claim 1824) suggests capex intensity will persist for years.

One structural signal warrants attention: in January 2026, the company amended its minimum liquidity covenant (delayed status, claim 3911) and postponed debt service coverage ratio testing (claim 3912). These revisions occurred after the December 2025 convertible offering and substantial prior capital raises, suggesting early covenant pressure despite capital abundance. The company is simultaneously unprofitable—Q2 2026 net loss of $626 million on $2.58 billion revenue (claims 3014, 8183)—with quarterly interest expense of $640 million (claim 7374), consuming 25 percent of revenue. Debt stands at six times equity (claim 7378), and total liabilities of $72 billion dwarf equity of $5 billion (claims 637, 638).

A sophisticated reader should monitor whether further covenant amendments occur and track the path to operating cash flow generation. The $104 billion backlog is valuable only if customer contracts renew and do not shift to competitors or in-house solutions. Convertible dilution and strategic investor behavior—whether NVIDIA and Meta remain committed capital partners or become sellers—will signal confidence in the infrastructure thesis holding.

CoreWeave executed an aggressive but strategically coherent capital plan post-IPO in May 2025, raising more than $30 billion through coordinated debt and equity issuance, per claim 8735. The financing mix included $1.4 billion in IPO proceeds (claim 4967), convertible notes totaling $2.5 billion (December 2025, claim 3981) and $3 billion (April 2026, claim 2126), senior notes in four tranches ($1.75B, $1.5B, $1B, $3.5B per claims 4658, 4700, 2056, 8927), and infrastructure-backed term loans ($3.1 billion DDTL 5.0 facility, claim 9116). This was not opportunistic patching but customer-led infrastructure financing: Meta committed $21 billion (claim 2124) and OpenAI $6.5 billion (claim 4297), with the stated capex guidance of $32.5–$37 billion for 2026 (claims 7925, 7890) closely tracking capital raised.

Yet performance reveals structural stress. Q2 2026 generated $2.58 billion in revenue, up 112 percent year-over-year (claim 8177), but posted a net loss of $626 million (claim 3014). Interest expense ran $640 million (claim 7341)—approximately 25 percent of revenue. The debt-to-equity ratio stands at 6x (claim 3162). The $104 billion sales backlog (claim 1493) is substantial but should be read alongside covenant amendments in early 2026 postponing debt service coverage and minimum liquidity testing (claims 3912, 3911), which suggests lenders anticipated the cash-burn profile.

CoreWeave operates on an infrastructure-as-deployed model, with customer prepayments funding capex ($9 billion in prepayments, per claim 8575). This is not a conventional balance sheet, but a customer-financed build-to-utilization structure. A sophisticated reader should track capex execution velocity, customer concentration risk (a handful of mega-customers underwriting the entire facility), refinancing calendars for debt maturing through 2030–2032, and the trajectory of net losses as capex moderates and infrastructure utilization ramps. The company bet heavily on customer lock-in and utilization economics to sustain the leverage. Whether that thesis holds will determine whether this financing strategy was prescient or over-levered.

CoreWeave has executed a remarkable but structurally strained capital raise: $30 billion-plus in 2026 alone, per claim 8735, deployed against a $104 billion customer backlog (claim 647). The financing cascade reveals a company with demand it cannot ignore but insufficient self-generated cash to meet it. IPO proceeds of $1.4 billion (claim 4967) in May 2025 proved a placeholder; the real work began in early 2026 when customer commitments—$21 billion from Meta (claim 2124), $6.5 billion from OpenAI (claim 4297), $2 billion from NVIDIA equity (claim 3891)—made the capital structure existential. The mix suggests not opportunism but rather necessity-driven coherence: convertible debt, infrastructure-backed delayed draw term loans (the first publicly syndicated model, per claim 9117), and customer equity stakes all serve a single function—funding GPU deployments against contracted revenue. The company is deployed capital-intensive but revenue-rich if illiquid: Q2 2026 revenue of $2.58 billion (claim 8183) grew 112 percent year-over-year (claim 8185), yet the net loss widened to $626 million (claim 3014) and interest expense alone consumed $640 million (claim 7374). The trajectory is unsustainable without improvement. Debt-to-equity of 6x (claim 7378) mirrors capital-structure debt of roughly $30 billion (claim 1525) against equity of $5 billion (claim 638), a ratio justified only by the visibility of customer contracts with three-year terms (claim 8731). Covenant amendments in January 2026—postponing debt service coverage ratio and contract realization ratio testing (claims 3912, 3913)—flagged early stress; the company appears to have negotiated relief rather than default. Watch whether capex guidance of $32.5 billion to $37 billion (claims 7890, 7925) holds as interest costs escalate and customer prepayments (cited as $9 billion in 2026, claim 8575) prove insufficient ballast. The question is not whether demand exists but whether customer concentration and negative unit economics—losing money on billion-dollar quarterly revenues—can reverse before liquidity pressure forces a restructuring.

CoreWeave's sixteen-month capital trajectory reveals a deliberate infrastructure-financing model rather than opportunistic patching. From May 2025 onward, the company secured anchor customer commitments—OpenAI $6.5 billion (claim 4297) and Meta $21 billion (claim 2124)—then systematically raised debt and convertibles against those contracts. By August 2026, CoreWeave had secured over $30 billion in total capital (claim 8735), weighted heavily toward debt: senior notes ($1.5B, $1.75B, $1.75B, $3.5B per claims 4700, 4697, 4658, 8927), convertible notes ($2.587B per claim 3980), and delayed-draw term loans ($2.6B, $3.0B, $8.5B, $3.1B, $2.6B per claims 4592, 4277, 2137, 9116, 8729). Equity came primarily from strategic investors—NVIDIA twice at $2 billion each (claims 3891, 9316) and Jane Street $1 billion (claim 2080)—not public dilution. The IPO (May 2025, $1.4B net proceeds per claim 4967) was modest relative to total capital raised. This is coherent strategy, not patch-work: customer contract → syndicated debt financing → strategic equity validation.

However, announced capex deployment tracks poorly against operating reality. CoreWeave stated it would finance customer infrastructure and GPU servers (claim 2138), and reported Q2 2026 capex of $9 billion with full-year guidance of $37 billion (claims 1524, 7890), supported by $9 billion in customer prepayments (claim 8575). Yet the $104 billion sales backlog (claim 962) sits alongside devastating unit economics: Q2 2026 net loss of $626 million despite $2.58 billion revenue (claims 3014, 8183), negative net margin of negative 25 percent (claim 7371), return on equity of negative 48 percent (claim 7344), and negative free cash flow of $6 billion (claim 948). Quarterly interest expense of $640 million (claim 7341) consumes nearly a quarter of revenue. The company is unprofitable at scale.

A sophisticated reader should monitor three vulnerabilities. First, whether operating leverage emerges as utilization rises—currently, CoreWeave loses money on growth. Second, debt covenant compliance; early 2026 required amendments (claims 3912, 3911) suggesting tightness. Third, the $36 billion in future debt principal payments (claim 623) against negative free cash flow creates refinancing risk. The model works only if customer contracts accelerate and operating margins normalize sharply. Otherwise, this is capital-intensive growth funded by financial engineering atop structurally unprofitable operations.

CoreWeave's financing arc from May 2025 through mid-2026 shows strategic coherence rather than reactive patching, though with emerging structural strain. The company raised approximately $30 billion in debt and equity over twelve months (claim 8735), following an IPO that netted $1.4 billion (claim 4967). The capital stack evolved deliberately: after public debut, the company layered multiple debt instruments—senior notes of $1.5 billion and $1.75 billion (claims 4700, 4697)—alongside a $2.59 billion convertible offering (claim 3980), structured as long-duration instruments maturing 2030–2032. This reflects not opportunistic funding gaps but staged capital deployment matched to customer contract cycles. Strategic equity arrived late and concentrated: NVIDIA's $2 billion (claim 3891) in January 2026 and Jane Street's $1 billion (claim 2080) in April, suggesting partnerships rather than generic venture capital.

The deployment thesis is infrastructure-backed and customer-anchored. CoreWeave described the $2.6 billion delayed-draw term loan closed in August 2026 as the first publicly syndicated HPC infrastructure-backed financing (claim 9117), with an $8.5 billion facility similarly positioned (claim 9122). Use of proceeds tracked stated purpose: customer contract capital expenditures and GPU servers (claim 2138), with $9 billion in capex during Q2 2026 alone (claim 1524) and full-year guidance near $32.5 billion (claim 7925). Customer prepayments are funding capex directly, per disclosure (claim 8575). This alignment between capital raised and capital deployed is precise.

Structural stress emerged early. Covenant amendments and testing postponements in January 2026 (claims 3911, 3913) occurred well before debt maturity, signaling tightness within six months of IPO. Q2 2026 results show net loss of $626 million (claim 3014) despite $2.58 billion in quarterly revenue (claim 8183), free cash flow negative $6 billion (claim 1233), and debt-to-equity of approximately 6x (claim 8202). Interest expense of $640 million quarterly (claim 7374) is material but appears sustainable given a $104 billion sales backlog (claim 1493). The negative returns on equity and negative margins reflect a company in aggressive capex mode, not operational distress.

What requires watching: customer concentration. OpenAI's $6.5 billion commitment (claim 4297) and Meta's $21 billion (claim 2124) may represent the majority of that backlog. Should either customer defer, curtail, or renegotiate, the capex trajectory and debt service capacity shift materially. Refinancing needs cluster around 2030–2032, requiring the business to have moved to positive operating cash flow by then. The debt maturity ladder and rising interest burden leave limited margin for revenue disappointment.

CoreWeave raised over $30 billion from May 2025 through August 2026 (claim 8735), but the composition reveals a shift from balanced IPO positioning to heavy leverage in under eighteen months. The May 2025 public offering yielded $1.4 billion in net proceeds (claim 4967) alongside revolving credit expansion of $1.5 billion (claim 5010) and initial senior debt of $1.5–$1.75 billion (claims 4700, 4697). By late 2025, convertibles became dominant—$2.59 billion in December 2025 (claim 3980) and $3 billion in April 2026 (claim 2126)—while strategic equity checks from NVIDIA ($2 billion, claim 3891) and Jane Street ($1 billion, claim 2080) totaled $4.4 billion. Cumulative debt substantially outpaced equity: delayed-draw term loans reached $14.1 billion (claims 2137, 9116, 8729), supplemented by $8+ billion in senior notes (claims 8927, 2056, 2057). This is not opportunistic layering but rather a deliberate infrastructure-finance model. The May 2026 DDTL was explicitly the "first publicly syndicated HPC infrastructure-backed financing" (claim 9117), collateralized by customer contracts averaging three years (claim 8731), where prepayments fund capex and contracted cash flows service debt.

Use of proceeds tracks deployment precisely. Capital was designated for customer-contracted GPU infrastructure (claim 2138); Q2 capex of $9 billion against full-year guidance of $32.5–$37 billion (claims 1524, 7925) aligns with $9 billion in customer prepayments (claim 8575). Yet operational metrics reveal structural tension. Q2 2026 revenue of $2.58 billion (claim 8183) supported interest expense of $640 million (claim 7374)—one-quarter of revenue—against negative 25 percent net margins (claim 7376), negative $6 billion free cash flow (claim 1233), and 6:1 debt-to-equity (claim 7378). Current and quick ratios near zero (claim 7380) provide no liquidity cushion. The Meta commitment of $21 billion (claim 2124) concentrates customer risk in a single counterparty. Covenant amendments in January 2026 (claims 3912, 3913, 3911) and a May debt service coverage ratio of exactly 1.0 (claim 1324) signal limited headroom. Watch whether customer prepayments sustain at disclosed levels, whether Meta deploys the $21 billion on schedule, whether quarterly interest expense exceeds $640 million, and whether the company requires further covenant relief before reaching operating leverage.

CoreWeave's capital strategy reflects the specific capital intensity of AI infrastructure: after a $1.4B May 2025 IPO, the company rapidly assembled over $30B in debt and equity financing through mid-2026, per claims 4967 and 8735. The composition tells a story of operational pressure meeting market opportunity. The company began with traditional senior notes ($5B+ aggregated across May-August 2025, claims 4824, 4700, 4697), moved to convertible securities ($2.6B in December 2025, per claim 3981), then shifted to equity checkpoints—$2B from NVIDIA, $1B from Jane Street—paired with customer revenue commitments that function as financing (Meta's $21B through April 2026, per claim 2124). This is not opportunistic patching but a choreographed de-risking: each capital layer narrows as collateral or customer certainty improves. Yet the financing sequence masks a structural tension. Q2 2026 interest expense ran $640M quarterly (claim 7341), annualizing to $2.56B, against quarterly revenue of $2.58B (claim 8183). The debt-to-equity ratio sits at 6x (claim 8202), and Q2 free cash flow was -$6B (claim 1233), even as capex guidance expanded to $32.5B-$37B for full-year 2026 (claim 7925). The postponement of debt-service-coverage-ratio testing in early 2026 (claim 3912) signals that the company did not yet meet existing covenants on drawn debt. Use of proceeds tracks to capex—$9B deployed in Q2 alone (claim 621)—but customer prepayments fund $9B of that (claim 8575), meaning debt capital is not yet self-sustaining. What changed mid-2026 is the repricing of risk. The $3.1B term loan closing in May 2026 was investment-grade rated (claim 8732), marking the first syndicated HPC infrastructure-backed securitization (claim 9117). This suggests lenders now view cash flow collateral as bankable; it also implies CoreWeave believes it can absorb additional maturity through 2031-2032 without new equity. A reader should track whether customer prepayments sustain as capex tapers, whether the next covenant test triggers renegotiation, and whether guidance revisions signal margin pressure from infrastructure oversupply.

CoreWeave's post-IPO capital strategy reveals coherent structure underneath apparent frenzy. The company raised $1.4 billion in net IPO proceeds (claim 4967) in May 2025, then immediately stacked debt: senior notes of $1.5 billion (claim 4700) and $1.75 billion (claim 4697), escalated by $250 million (claim 4698), a $2.542 billion convertible net proceeds (claim 3981), and a $340 million capped call (claim 3986). Three tranches of delayed-draw facilities followed, totaling approximately $14.6 billion (claims 4277, 2137, 1319), capped by a $3.5 billion senior notes offering (claim 8927). By August 2026, CoreWeave had assembled more than $30 billion in debt and equity year-to-date (claim 8735). Rather than diluting shareholders post-IPO, the company locked strategic equity stakes: NVIDIA's $2 billion investment (claim 3891) and Jane Street's $1 billion (claim 2080). The capital mix—predominantly debt, minimal secondary equity dilution, strategic stakes from hyperscalers—reflects disciplined structuring, not opportunistic patching.

The announced use of proceeds tracks deployment closely. Two anchoring customer contracts frame the strategy: OpenAI's $6.5 billion committed revenue (claim 4297) and Meta's $21 billion payment commitment (claim 2124). The $8.5 billion delayed-draw facility explicitly targets customer contract capital expenditures and GPU servers (claim 2138). Q2 2026 capital spending reached $9 billion (claim 1524), supported by $9 billion in customer prepayments funding capex for the full year (claim 8575). The $104 billion sales backlog (claim 1493) justifies the capex trajectory.

However, the capital raise masks an acute profitability challenge. Q2 net loss was $626 million on $2.58 billion revenue (claims 1480, 8183), while quarterly interest expense reached $640 million (claim 1482)—a burn rate where financing costs consume 25 percent of quarterly revenue. Debt-to-equity ratios between 6 and 9 times (claims 7350, 1484) underscore the leverage. More revealing: January 2026 amendments to minimum liquidity covenants and postponements of debt-service-coverage testing (claims 3911, 3912) signaled lender concern immediately after the capital raise. The critical watch is whether customer prepayments scale faster than interest accruals, or whether covenant pressures resurface.

CoreWeave has executed a capital-intensive financing strategy calibrated to fund $37B in 2026 capex (claim 7890), raising more than $30B in debt and equity year-to-date (claim 8735). The sequencing is deliberate: IPO provided public currency with $1.4B net proceeds (claim 4967), followed by layers of subordinated equity—$2.59B in convertible notes (claim 3981) with a $340M capped call (claim 3986) to limit dilution, strategic investments from NVIDIA ($2B, claim 3891) and Jane Street ($1B, claim 2080)—then senior unsecured debt ($7B+ across five offerings through Q2 2026). This is not opportunistic patching but a coherent structure aligned to funding infrastructure capex for customer contracts averaging three years (claim 8731).

The capital stack shows discipline. Convertibles arrived after the IPO established pricing and public currency. Strategic equity from partners suggests customer alignment over equity-market opportunism. Senior debt arrived once revenue trajectory was visible: Q2 2026 generated $2.6B revenue on 112% year-over-year growth (claim 8185) against a $104B backlog (claim 1493). Deployment appears to track stated use: customer prepayments are funding $9B of capex in 2026 (claim 8575), with debt facilities explicitly targeting customer-contract capital expenditures and GPU servers (claim 2138). Yet the math requires scrutiny. Capex guidance of $37B against $30B+ raised and $9B in customer prepayments leaves a gap—suggesting either upcoming tranches or guidance revision.

The balance sheet carries execution risk. Quarterly interest expense of $640M (claim 1482) consumes 25% of quarterly revenue ($2.6B, claim 1479), while the company posted a $626M net loss and $6B negative free cash flow in Q2 2026 (claims 1480, 948). Debt-to-equity reaches 9:1 (claim 1484). This leverage is serviceable only if capex converts to revenue as promised and near-term profitability emerges. Covenant management has required amendment: minimum liquidity covenants were adjusted in January 2026 (claim 3911), and debt service coverage and contract realization testing were postponed (claims 3912, 3913), signaling cash tightness despite capital abundance. Customer concentration is pronounced—OpenAI ($6.5B committed, claim 4297) and Meta ($21B, claim 2124) dominate the backlog. A sophisticated reader should monitor whether capex guidance holds or resets, whether interest coverage improves as revenue scales into profitability, whether customer concentration narrows through diversification, and whether the refinancing cadence required as debt matures remains feasible in a tightening capital market.

CoreWeave's capital structure reflects a coherent, if capital-intensive, strategy matched to a high-backlog business model. The company raised over $30 billion in debt and equity year-to-date through August 2026, per claim 8735, anchored by two megadeals: a $21 billion payment commitment from Meta (claim 2124) and a $6.5 billion revenue commitment from OpenAI (claim 4297). These commitments front-load customer relationships, enabling the company to securitize capex through contractual visibility. Delayed draw term loan facilities—$8.5B via claim 2137, $3.1B via claim 1319, and $2.6B via claim 8729—are structured on approximately 3-year maturities matching customer contract duration (claim 8731), suggesting infrastructure financing designed to synchronize debt service with cash inflows. The capital stack itself—$30 billion in long-term debt against $5 billion in equity, creating a 6:1 debt-to-equity ratio (claim 1484)—is extreme but defensible if backlog converts and prepayments sustain (claim 8575 cites $9 billion in customer prepayments funding capex in 2026).

The covenant amendments in January 2026 (claim 3911 on minimum liquidity, claim 3913 on contract realization ratio testing) signal early stress. Q2 results expose the core tension: $2.58 billion in revenue with 112% year-over-year growth (claim 8183), yet a $626 million net loss (claim 1480) and negative free cash flow of $6 million (claim 1233). Interest expense alone ran $640 million in Q2 (claim 1482), consuming the operating profit margin and forcing dependence on continuing customer prepayments and undrawn debt capacity to fund the $32.5 billion to $37 billion capex budget. Equity raises came from NVIDIA ($2B via claim 3891) and Jane Street ($1B via claim 2080), reinforcing a customer-as-shareholder dynamic rather than traditional venture or public market demand. The convertible offering in December 2025 ($2.587B issued, net proceeds $2.542B per claims 3980 and 3981) was further dilution masked as debt.

A sophisticated reader should monitor three vectors: first, whether the $104 billion sales backlog (claim 1824) converts on guidance and whether customer prepayments remain durable; second, refinancing risk as delayed draw facilities and $3.5 billion in convertible notes due 2032 (claim 2126) approach maturity; and third, whether positive free cash flow materializes at scale, which remains absent despite explosive revenue growth. The company has optimized for capex velocity and customer acquisition rather than unit profitability—a sustainable pattern only if customer commitments hold and refinancing remains available at current spreads (SOFR + 5.50% per claim 8733).

CoreWeave's eighteen-month capital raise—exceeding $30 billion through August 2026, per claim 8735—reflects a customer-driven financing strategy rather than a pre-planned capital structure. The company progressed from IPO in May 2025 (claim 4967: $1.4B net proceeds) through a layered funding model: senior debt issued across multiple tranches between May and July 2025, a $2.587B convertible offering in December (claim 3980), strategic equity from NVIDIA at $2 billion in January 2026 (claim 3891), and Jane Street's $1 billion commitment in April (claim 2080). The most revealing component is the shift toward customer-linked infrastructure financing—$8.5 billion (claim 2137), $3.1 billion (claim 1319), and a final $2.6 billion DDTL facility closed in August (claim 8729)—each sized to specific customer commitments (Meta's $21 billion payment commitment per claim 2124; OpenAI's $6.5 billion revenue commitment per claim 4297). This structure indicates not opportunistic patching but rather a deliberate model in which debt duration and size match customer contract tenors.

However, deployment appears constrained relative to raised capital. CoreWeave guided 2026 capex at $32.5 billion to $37 billion (claims 7925, 7890), yet Q2 2026 showed only $9 billion in actual capital spending (claim 1229) with negative $6 billion free cash flow (claim 1233)—suggesting a widening gap between committed capital and deployed capital. Q2 net loss reached $626 million against $2.58 billion revenue (claims 3014, 8183), driven by $640 million interest expense alone (claim 3015), implying annual debt service exceeding $2.5 billion on a quarterly revenue run rate of $10.3 billion. The 6:1 debt-to-equity ratio (claim 8202) reflects a capital structure designed for infrastructure, not profitability.

The covenant amendments in January 2026 merit attention: postponements of debt service coverage ratio testing and contract realization ratio testing (claims 3912, 3913) arrived before the Meta and subsequent financings were publicly announced, signaling early pressure on leverage metrics. Whether capex accelerates to justify the raised capital, or whether the company faces revised deployment timelines and covenant renegotiations, remains the critical watch. The negative net margin of 25 percent (claim 8179) despite 112 percent year-over-year growth (claim 8185) indicates the capital productivity question is structural, not temporary.

CoreWeave's post-IPO capital strategy reveals less a coherent playbook than a company forced to layer financing instruments in real time around customer demands. The company raised $1.4 billion in net IPO proceeds in May 2025 (claim 4967), but immediately began issuing debt at scale—approximately $5 billion in senior notes across May through July (claims 4824, 4700, 4697, 4658). By December, it added a $2.587 billion convertible offering (claim 3980) with a $340 million capped call (claim 3986), followed by strategic equity from NVIDIA ($2 billion, claim 3891) and Jane Street ($1 billion, claim 2080). This sequencing is not opportunistic patching; it is customer-anchored infrastructure financing.

The Meta commitment of $21 billion in April 2026 (claim 2124) rewired the capital structure entirely. The $8.5 billion delayed-draw term loan facility (claim 2137) is explicitly tied to "customer contract capital expenditures and GPU servers." An additional $3 billion convertible (claim 2126) and $1.25 billion senior notes (claim 2125) followed within days. Customer prepayments of $9 billion are disclosed as funding capex in 2026 (claim 8575), and full-year capital expenditure guidance runs $32.5 to $37 billion (claims 7925, 7890). Use of proceeds tracks deployment: capital raises are synchronized with customer purchase commitments and capex execution.

The structural vulnerability is profitability. Q2 2026 revenue was $2.575 billion (claim 1479) against a net loss of $626 million (claim 1480) and interest expense of $640 million (claim 1482). Net margin is negative 25 percent (claim 8179). Total debt stands at $35 billion (claim 1798) against stockholders' equity of $5 billion (claim 1515), producing a debt-to-equity ratio of 9 (claim 1484). Free cash flow burn in that single quarter reached $5.74 billion (claim 1486).

A revenue backlog of $104 billion (claim 1513) and 115 percent trailing-twelve-month growth (claim 1815) confirm demand is genuine. But the company is currently incurring $1 of net loss for every $4 of revenue while servicing $35 billion in debt. Covenant postponements in January 2026 (claims 3911, 3913) signal early stress on debt metrics. The central question is execution velocity: whether capex deployment converts backlog into positive operating leverage before debt-service capacity becomes unsustainable. Customer concentration risk—Meta's scale within the backlog—remains unrevealed but material.

CoreWeave's capital structure from May 2025 through mid-2026 reveals a deliberately debt-heavy strategy anchored to customer commitments rather than traditional equity dilution. The company raised approximately $2.6B in net convertible proceeds (claim 3981) and layered multiple tranches of senior notes—9.25%, 9%, and $1.25B due 2031 (claims 4824, 4697, 2125)—alongside three distinct delayed-draw term loan facilities totaling $8.5B by March 2026 (claim 2137). This debt-first approach was offset by three rounds of strategic equity: the IPO at $1.4B net (claim 4967), a $2B NVIDIA investment (claim 3891), and $1B from Jane Street (claim 2080). The pattern is coherent: debt services the capex, equity provides validation from tier-one technology buyers. Two December 2025 amendments—minimum liquidity covenant relaxation and postponement of debt service coverage testing (claims 3911, 3912)—signal early working-capital pressure despite the fundraising pace. The announced use of proceeds (GPU infrastructure for OpenAI at $6.5B committed revenue and Meta at $21B committed, per claims 4297, 2124) maps directionally to actual deployment: Q2 2026 capex reached $9B with full-year 2026 guidance of $32.5–$37B (claims 1524, 7890), funded partly through $9B in customer prepayments (claim 8575). However, the cash dynamics are troubling. Q2 2026 free cash flow was negative $6B despite $2.58B in quarterly revenue and 112% YoY growth (claims 948, 8183, 8185), while net interest expense consumed $640M per quarter (claim 7374). The debt-to-equity ratio reached 6–9x by Q2 (claims 1484, 3162), and absolute debt stands near $35B (claim 1798), creating a serviceability question if customer prepayments slow or capital deployment timelines slip. The weighted-average cost of borrowing fell approximately 300 basis points late in the period (claim 1803), suggesting earlier tranches carried higher rates—a potential lock-in cost if refinancing needs accelerate. A sophisticated reader should track whether the $104B sales backlog (claim 1524) sustains prepayment velocity and whether operating margins inflect positive as capex moderates; the current structure assumes backlog-to-cash conversion remains resilient and capex per dollar of revenue continues to decline.

CoreWeave's capital trajectory traces a company building infrastructure at scale through a layered strategy that became increasingly leveraged, then structurally stressed within months. The IPO in May 2025 raised $1,400M in net proceeds (per claim 4967), accompanied immediately by $1,500M in revolving credit expansion and senior notes issuance, establishing a debt-forward posture from the start. Through mid-2025, the company stacked convertible notes ($2,587M gross offering by December, per claim 3980, yielding $2,542M net per claim 3981) and serial tranches of senior notes at rising rates, while customer-specific secured financings—an $8.5B delayed-draw term loan tied to Meta infrastructure (per claim 7563) and $3.1B facilities (per claim 1319)—became the dominant funding source. By August 2026, total debt reached $35B (per claim 1798), against equity of roughly $5B, implying a debt-to-equity ratio near 7:1. The equity mix shifted notably: initial IPO and over-allotment represented roughly $1.5B, but NVIDIA's $2B strategic investment (January 2026, per claim 3891) and Jane Street's $1B (April 2026, per claim 2080) arrived after covenant amendments on debt-service coverage and contract realization ratios were already postponed (January 2026, per claims 3912-3913), signaling capital structure stress less than eight months after IPO. The stated use of proceeds—customer-facing capex, particularly Meta's infrastructure—does map to actual quarterly deployments of $9.4B in capital expenditures (Q2 2026, per claim 1801), but the deployment model contains a critical tension: CoreWeave reported Q2 2026 revenue of $2,580M (per claim 8183) against a net loss of $626M and interest expense of $640M (per claims 1480, 1482), implying negative net margins and operating margins barely positive. The $104B sales backlog (per claim 1513) and OpenAI's $6.5B committed revenue (claim 4297) suggest committed capacity, but the company burned $5.7B in free cash in a single quarter (per claim 1486) while posting negative return on equity of -48% (per claim 8178). A reader should track whether debt service coverage covenant tests in 2026-2027 hold or trigger further amendments, whether capex-to-revenue conversion actually yields profitability, and whether the cost of debt (weighted-average borrowing cost, per claim 1803) continues falling enough to ease the burden of $35B in principal across a rising-rate environment. The current model funds growth through customer commitments, not cash generation; it requires either revenue-to-profitability transition or continued equity infusions to sustain.

CoreWeave's capital structure reveals methodical debt layering matched to customer-specific infrastructure deployments, not opportunistic patching, but execution risk is rising. The company IPO'd in May 2025 with $1.4B net proceeds (claim 4967) embedded within $17.2B total capital raised (claim 4968), then scaled debt aggressively: senior notes at 9.25% and 9% (claims 4824, 4697), a $2.59B convertible offering (claim 3980) with $340M capped-call hedges (claim 3986) to limit dilution, and sequential delayed-draw facilities of $2.6B, $8.5B, and $3.1B (claims 4592, 2137, 1319) tracked against customer commitments from OpenAI ($6.5B, claim 4297) and Meta ($21B, claim 2124). By Q2 2026, total debt reached $35B (claim 1798) against $5B equity (claim 1515), a 7:1 ratio sustainable only if backlog converts to revenue.

The $104B sales backlog (claim 1513) supports this thesis. Revenue grew 112% to $2.58B in Q2 (claims 8183, 1815), demonstrating traction. But cash generation is deeply negative: the company burned $5.74B free cash flow (claim 1486) in Q2 because capex of $9.4B (claim 1801) and interest expense of $640M (claim 1799)—25% of quarterly revenue—overwhelmed operations. Full-year 2026 capex guidance of $32.5–37B (claims 7925, 7890) implies the trajectory will worsen before improving. Critically, the company is unprofitable: Q2 net loss was $626M (claim 1480) on a -25% net margin (claim 7343).

Covenant erosion signals tightening. In January 2026, CoreWeave postponed debt service coverage ratio testing (claim 3912) and required a liquidity covenant amendment (claim 3911)—standard tells that cash flow fell short of lender expectations. The postponements bought time to demonstrate backlog-to-revenue conversion but signal lenders are monitoring closely. The path forward is binary: if the company deploys its $32.5–37B capex efficiently and converts backlog into cash-generative revenue by 2027, it can begin deleveraging before maturities arrive (senior notes due 2030 onward, per claim 4824). If backlog realization stalls or capex exceeds plan, the company faces refinancing pressure on $35B outstanding at higher rates given negative earnings. Customer fulfillment by Meta and OpenAI and quarterly FCF-to-capex trending are the metrics to watch; any material slowdown would force capital markets activity or equity dilution at reduced valuations.

CoreWeave has executed a high-leverage capital strategy since its May 2025 IPO that exhibits internal coherence around a customer-financed infrastructure model but carries material execution risk. Total gross capital raised through September 2026 exceeds $30 billion: equity of $4.8 billion including the IPO (claim 4967), NVIDIA $2 billion (claim 3891), and Jane Street $1 billion (claim 2080); senior debt of approximately $8.5 billion across multiple issuances; convertibles of $5.6 billion (claims 3980, 2126); and undrawn delayed-draw term facilities totaling $15 billion (claims 4592, 2137, 1319). The dilution trajectory moved immediately post-IPO rather than preserving capital: within months, the company layered convertibles and facilities against announced customer commitments—Meta's $21 billion capital commitment (claim 2124) and OpenAI's $6.5 billion (claim 4297)—suggesting these large customers effectively collateralize the deeper leverage tiers.

The capital mix reflects discipline rather than opportunism. Strategic equity from semiconductor and trading firms, senior debt positioning, and convertibles with capped-call hedges ($340 million per claim 3986) show intentional structure design. Use of proceeds tracks coherently to deployment: the announced capex guidance of $32.5-37 billion for 2026 (claim 7925) aligns with serving the $104 billion sales backlog (claim 1513), and Q2 capital spending of $9.4 billion (claim 1801) is proportionate to $2,575 million in quarterly revenue (claim 1479).

Yet deployment has not produced the margin inflection the capital structure requires. Q2 2026 reported a $626 million net loss (claim 3014) with $640 million in interest expense (claim 1482)—financing costs now approximate reported operations. Debt-to-equity reaches 9x (claim 1484), and free cash flow reflected $5.7 billion burn in the quarter (claim 1486). A covenant amendment was postponed in January 2026 (claim 3913), only nine months after the IPO, indicating initial financial models assumed faster operating margin recovery than materialized.

Readers should monitor whether operating margins inflect positive as capex moderates in late 2026, validating the leverage thesis. Refinancing risk on the 2029-2032 maturity wall requires attention if profitability remains delayed. Customer concentration—Meta may represent substantial backlog fraction—warrants scrutiny. If interest burden grows faster than revenue margin expansion, financial flexibility to maintain capex pace through demand cycles may constrain.

CoreWeave's capital trajectory from its May 2025 IPO ($1.4 billion net proceeds, claim 4967) through Q2 2026 shows a company financing infrastructure deployment faster than it generates operating cash flow. In 16 months, it raised roughly $17 billion cumulatively through a sequence that appears opportunistic: traditional bank facilities first ($1.5 billion revolver, claim 5010), senior notes as the IPO window closed (claim 4697: $1.75 billion in July 2025), convertibles later ($2.587 billion in December, claim 3980; $3 billion in April 2026, claim 2126), and equity when available (NVIDIA $2 billion, claim 3891; Jane Street $1 billion, claim 2080). The pattern—equity when signalable, debt when abundant—is orderly but reveals capital constraints shaped choices more than strategy.

Use of proceeds aligns to specific customer needs. The $8.5 billion delayed-draw facility targets customer capex and GPU servers (claim 2138); Meta's $21 billion payment commitment (claim 2124) and OpenAI's $6.5 billion revenue commitment (claim 4297) appear to anchor deployment. Q2 2026 capex of $9.4 billion (claim 1801) reflects this model. But financial results expose the tension. Q2 revenue of $2.575 billion (claim 1479) produced a $626 million net loss (claim 1480) with $640 million in interest expense (claim 1799)—debt service consumed the entire quarter's operating income. Free cash flow was negative $5.74 billion (claim 1486). Total debt reached $35 billion (claim 1798) against $8 billion trailing-twelve-month revenue (claim 6223) and a $2 billion TTM net loss (claim 6224), yielding debt-to-equity of 9x (claim 1484).

Covenant amendments in January 2026 signal strain: minimum liquidity thresholds loosened and debt service coverage and contract realization testing postponed (claims 3911, 3912, 3913). Yet lenders continued funding and weighted-average borrowing costs fell 300 basis points (claim 1803). Debt concentrates $15 billion due after 2030 (claim 6194).

A sophisticated reader should focus on whether operating cash flow inflects before 2030. The company bets on customer volume ramping and driving down cost-per-unit (the $104 billion sales backlog per claim 1824 is suggestive but unaudited). If customer commitments slip or unit economics remain poor, refinancing becomes acute. The question is not whether CoreWeave can raise capital—markets have repeatedly acquiesced—but whether deployed infrastructure can achieve the operating margins necessary to service $35 billion debt from $8 billion annual revenue.

CoreWeave's capital structure reveals a company building real infrastructure amid fragile financial engineering. From IPO completion in November 2025 through Q2 2026, the company raised approximately $35 billion in total debt and equity, but the mix betrays opportunism over coherent strategy. Debt dominates: senior notes totaling $5.5 billion (per claims 2125, 2056, 2057), convertibles of $2.587 billion (claim 3980) and $3 billion (claim 2126), plus delayed-draw term facilities now standing at $20.2 billion arranged across five successive facilities (claims 4592, 4277, 2137, 1319). Equity infusions were selective—NVIDIA's $2 billion (claim 3891) in January 2026 and Jane Street's $1 billion (claim 2080) in April came after public market stress, not before it. By Q2 2026, debt reached $35 billion against $5 billion equity (claims 1798, 638), a 7:1 leverage ratio (claim 1484 reports 9:1). This is not strategic sizing; it is layering.

The announced deployment tracks reality, but the arithmetic is brutal. CoreWeave reported $2.575 billion Q2 revenue (claim 1479) against $9.4 billion capex (claim 1801), a debt-fueled infrastructure build targeting announced mega-customer commitments: Meta's $21 billion (claim 2124) and OpenAI's $6.5 billion (claim 4297). The $104 billion backlog (claim 1513) appears genuine. However, Q2 free cash flow burned $5.7 billion (claim 1486) even as interest expense topped $640 million (claim 1482)—a quarterly cash drain exceeding quarterly revenue by 2.2x. Interest expense guidance jumped to $860 million in Q3 (claim 2965), a 34 percent increase on new debt layers. More concerning: in January 2026, just six weeks after closing a $2.587 billion convertible offering, CoreWeave postponed debt service coverage ratio testing and amended minimum liquidity covenants (claims 3912, 3911). A notice of delayed interest payment appeared simultaneously (claim 4659). These are not routine administrative moves.

A sophisticated reader should track whether CoreWeave can scale revenue to approach serviceability. At current burn, the company needs annual revenue approaching $12–15 billion to service debt sustainably; trailing twelve-month revenue stands at $8 billion (claim 6223). The backlog is material, but conversion risk and GPU market cyclicality matter enormously. Both NVIDIA and Meta are strategic investors and demand drivers—if either customer cycles or deploys in-house, CoreWeave's capex thesis collapses. Watch covenant amendment frequency and the weighted-average interest rate (they reduced it by 300 basis points per claim 1803, suggesting active refinancing), both signals of liquidity pressure beneath guidance.

CoreWeave's dilution trajectory reflects a deliberate sequencing: an IPO raising $1.4 billion in May 2025 (claim 4967), followed by a $2.587 billion convertible offering in December 2025 (claim 3980) and strategic equity injections from NVIDIA ($2 billion, claim 3891) and Jane Street ($1 billion, claim 2080). Total dilutive equity across the period reached roughly $4.8 billion against an estimated $35 billion in total debt raised (claim 1798), suggesting measured shareholder dilution relative to total capital deployed.

The capital mix demonstrates coherence rather than opportunistic patching. CoreWeave layered senior notes—$1.5 billion and $1.75 billion in July 2025 (claims 4700, 4697)—alongside multiple delayed draw facilities ($8.5 billion, $3.1 billion, and others spanning 2025-2026, per claims 2137, 1319). A July 2025 attempt to increase an offering by $250 million was delayed (claim 4699), indicating market receptivity had limits, yet the December convertible offering succeeded at scale. The $340 million capped call transaction (claim 3986) reflects embedded option value pricing.

Use of proceeds tracks deployment. CoreWeave stated delayed draw facilities would fund customer contract capital expenditures and GPU servers (claim 2138); Q2 2026 capital spending of $9.4 billion (claim 1801) and full-year guidance of $32.5 billion to $37 billion (claim 7925) align with committed customer infrastructure. The OpenAI $6.5 billion revenue commitment (claim 4297) and Meta $21 billion payment commitment (claim 2124) appear to underpin the debt strategy, providing visibility for debt service.

Strain points require scrutiny. Debt-to-equity reached 7x by Q2 2026 (claim 1484), with $35 billion in total debt against $5 billion equity (claims 1798, 1515). Interest expense of $640 million quarterly (claim 1482) consumes roughly 320 basis points of trailing revenue (~$8 billion, claim 1510). More acutely, Q2 free cash flow was negative $5.74 billion (claim 1486) despite a $104 billion backlog (claim 1513), indicating CapEx outpaces cash generation. Debt covenant amendments in January 2026 (claims 3912, 3911) suggest the initial structure assumed faster cash conversion.

A sophisticated reader should monitor whether the $104 billion backlog converts to revenue quickly enough to sustain debt service and $32-37 billion annual CapEx. The $300 basis point reduction in weighted average borrowing cost (claim 1803) hints at refinancing optionality, but persistent negative free cash flow may necessitate equity dilution or asset sales if revenue growth falters.

CoreWeave's capital raising since May 2025 reflects crisis-driven financing rather than strategic architecture. The company raised approximately $35 billion in debt (per claim 1514) against $5 billion in equity through the IPO and strategic investors, establishing a 7:1 leverage ratio. The debt mix itself fragments across four tiers: revolving facilities ($1.85 billion, per claims 5010, 5011), term loans ($17.2 billion across four facilities per claims 4592, 4277, 2137, 1319), senior notes ($7-8 billion per claims 4824, 4700, 4697, 2125, 2057, 2056), and convertible notes ($5.6 billion per claims 3980, 2126, 3133). The sheer number of separate facilities—three DDTL tranches in nine months (July 2025 through May 2026)—suggests capital needs escalated beyond initial underwriting. The announced use of proceeds aligns narrowly with observed deployment: claim 2138 specifies financing for customer contract capital expenditures and GPU servers, and Q2 2026 capex of $9.4 billion (claim 1801) with H1 total capex of $16 billion (claim 3038) tracks the infrastructure-as-a-service model. Yet the operational tension is acute. Q2 2026 revenue reached $2.575 billion (claim 1479) against a $104 billion sales backlog (claim 1513), generating a $626 million net loss (claim 1480) while interest expense alone hit $640 million (claim 1482). The weighted-average cost of borrowing declined approximately 300 basis points (claim 1803), indicating refinancing activity, but covenant amendments in January 2026 (claims 3911, 3913) reveal the original facility terms became unworkable almost immediately—DSCR and contract realization ratio testing were postponed and liquidity covenants amended. The inflection points warrant attention: the OpenAI commitment of $6.5 billion (claim 4297) and Meta commitment of $21 billion (claim 2124) preceded the largest debt tranches, suggesting customer contracts drive facility sizes rather than operating performance. Two strategic equity checks—NVIDIA's $2 billion (claim 3891) and Jane Street's $1 billion (claim 2080)—followed these commitments, implying management accepted dilution only after securing customer capital certainty. The free cash flow burn of $5.74 billion in Q2 2026 alone (claim 1486) creates a path dependency: sustained backlog realization is not optional but required to service $2.5+ billion in annualized interest expense. A sophisticated reader should monitor whether management can hold covenant thresholds through 2027 without further amendment, track the cash conversion cycle on the $104 billion backlog, and assess whether customer concentration risk (two contracts representing a substantial backlog portion) creates refinancing hazard if either contract delays or downsizes.

CoreWeave's capital pattern through 2026 reveals a financing structure that has shifted repeatedly in response to unanticipated capital intensity. The company IPO'd in May 2025 with $1.4 billion in net proceeds (claim 4967), paired with $1.5 billion in revolving credit expansion (claim 5010), but within months pivoted toward debt financing and convertible securities. By December 2025, just seven months after going public, CoreWeave issued $2.587 billion in convertible senior notes (claim 3980) and required a minimum liquidity covenant amendment (claim 3911). This January 2026 amendment—triggered mere weeks after the IPO—indicates the capital structure was inadequate for the company's actual deployment requirements.

The deployment story clarifies the pressure. CoreWeave reported $9.4 billion in capital expenditures during Q2 2026 alone (claim 1801), and raised full-year capex guidance to $35 billion (claim 2963). Against trailing twelve-month revenue of only $8 billion (claim 1510), this capex cadence is extreme. The company's negative free cash flow of $5.74 billion in Q2 2026 (claim 1486) and negative trailing net income of $2 billion (claim 1511) show no organic funding source for this build-out. Strategic equity became necessary: NVIDIA invested $2 billion in January 2026 (claim 3891), and Jane Street committed $1 billion in April 2026 (claim 2080). These mid-year injections suggest the May IPO and subsequent debt raises had not provided adequate cushion.

The debt stack by Q2 2026 stood at $35 billion (claim 1798) against stockholders' equity of $5 billion (claim 1515), yielding a 7:1 debt-to-equity ratio (claim 1484). Interest expense climbed to $640 million in Q2 (claim 1482), with Q3 guidance to $860 million (claim 2965), compressing an already-negative operating margin. The company's quick and current ratios both registered near zero (claims 7380, 7379), signaling minimal liquid cushion despite $104 billion in announced sales backlog (claim 1513). A debt maturity of $6 billion looms in 2027 (claim 3041).

This financing arc does not reflect premeditated strategy but rather reactive capital raising to fund infrastructure commitments—particularly the OpenAI $6.5 billion revenue contract (claim 4297) and Meta $21 billion infrastructure commitment (claim 2124). The August 2026 convertible offering ($3 billion with optional $450 million expansion, claims 3133, 3134) followed by Q3 interest expense guidance implies management expects further covenant pressure. Sophisticated investors should monitor whether backlog-to-revenue conversion accelerates sharply or whether the company must undertake additional dilutive capital raises to manage the 2027 debt maturity and sustain current capex guidance.

CoreWeave's 2026 financing pattern reflects a methodical execution of a capital-intensive GPU infrastructure strategy backed by concrete customer commitments, but reveals strain in converting that backlog into near-term profitability. Between March and May, the company secured approximately $20 billion in capital: $11.6 billion in delayed draw term loan facilities ($8.5 billion in March, per claim 2137, and $3.1 billion in May, per claim 1319), $4 billion in senior notes, convertible notes totaling $3.45 billion with a portion due 2032 (claims 2126, 2127), and a $1 billion equity stake from Jane Street (claim 2080). This mix reflects a deliberate preference for leveraged growth over equity dilution. The $21 billion Meta payment commitment announced in April (claim 2124) provides a credible anchor for the capital intensity, and the May DDTL facility explicitly ties deployment to customer contract capex and GPU servers (claim 1320), signaling financial discipline around use of proceeds.

The capital deployment aligns with the financing narrative. Q2 2026 capex reached $9.4 billion (claim 1801), consistent with a business scaling GPU server delivery to major customers. The $104 billion sales backlog (claim 1513) towers over $2.6 billion quarterly revenue (claim 1479), suggesting infrastructure investment targets tangible demand rather than speculative buildout.

The subsurface dynamics demand skepticism. Q2 net loss reached $626 million (claim 1480) while operating and net losses widened (claim 1495) despite revenue growth. Free cash flow burned $5.74 billion that quarter (claim 1486), implying the company consumes cash faster than customer payments can offset. Interest expense of $640 million quarterly (claim 1799) on $35 billion total debt (claim 1798) creates a debt-to-equity ratio of 9x (claim 1484), while the debt service coverage ratio covenant stands at exactly $1 (claim 2141)—the minimal sustainable threshold. The company did secure a 300-basis-point reduction in its weighted-average borrowing rate (claim 1803), easing pressure, but such concessions typically depend on lender confidence in an imminent path to profitability that these numbers have not yet demonstrated.

The question for sophisticated investors: does Q2 mark the trough of negative FCF and losses, or the beginning of sustained cash burn? The 40x backlog-to-revenue ratio cannot persist indefinitely without profitability momentum. Watch for evidence that revenue acceleration outpaces capex growth and cash burn narrows materially in coming quarters.

CoreWeave's record splits cleanly into three tiers, and only the top one carries contractual weight. The signed agreements are the order forms executed under existing Master Services Agreements: OpenAI (4296, building on 19820), NVIDIA (4352), and Meta (4282, 2123). These are filed contracts with dollar ceilings. OpenAI's tranches of $11.9 billion, $4 billion and $6.5 billion (19820, 25573, 20198) sum exactly to the stated $22.4 billion total. Meta's initial $14.2 billion (4282) was followed by a roughly $21 billion expansion through 2032 (10340, 2123, 9317). Even here, the language is "up to" and "initially committed," so these are ceilings rather than floors. The $11.2 billion backlog contribution attributed to OpenAI (4969), against the $11.9 billion headline (19820), suggests some of the headline value had not converted to booked backlog. Jane Street's roughly $6 billion (10304) and the Anthropic multi-year agreement (10731) are named and described as commitments, but the Anthropic deal carries no disclosed value, and neither appears here as a filed order form.

The second tier is unnamed. The "new hyperscaler customer" (25574) and the "leading AI enterprise" financed by DDTL 4.0 (11224) are never identified. The DDTL 3.0 customer (4595) is resolved to OpenAI only through a separate disclosure (26687). The third tier is relationship language. The January 2026 NVIDIA expansion is explicitly a "collaboration framework" (3892) that includes testing and validation work (17598) and a 5 GW aspiration (27948), distinct from the $6.3 billion capacity order (20209). IBM (4971), CrowdStrike (11768), Core Scientific (20313), Genesis Mission (11778) and the Aston Martin sponsorship (25576) carry no economics. The logo lists (25575, 11767, 9320) name counterparties without contract terms.

Concentration has shifted rather than eased. In early 2025, OpenAI alone was 43% to 46% of the $25.9 billion backlog (4970). By the end of 2026's first quarter, Meta's cumulative ceiling of about $35 billion exceeds OpenAI's $22.4 billion. NVIDIA is both supplier and customer, which adds circularity risk.

The arithmetic roughly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555). Named 2025 additions from OpenAI, Meta and NVIDIA total about $42.9 billion, which implies roughly $3 billion plus revenue burn-off from the hyperscaler and smaller logos. That residual is plausible but unverified, because the "up to" ceilings may overstate what is booked. Meta's $21 billion, Jane Street and Anthropic all postdate the $60.7 billion figure and should surface in 2026 RPO.

CoreWeave's record separates cleanly into contracts executed under master services agreements and a larger body of partnership and logo language. The signed tier is narrow. OpenAI is documented through an order form under an existing MSA (4296), with announced tranches of $11.9 billion (19820), $4 billion (25573) and up to $6.5 billion (20198), summing to the stated $22.4 billion (20198). Meta's order forms are filed for up to $14.2 billion through 2031 (4282, 2123) and a further roughly $21 billion through 2032 (10340, 17654). NVIDIA's September 2025 order form (4352) corresponds to the $6.3 billion arrangement (20209). Jane Street's roughly $6 billion commitment (10304) is named and quantified. The Anthropic agreement (10731) is named but carries no value. By contrast, NVIDIA's January 2026 'collaboration framework' toward more than 5 GW by 2030 (3892, 17596, 27948) is a framework, not contracted revenue. The IBM (4971), CrowdStrike (11768, 10330), Aston Martin (25576), Genesis Mission (11778) and Core Scientific (20313) items are partnership or participation language. The 'partner of choice' logo lists (25575, 11767, 10338, 9320) disclose no terms. Two counterparties remain unidentified: a 'new hyperscaler customer' (25574) and the 'leading AI enterprise' behind DDTL 4.0 (11224). The DDTL 3.0 'customer contract' (4595) is resolved as OpenAI by 26687.

Concentration has shifted rather than diluted. At March 2025, OpenAI's tranche was roughly 45 percent of the $25.9 billion backlog (4970). By December 2025, OpenAI's $22.4 billion and Meta's $14.2 billion together approximate 60 percent of $60.7 billion in RPO (3555). Meta's 2026 expansion lifts its headline total to about $35 billion, likely making it the largest single counterparty. Jane Street is the only sizable new non-hyperscaler name.

The arithmetic broadly reconciles, with caveats. Year-end 2024 RPO of $15.1 billion (3572) plus the 2025 named deals ($22.4B, $14.2B and $6.3B, about $42.9 billion) yields roughly $58 billion, near the reported $60.7 billion. That fit holds only if the 'up to' ceilings count at full value and burn-down is offset by unnamed wins. The first OpenAI tranche is cited at $11.9 billion (19820) but only $11.2 billion of backlog was added (4969), an unexplained $0.7 billion gap. RPO also fell to $14.7 billion in Q1 2025 (4970), implying that OpenAI value initially sat in conditional, non-RPO backlog. No post-2025 RPO figure is on record to test the 2026 Meta and Jane Street additions.

CoreWeave's signed contracts are concentrated in a few order forms filed under existing master services agreements. These are with OpenAI (claim 4296), Meta (claims 4282 and 2123) and NVIDIA (claim 4352). The disclosed values are binding, though the Meta and OpenAI figures are framed as "up to" amounts (claims 4282, 20198), which makes them ceilings rather than floors. OpenAI's tranches of $11.9 billion, $4 billion and $6.5 billion (claims 19820, 25573, 20198) add up to the stated $22.4 billion total. The earlier backlog description of the first tranche as $11.2 billion (claim 4969) is $0.7 billion short of the headline and is not explained. Jane Street's roughly $6 billion commitment (claim 10304) is named and sized. Anthropic's "multi-year agreement" (claim 10731) is named but unpriced and appears only through secondary reporting.

Much of the remaining record uses softer language. The NVIDIA 5 GW arrangement is explicitly a "collaboration framework" (claims 3892, 27948). The IBM, CrowdStrike, Core Scientific and Aston Martin items (claims 4971, 11768, 20313, 25576) are partnerships with no disclosed economics. Genesis Mission (claim 11778) is programme participation. The logo lists (claims 25575, 11767, 9320) carry no contract values. Two material counterparties remain unidentified: the "new hyperscaler customer" (claim 25574) and the "leading AI enterprise" financed by DDTL 4.0 (claim 11224). The counterparty behind DDTL 3.0 was later identified as OpenAI (claims 26687, 4595).

Concentration has broadened at the top but not in depth. In March 2025, backlog was $25.9 billion against RPO of $14.7 billion (claim 4970). The $11.2 billion gap equals the OpenAI addition, so OpenAI was effectively the entire non-RPO backlog. By late 2025, Meta, OpenAI and NVIDIA together accounted for roughly $42.9 billion of announced commitments. Meta's further $21 billion order through 2032 (claims 2123, 10340) lifts its cumulative headline to about $35 billion, overtaking OpenAI. NVIDIA's role as supplier, collaborator and $6.3 billion customer (claim 20209) is a circularity that investors should weigh.

The arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3560), a $45.6 billion net increase. Named 2025 deals total $42.9 billion. Revenue recognised during the year also drew the balance down, so unnamed bookings, plausibly including the hyperscaler, likely contributed several billion dollars that are not attributed to any counterparty. Several records also carry unit artifacts that should not be read literally, such as "$21,000" in claim 10729 and "$61" in claim 3555.

CoreWeave's record has few soft MOU-style announcements, but a lot of repetition, and the figures for the signed contracts need careful handling. The executed agreements are order forms under existing master services agreements. These are OpenAI (19820, then 4296), Meta (4282, then 2123 for the March 31, 2026 order form) and NVIDIA (4352). Jane Street's roughly $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) are also described as signed. Anthropic's has no disclosed value. Several items carry partnership language without disclosed economics: IBM (4971), CrowdStrike (11768, repeated as 10330), Aston Martin (25576), Core Scientific (20313) and the Genesis Mission (11778). The January 2026 NVIDIA expansion is explicitly a collaboration framework (3892, 17596, 27948) aimed at more than 5 GW by 2030, so it should not be read as contracted revenue. Exemplar Cloud status (10332, 9324) is a designation, not a contract. Some disclosures leave the counterparty unnamed. One is a hyperscaler, described as signed and expanded (25574). Another is a 'leading AI enterprise' whose contract backs DDTL 4.0 (11224). DDTL 3.0 at first referred only to 'a customer contract' (4595) but was later tied to OpenAI (26687). The logo lists (25575, 11767, 9320) name customers but give no dollar amounts.

Concentration is severe and growing. OpenAI commitments of $11.9 billion, plus $4 billion, plus $6.5 billion sum exactly to the stated $22.4 billion (20198). One mismatch remains: the first quarter booked the original OpenAI deal as $11.2 billion of backlog (4969), not $11.9 billion. Meta's initial commitment was up to $14.2 billion (4282). The later $21 billion order form is called a 'new' commitment (9317), which implies a cumulative Meta exposure near $35 billion. That figure is unconfirmed, because coverage repeats the same deal (10729, 10694, 10340, 17654).

On the arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of $45.6 billion before netting out recognized revenue. The 2025 OpenAI, Meta and NVIDIA deals total about $42.9 billion. That leaves the unnamed hyperscaler and smaller logos to explain the remainder plus whatever revenue was recognized in the year. OpenAI and Meta alone account for roughly 60 percent of year-end RPO, and about 70 percent once NVIDIA is added. NVIDIA is also a supplier and an investor. Many figures are stated as 'up to' ceilings, which is hard to square with RPO's committed-only definition. The roughly $27 billion of 2026 signings (Meta and Jane Street) has not yet appeared in any reported RPO.

CoreWeave's dollar-denominated commitments sit almost entirely with a few named counterparties under signed order forms. The firmest are the OpenAI order form of September 2025 (claim 4296, sized at up to about $6.5 billion in claim 20198), the Meta order form committing up to about $14.2 billion through December 2031 (claims 4282, 17655), the March 31, 2026 Meta order form of about $21 billion through 2032 (claims 2123, 10340, 9317), and the NVIDIA capacity order form (claim 4352, described as $6.3 billion in claim 20209). The original $11.9 billion OpenAI agreement (claim 19820) and the $4 billion expansion (claim 25573) are described as executed agreements. The Anthropic deal is called signed (claim 9318) but carries no value. The roughly $6 billion Jane Street commitment (claims 10304, 10730) is quantified but rests on press-release language rather than a filed order form.

A second tier uses softer language. The January 2026 NVIDIA arrangement is a 'collaboration framework' toward more than 5 GW by 2030, with testing and validation aims, and carries no stated payment obligation (claims 3892, 17596, 27948, 17598). The same applies to the IBM partnership (claim 4971), CrowdStrike 'global partnership' (claims 11768, 10330), Aston Martin sponsorship (claim 25576), Genesis Mission participation (claim 11778) and Core Scientific collaboration (claim 20313). The logo lists in claims 25575, 11767, 10338 and 9320 give no values. Three counterparties are not identified at all: the 'new hyperscaler customer' (claim 25574), the DDTL 3.0 'customer contract' (claim 4595), which claim 26687 separately ties to OpenAI, and the 'leading AI enterprise' financed by DDTL 4.0 (claim 11224).

Concentration has moved but not eased. OpenAI's $11.2 billion was roughly 43 percent of the $25.9 billion backlog in March 2025 (claims 4969, 4970). By December 2025, OpenAI's $22.4 billion plus Meta's $14.2 billion equal about 60 percent of $60.7 billion RPO (claim 3555), and about 71 percent with NVIDIA included. Meta's March 2026 addition makes Meta, not OpenAI, the largest exposure.

The arithmetic broadly reconciles, with caveats. The OpenAI figures sum correctly ($11.9B + $4B + $6.5B = $22.4B), but the initial deal was booked as $11.2 billion against a $11.9 billion headline. RPO also fell from $15.1 billion to $14.7 billion in Q1 2025 (claims 3572, 4970), which suggests the OpenAI deal first appeared in delivery-contingent backlog. The $30.6 billion rise from June 2025 backlog (claim 25586) to December RPO roughly matches about $27 billion of named deals. However, this compares backlog with RPO, and the 'up to' wording means these figures are ceilings rather than floors.

CoreWeave's record separates cleanly into dollarized order forms and softer relationship language. The signed customer agreements are the OpenAI contracts under a master services agreement: $11.9 billion initially (claim 19820), a $4 billion expansion (25573) and a September 2025 order form (4296) of up to $6.5 billion, which brings OpenAI to up to $22.4 billion (20198). That arithmetic adds up exactly, although the first quarterly print booked only $11.2 billion of the initial deal into backlog (4969). Meta signed an order form worth up to $14.2 billion through 2031 (4282) and a roughly $21 billion order form on March 31, 2026 (2123, 10340, 9317). NVIDIA's $6.3 billion capacity order form (4352, 20209) is also signed, though it reads more like a backstop for unused capacity than end demand. Jane Street committed about $6 billion (10304). The Anthropic agreement is described as signed and multi-year (9318), but no value was disclosed. Everything else is framework or partnership language with no committed dollars attached. That includes IBM (4971), Aston Martin (25576), Core Scientific (20313), CrowdStrike (11768, 10330), the Genesis Mission (11778), the January 2026 NVIDIA collaboration framework and its 5 GW ambition (3892, 17596, 27948), Exemplar Cloud status (10332), and the logo lists (25575, 11767, 10338). Several counterparties go unnamed. The record mentions a 'new hyperscaler customer' (25574) and a 'leading AI enterprise' whose contract DDTL 4.0 financed (11224). DDTL 3.0's 'customer contract' (4595) is identifiable only because a separate disclosure ties it to OpenAI (26687). Concentration has deepened. OpenAI, Meta and NVIDIA together account for about $42.9 billion of commitments disclosed through 2025, roughly 70% of the $60.7 billion of RPO at year-end (3555). The 2026 additions skew further toward Meta. On reconciliation, RPO actually dipped from $15.1 billion (3572) to $14.7 billion by March 2025 even as headline backlog reached $25.9 billion (4970). That gap reflects non-RPO amounts that depend on delivery conditions. Taking the $15.1 billion opening RPO and adding the three named 2025 deals gives about $58 billion before any revenue burn-down. That broadly squares with $60.7 billion, provided unnamed customers contributed a few billion more. The record does not include revenue recognized during 2025, so the bridge cannot be closed precisely. Readers should also be aware that the Meta expansion is triple-counted across press coverage (10694, 10729, 17654), and one of those entries is dated before the March signing.

CoreWeave's signed book rests on three named counterparties. OpenAI has executed order forms under a master services agreement: $11.9 billion (19820), a $4 billion expansion (25573) and a further order form of up to roughly $6.5 billion (4296, 20198). Those three sum to the stated $22.4 billion, so that tranche reconciles. The first leg, however, entered backlog at $11.2 billion rather than $11.9 billion (4969). Meta holds an executed order form of up to $14.2 billion through 2031 (4282). A March 2026 order form adds about $21 billion through 2032 (2123, 10340, 9317), which takes Meta past OpenAI as the largest counterparty. NVIDIA's $6.3 billion capacity purchase (4352, 20209) is a signed order form, but it comes from CoreWeave's own GPU supplier and should be read as circular. Jane Street's roughly $6 billion (10304) and the Anthropic multi-year agreement (9318) are described as signed, but the Anthropic agreement carries no disclosed value.

Much of the remaining language is framework or sponsorship rather than revenue contract. This covers the IBM 'partnership' (4971), the NVIDIA 'collaboration framework' toward 5 GW (3892, 17596, 27948), Core Scientific 'collaboration' (20313), the CrowdStrike 'global partnership' (11768), the Genesis Mission (11778), the Aston Martin F1 designation (25576) and 'Exemplar Cloud' status (10332). Logo lists such as Cohere, Mistral, Cursor and Midjourney (25575, 11767, 9319) name customers but attach no values. Three counterparties remain unidentified: a 'new hyperscaler customer' (25574), the DDTL 3.0 'customer contract' (4595, later tied to OpenAI by 26687), and the 'leading AI enterprise' behind DDTL 4.0 (11224).

Concentration has deepened rather than diversified. OpenAI, Meta and NVIDIA together account for about $42.9 billion of the $60.7 billion RPO at year-end 2025 (3555), roughly 70 percent. The 2026 Meta expansion lifts that counterparty alone to as much as $35 billion.

The revenue arithmetic only partially reconciles. RPO fell from $15.1 billion (3572) to $14.7 billion by March 2025 (4970), despite the OpenAI signing. That implies most of the OpenAI deal sat in the non-RPO, delivery-contingent portion of the $25.9 billion backlog. The bridge from $15.1 billion to $60.7 billion leaves an unexplained residual in the high teens of billions after named deals and 2025 recognition, plausibly the unnamed hyperscaler. Many figures are 'up to' caps. Some source values carry unit errors (10729, 3555), and claim 9320 was truncated in the record.

CoreWeave's record is unusually rich in signed, filed agreements, but they cluster around very few counterparties. The hard contracts are order forms under existing master services agreements, disclosed in 8-Ks. These cover OpenAI (claims 19820, 4296), NVIDIA (4352) and Meta (2123). The headline values attached to them are $11.9 billion, then $4 billion and $6.5 billion expansions for OpenAI (25573, 20198), $6.3 billion for NVIDIA (20209), and $14.2 billion followed by roughly $21 billion for Meta (4282, 10340). Several of these are framed as 'up to' amounts (4282, 20198), so they are ceilings rather than fixed take-or-pay figures. The opening OpenAI number also does not match across documents: $11.9 billion in the announcement (19820) against $11.2 billion added to backlog (4969). The Jane Street commitment of about $6 billion (10304) and the Anthropic agreement (9318) are described as signed. Anthropic carries no disclosed value, and neither deal appears in the record as a filed order form. Everything else is softer. The IBM 'partnership' (4971), the NVIDIA 'collaboration framework' targeting 5 GW (3892, 17596, 27948), the CrowdStrike and Core Scientific relationships (11768, 20313), the Aston Martin sponsorship (25576) and the Genesis Mission participation (11778) are not customer contracts. The recurring logo lists (25575, 11767, 9320, 9319) name customers but give no economics. Anonymity is an issue in three places. The 'new hyperscaler customer' (25574) is never identified. The DDTL 3.0 'customer contract' (4595) was later tied to OpenAI (26687). The 'leading AI enterprise' financed under DDTL 4.0 (11224) remains unnamed. Concentration has broadened in name only. OpenAI and Meta together account for roughly $57 billion of announced commitments. NVIDIA is both CoreWeave's supplier and a buyer of its capacity (4352), which makes it a circular counterparty. The arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555). The named 2025 deals sum to about $43 billion: OpenAI $22.4 billion, Meta $14.2 billion and NVIDIA $6.3 billion. That brings the total to roughly $58 billion before deducting revenue recognized during the year. Once burn-off is subtracted, a residual of several billion dollars is left for the unnamed hyperscaler and smaller customers, which is plausible but unverified. Backlog (4970, 25586) and RPO also use different definitions, and backlog adds amounts that depend on delivery conditions. The source record ends partway through claim 9317, so later 2026 disclosures are not assessed here.

CoreWeave's contracted book rests on a small number of named order forms executed under master services agreements, and these are the claims that bear weight. OpenAI is documented through an order form under its MSA (4296), with headline values of $11.9 billion (19820), a $4 billion expansion (25573) and a further up-to-$6.5 billion deal. The company's own total of up to roughly $22.4 billion (20198) reconciles exactly with those three figures. The one wrinkle is that the first-quarter release credited only $11.2 billion of backlog to the original deal (4969), against the $11.9 billion headline. Meta is likewise papered by order forms, one for up to about $14.2 billion through December 2031 (4282, 17655) and a second dated March 31, 2026 for about $21 billion through 2032 (2123, 10340). NVIDIA's $6.3 billion capacity purchase also sits on an order form (4352, 20209). Jane Street's roughly $6 billion is described as a commitment without a disclosed contract form (10304), and the Anthropic agreement carries no value at all (10731).

A second tier uses partnership or framework language that does not evidence revenue. This includes the NVIDIA collaboration framework and the 5 GW ambition (3892, 17596, 27948), the reference-architecture testing work (17598) and Exemplar Cloud status (10332, 9324). It also covers IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), Aston Martin (25576) and the Genesis Mission (11778). The logo lists (25575, 11767, 9320) and the claim of expanded relationships with Cohere and Mistral (9319) name counterparties but attach no economics.

The unidentified counterparties deserve attention. The "new hyperscaler customer" (25574) has never been named. The customer behind DDTL 3.0 (4595) was only later identified as OpenAI (26687), and the "leading AI enterprise" financed by DDTL 4.0 (11224) remains anonymous.

Concentration has shifted from OpenAI-dominant to an OpenAI-Meta duopoly. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), and the named 2025 deals with OpenAI, Meta and NVIDIA total about $42.9 billion. Measured against year-end RPO, that is roughly 70 percent. Adding the 2026 Meta order lifts Meta's cumulative headline to about $35 billion, likely making it the largest single obligor.

The arithmetic broadly reconciles, but with caveats. The $45.6 billion RPO increase exceeds the named deals by only a few billion before deducting revenue recognized in 2025, which leaves modest room for the unnamed hyperscaler and smaller logos. Several headline figures also carry "up to" qualifiers (4282, 20198), so recognized RPO may understate or overstate the announced totals.

CoreWeave's record has a small core of signed, dollar-denominated contracts surrounded by a larger layer of partnership and logo language. The hard contracts are OpenAI, Meta, Jane Street and NVIDIA. OpenAI's tranches of $11.9 billion, $4 billion and $6.5 billion sum to the stated $22.4 billion (19820, 25573, 20198), with later order forms under the existing MSA (4296). Meta committed $14.2 billion through December 2031 (4282) and then about $21 billion through December 2032 under a March 31, 2026 order form (2123, 10340). Several claims repeat that $21 billion figure (10694, 10729, 17654, 9317) and should not be counted twice. Jane Street committed about $6 billion (10304), and NVIDIA signed a capacity order form (4352). Anthropic is described as a signed multi-year agreement (9318), but no value is disclosed. Everything else is softer. That includes the NVIDIA 'collaboration framework' and the 5 GW-by-2030 ambition (3892, 17596, 27948), the IBM partnership (4971), CrowdStrike's 'global partnership' (11768, 10330), the Aston Martin sponsorship (25576), Genesis Mission participation (11778), the Core Scientific 'collaborate on shared opportunities' language (20313), and logo lists with no amounts (25575, 11767, 9320). Unidentified counterparties remain. The 'new hyperscaler customer' (25574) and the 'leading AI enterprise' behind DDTL 4.0 (11224) are never named. The DDTL 3.0 contract (4595) was tied to OpenAI only by a separate disclosure (26687). Concentration has moved from a single dominant AI-lab relationship toward a two-pillar book. By 2026 the named commitments are roughly $35 billion from Meta and $22 billion from OpenAI, with Jane Street and Anthropic as a smaller second tier. The arithmetic partly reconciles. OpenAI's first deal is $11.9 billion in the announcement (19820) but $11.2 billion in backlog (4969), a $0.7 billion gap that is not explained. Reported RPO was $15.1 billion at December 2024 and $14.7 billion at March 2025 (3572, 4970), which suggests the OpenAI award first sat in non-RPO backlog. RPO of $60.7 billion at December 2025 (3555) roughly fits OpenAI's $22.4 billion plus Meta's $14.2 billion plus the legacy book, which presumably includes Microsoft, a name absent from this record. Meta's $21 billion and Jane Street's $6 billion came after that date. They imply RPO well above $80 billion before revenue recognition, but no later RPO figure appears in the claims provided, which end mid-record at claim 9319.

CoreWeave's signed-contract core is narrower than its partnership headlines suggest. Two of its counterparties are documented at order-form level under existing master services agreements: OpenAI (4296, 19820, 25573, 20198) and Meta (4282, 2123). NVIDIA's capacity purchase commitment is also an order form (4352), which later appears as a $6.3 billion 'strategic collaboration' (20209). Because NVIDIA is also CoreWeave's key supplier, that contract reads more like a capacity backstop than independent end demand. Jane Street's roughly $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) are described as signed, but no dollar value has been disclosed for Anthropic. Much of the rest is softer language. The NVIDIA 'collaboration framework' and the 5 GW buildout (3892, 17596, 27948) carry no stated payment terms, and neither does the reference-architecture validation (17598). The same is true of the IBM (4971), CrowdStrike (11768, 10330), Aston Martin (25576), Core Scientific (20313) and Genesis Mission (11778) partnerships. The logo lists (25575, 11767, 9320) name customers but attach no values. Several material counterparties went unidentified when disclosed. These include the 'new hyperscaler customer' (25574) and the DDTL 3.0 'customer contract' (4595), which was only later tied to OpenAI (26687). The DDTL 4.0 'leading AI enterprise' (11224) remains unnamed. Concentration has shifted from one dominant customer to two. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 31, 2025 (4969, 4970). By year-end, OpenAI's up to $22.4 billion plus Meta's up to $14.2 billion equaled roughly 60% of the $60.7 billion in RPO (3555). Meta's further $21 billion in March 2026 (10340, 9317) deepens that pairing. Jane Street and Anthropic diversify at the margin. Repeated coverage of the same Meta deal (10729, 10694, 17654, 17655) inflates the apparent number of announcements, and one item dated before the March 31 order form is effective suggests metadata error. The arithmetic only partly reconciles. The OpenAI deal was announced at $11.9 billion but booked as $11.2 billion of backlog (19820 vs. 4969). Reported RPO slipped from $15.1 billion (3572) to $14.7 billion (4970) even as backlog rose, which places most of the OpenAI value in a non-RPO bucket that depends on delivery conditions. The 2025 RPO increase of $45.6 billion broadly matches the named 2025 adds of about $42.9 billion, but that total relies on 'up to' figures and an unnamed hyperscaler. Read the reconciliation as directional, not exact.

CoreWeave's signed, dollar-denominated contracts come from a small number of named counterparties. OpenAI is the clearest. The March 2025 agreement was announced at $11.9 billion (claim 19820), although only $11.2 billion entered backlog (claim 4969). It was followed by a $4 billion expansion (claim 25573) and an up-to-$6.5 billion order. Management gave the cumulative figure as up to roughly $22.4 billion (claim 20198), and that total reconciles exactly with the three tranches. The September 2025 order form under the existing MSA (claim 4296) shows these are executed documents, not letters of intent. Meta is the second pillar. Its September 2025 order form committed up to about $14.2 billion through December 2031 (claims 4282, 4283), and a March 31, 2026 order form added roughly $21 billion through 2032 (claims 2123, 10340, 9317). Readers should note that some secondary coverage blurs whether the $21 billion is incremental or a restatement (claims 17654, 17655, 10729); the filing language supports incremental. The NVIDIA $6.3 billion capacity order (claims 4352, 20209) is also a signed order form. Because NVIDIA is simultaneously CoreWeave's supplier, the circularity deserves scrutiny. Anthropic (claim 9318) is described as signed but carries no disclosed value. Jane Street's roughly $6 billion (claim 10304) rests largely on press characterization (claims 10730, 10695).

The softer layer is extensive. The NVIDIA "collaboration framework" toward 5 GW (claims 3892, 17596, 27948) and the reference-architecture validation (claim 17598) are non-binding. So are the IBM partnership (claim 4971), Core Scientific collaboration (claim 20313), CrowdStrike global partnership (claims 11768, 10330), Genesis Mission (claim 11778) and the Aston Martin sponsorship (claim 25576). Logo lists (claims 25575, 11767, 9320) name counterparties without values. Several material counterparties remain unidentified: the "new hyperscaler customer" (claim 25574) and the "leading AI enterprise" behind DDTL 4.0 (claim 11224). The DDTL 3.0 customer (claim 4595) was later disclosed as OpenAI (claim 26687).

On arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), an increase of about $45.6 billion. The named 2025 signings explain most of it. OpenAI's roughly $22.4 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion total about $42.9 billion. The unnamed hyperscaler and smaller logos plausibly cover the remainder, net of recognized revenue, so the bridge broadly holds. The same math also implies roughly 70 percent of year-end RPO sits with three counterparties. Meta's March 2026 addition and Jane Street have since diversified the book only modestly.

CoreWeave's signed, named contract base is narrower than its announcement flow suggests. Three counterparties have executed order forms under existing Master Services Agreements. OpenAI signed for $11.9 billion (claim 19820), expanded by $4 billion (claim 25573), and signed a September 2025 order form (claim 4296) worth up to $6.5 billion, for a stated total of up to $22.4 billion (claim 20198). Meta signed for up to $14.2 billion through 2031 (claims 4282, 20197) and then signed a further order form of roughly $21 billion on March 31, 2026 (claims 2123, 10340, 9317). NVIDIA signed a capacity order form (claim 4352) valued at $6.3 billion (claim 20209). Jane Street's roughly $6 billion commitment (claims 10304, 10730) and the Anthropic agreement (claims 10731, 9318) are described as committed or signed, but no filed order form is on record here, and Anthropic carries no value. Much of the remaining material is softer. The 2026 NVIDIA arrangement is expressly a collaboration framework (claim 3892) with aspirational gigawatt targets and 'work towards' language (claims 27948, 17598). IBM, CrowdStrike, Aston Martin, Genesis Mission and Core Scientific are partnerships or affiliations without disclosed economics (claims 4971, 11768, 25576, 11778, 20313). The logo lists in claims 25575 and 11767 carry no contract values. Several counterparties were initially unidentified. The DDTL 3.0 'customer contract' (claim 4595) was later identified as OpenAI (claim 26687). The 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' behind DDTL 4.0 (claim 11224) remain unnamed. Concentration has shifted from a single customer to two. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 2025 (claims 4969, 4970). OpenAI, Meta and NVIDIA together total up to about $42.9 billion against $60.7 billion of RPO at year-end 2025 (claim 3555), roughly 70%. Meta has since added $21 billion. The arithmetic only loosely reconciles. OpenAI's $11.9 billion headline does not match the $11.2 billion booked to backlog. The headline values are 'up to' ceilings. The jump from $30.1 billion of backlog (claim 25586) to $60.7 billion of RPO compares two differently defined measures (claim 4970). Signed deals between the two dates account for roughly $27 billion of the $30.6 billion increase. Claim 10729 is dated before the March event it describes, and the record is truncated at claim 9318.

CoreWeave's disclosed book rests on a small number of signed, named order forms, surrounded by a larger halo of partnership and logo language that carries no stated dollar value. The documented contracts are executed order forms under existing MSAs. OpenAI's grew from $11.9 billion (claim 19820) to $4 billion more (claim 25573) and then up to $6.5 billion more, for up to about $22.4 billion in total (claims 20198, 4296). Meta's began at $14.2 billion through 2031 (claim 4282) and added a further $21 billion order form dated March 31, 2026 (claims 2123, 9317). NVIDIA's capacity order form is valued at $6.3 billion (claims 4352, 20209). Jane Street's roughly $6 billion commitment (claim 10304) and the Anthropic multi-year agreement (claim 9318) are named and described as signed, but Anthropic's value is undisclosed. The IBM 'partnership' (claim 4971), the NVIDIA 'collaboration framework' with 'work towards' language (claims 3892, 17598), CrowdStrike (claim 11768), the Genesis Mission (claim 11778), Aston Martin sponsorship (claim 25576) and the post-merger Core Scientific collaboration (claim 20313) are framework or relationship statements and should not be read as backlog. Customer lists (claims 25575, 11767) name logos but attach no amounts. Several counterparties were never identified. A 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' financed by DDTL 4.0 (claim 11224) remain unnamed. The DDTL 3.0 customer (claim 4595) was later identified as OpenAI (claim 26687). Concentration has shifted from a single-name dependency to a duopoly rather than diversifying. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 2025 (claims 4969, 4970). By December 2025, OpenAI's up-to $22.4 billion and Meta's $14.2 billion equaled roughly 60% of $60.7 billion in RPO (claim 3555), or about 71% once NVIDIA is included. The March 2026 Meta order lifts Meta's stated commitments to roughly $35 billion. The arithmetic broadly reconciles, with caveats. RPO of $15.1 billion (claim 3572) plus the OpenAI deal is close to the March backlog, but the company recorded $11.2 billion against a headline of $11.9 billion. The $4.2 billion rise to $30.1 billion (claim 25586) tracks the $4 billion expansion. The step to $60.7 billion is consistent with the $6.5 billion OpenAI, $14.2 billion Meta and $6.3 billion NVIDIA additions plus smaller wins, less revenue burn. Three cautions apply. The 'up to' figures are ceilings. Backlog includes amounts contingent on delivery (claim 4970), unlike RPO. Some secondary sources (claims 17654, 17655) blur whether Meta's $21 billion supersedes or adds to the $14.2 billion; the order-form disclosure (claim 2123) indicates it is additive.

CoreWeave's record splits cleanly between executed order forms and softer partnership language, and the executed column is dominated by two names. The firmest disclosures are 8-K order forms under existing master services agreements: OpenAI (claim 4296), Meta (claims 4282, 2123) and NVIDIA (claim 4352). OpenAI's stack runs from $11.9 billion (claim 19820) through a $4 billion expansion (claim 25573) and an up-to-$6.5 billion add-on to a stated total of up to roughly $22.4 billion (claim 20198), and those figures sum correctly. Meta's initial up-to-$14.2 billion commitment through 2031 (claim 4282) was followed by a roughly $21 billion order form signed March 31, 2026 (claims 2123, 10340), bringing named Meta commitments to about $35 billion. Jane Street's roughly $6 billion (claim 10304) is described as a commitment. Anthropic is called a signed multi-year agreement (claim 9318), but no value is disclosed. Note that 'up to' language on both OpenAI and Meta frames these as ceilings, not floors. Several items are partnerships rather than revenue contracts. These include IBM (claim 4971), CrowdStrike (claim 11768), Aston Martin's sponsorship-style designation (claim 25576), Genesis Mission participation (claim 11778), and the January 2026 NVIDIA 'collaboration framework' (claims 3892, 27948), whose operative verbs are 'test,' 'validate' and 'work towards' (claim 17598). Three counterparties went unnamed when first disclosed: a 'new hyperscaler customer' (claim 25574), a DDTL 3.0 'customer contract' (claim 4595), which was later identified as OpenAI (claim 26687), and a 'leading AI enterprise' financed through DDTL 4.0 (claim 11224), which remains unidentified. The logo lists (claims 25575, 11767) carry no dollar values. Concentration has shifted rather than diluted: OpenAI anchored 2025, while Meta's two tranches now exceed OpenAI's total. NVIDIA's dual role as supplier and $6.3 billion collaborator (claim 20209) adds circularity worth flagging. On arithmetic, the first OpenAI deal was announced at $11.9 billion but added $11.2 billion to backlog (claim 4969), an unexplained $0.7 billion gap. RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3555). The named 2025 signings (OpenAI $22.4 billion, Meta $14.2 billion, NVIDIA $6.3 billion) total about $42.9 billion. That is broadly consistent with the increase once 2025 revenue burn is netted, but it leaves no visible room for the unnamed hyperscaler unless the 'up to' amounts are not fully booked. The 2026 Meta, Jane Street and Anthropic deals postdate that RPO figure and await the next disclosure.

CoreWeave's contracted book rests on a small number of order forms executed under master services agreements, and these should be read separately from its partnership and logo announcements. The binding instruments on record are the OpenAI order forms (claims 19820, 4296), the Meta order forms of September 2025 and March 31, 2026 (claims 4282, 2123), and an NVIDIA capacity order form (claim 4352). Jane Street's roughly $6 billion commitment (claim 10304) and the Anthropic agreement, described as signed and multi-year but without a value (claim 9318), are also customer commitments. Several headline figures are framed as 'up to' amounts: OpenAI's cumulative $22.4 billion (claim 20198) and Meta's initial $14.2 billion (claim 4282). Meta's expansion beyond its initial commitment is an option, not an obligation (claim 4283). The IBM, Aston Martin, CrowdStrike, Core Scientific and Genesis Mission items (claims 4971, 25576, 11768, 20313, 11778) are partnership or affiliation language with no disclosed economics. The January 2026 NVIDIA 'collaboration framework' and its 5 GW buildout goal (claims 3892, 27948) are directional, not contracted revenue.

Some counterparties remain unidentified. These include the 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' behind DDTL 4.0 (claim 11224). The DDTL 3.0 'customer contract' (claim 4595) was later attributed to OpenAI (claim 26687). The logo rosters (claims 25575, 11767) name firms but attach no contract values. The Meta $21 billion expansion appears in at least seven records (claims 10729, 10694, 10340, 17654, 9317, 2123 and others). These describe one event and should not be counted more than once.

Concentration has shifted, but it has not widened. Through 2025, OpenAI, Meta and NVIDIA together account for about $42.9 billion of named commitments, roughly 70% of the $60.7 billion in RPO (claim 3555). OpenAI alone is about 37% of that figure. The March 2026 Meta order lifts Meta's named total to about $35 billion, overtaking OpenAI. Diversification therefore comes mainly from Jane Street and the unpriced Anthropic deal.

The arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3560), an increase of $45.6 billion. That is consistent with the $42.9 billion in named deals plus the unnamed hyperscaler, less revenue recognized during the year. Two caveats apply. First, OpenAI's $11.9 billion headline was booked as $11.2 billion of backlog (claims 19820, 4969). Second, RPO slipped to $14.7 billion at March 31, 2025, even after that deal (claim 4970). The 'up to' framing implies that realized revenue could fall below the headline figures.

CoreWeave's signed-contract core is narrower than its announcement volume suggests. Only a handful of deals are documented as executed order forms under master services agreements. OpenAI's are the original $11.9 billion contract (19820), a September 2025 order form (4296), and expansions of $4 billion (25573) and up to $6.5 billion (20198), which management totals at up to $22.4 billion. Meta's are an up-to $14.2 billion order form (4282) and a March 31, 2026 order form of roughly $21 billion (2123, 10340). NVIDIA has a capacity order form (4352), presented as a $6.3 billion collaboration (20209), and NVIDIA is also CoreWeave's chip supplier. Jane Street's roughly $6 billion commitment (10304) and the Anthropic multi-year agreement (9318), which carries no stated value, are described as signed but are not supported by filed order forms in this record. The rest is softer. The NVIDIA 'collaboration framework' (3892) and its aim to 'work towards' reference-architecture inclusion (17598) are statements of intent, not contracts. The same is true of the IBM (4971), CrowdStrike (11768), Core Scientific (20313), Aston Martin sponsorship (25576) and Genesis Mission (11778) items. Several material counterparties go unnamed: a 'new hyperscaler customer' (25574), a DDTL 3.0 'customer contract' (4595) that later disclosure ties to OpenAI (26687), and a 'leading AI enterprise' behind DDTL 4.0 (11224). The logo lists (25575, 11767, 9320) carry no dollar values, and the February and April lists repeat each other (10338). The same deals are also re-announced: the Meta $14.2 billion deal appears three times (4282, 20197, 17655), the $21 billion deal four times (10729, 10694, 10340, 17654), and Jane Street three times (10730, 10304, 10695). On concentration, OpenAI's $22.4 billion and Meta's $14.2 billion together come to about 60% of the $60.7 billion in remaining performance obligations at year-end 2025 (3555). Meta's March 2026 addition, if it is additive, would make Meta the largest named counterparty. On the arithmetic, RPO rose from $15.1 billion (3572) to $60.7 billion. The named 2025 deals add about $42.9 billion, which reaches roughly $58 billion before any 2025 revenue is subtracted from the opening balance. That leaves a gap of several billion dollars that is likely filled by the unnamed hyperscaler and unpriced logos. Several smaller inconsistencies remain. OpenAI's backlog contribution is given as $11.2 billion (4969) against an $11.9 billion contract (19820). The March 2025 backlog figure of $25.9 billion (4970) includes amounts beyond RPO that depend on delivery conditions. 'Up to' ceilings may overstate what is firmly committed. The record supplied ends mid-claim at 9319, so later expansions could not be assessed.

CoreWeave's record contains a small number of genuinely signed, filed customer contracts surrounded by a much larger volume of partnership and logo language. The hard contracts are the OpenAI order forms under an existing MSA: $11.9 billion (19820), a $4 billion expansion (25573), and up to $6.5 billion more, which brings the total to up to $22.4 billion (20198, 4296). The arithmetic checks out, although the Q1 deck booked the first deal at $11.2 billion of backlog rather than $11.9 billion (4969). Meta's order forms are similarly firm: up to $14.2 billion through December 2031 (4282, 20197), then an approximately $21 billion expansion through 2032 (2123, 10340). Claims 10729, 10694 and 17654 restate that same $21 billion deal and should not be counted again. The NVIDIA $6.3 billion order form (4352, 20209) is also contractual. NVIDIA is a supplier and strategic counterparty, however, and the January 2026 'collaboration framework' and 5 GW buildout language (3892, 17596, 27948) are framework commitments, not revenue contracts. Jane Street's roughly $6 billion (10304) is quantified. Anthropic's 'multi-year agreement' (10731) carries no value.

Several counterparties go unnamed. The 'new hyperscaler customer' (25574) is never identified. The DDTL 3.0 'customer contract' (4595) is later attributed to OpenAI (26687). The 'leading AI enterprise' behind DDTL 4.0 (11224) remains anonymous. Logo rolls such as BT, Cohere, Mistral, Cursor, Perplexity and others (25575, 11767, 9320, 9319) disclose no terms. IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), Aston Martin (25576) and the DOE Genesis Mission (11778) read as partnership, sponsorship or affiliation language rather than quantified backlog.

Concentration has broadened in name more than in dollars. OpenAI dominated in early 2025, and Meta became a co-anchor by Q3. Taken together, OpenAI's $22.4 billion and Meta's $14.2 billion equal about 60% of the $60.7 billion in RPO at year-end 2025 (3555). Adding NVIDIA raises that share to roughly 70%. The 2026 additions of Meta's $21 billion and Jane Street's $6 billion deepen the Meta exposure while adding only one other quantified name.

The revenue arithmetic broadly reconciles. RPO rose $45.6 billion from $15.1 billion (3572) to $60.7 billion. The named 2025 deals total about $42.9 billion, which leaves only a few billion for the hyperscaler and smaller customers before netting out revenue recognized during the year. That fit is tight enough to suggest the 'up to' figures are being booked close to their face value. Investors should therefore read them as ceilings that depend on delivery, a condition the backlog definition makes explicit (4970). The record as supplied ends mid-claim at 9324, so later items could not be assessed.

CoreWeave's record separates into a small set of filed order forms and a much larger layer of partnership and logo language. The contracts with filing-level support are order forms under existing master services agreements with OpenAI (claim 4296), Meta (claims 4282, 2123) and NVIDIA (claim 4352). The OpenAI tranches total $11.9 billion, $4 billion and $6.5 billion, which sum to the stated $22.4 billion (claims 19820, 25573, 20198). Meta's commitment is $14.2 billion plus roughly $21 billion (claims 4282, 10340, 9317). Both are framed as 'up to' amounts, so they are ceilings rather than floors. Jane Street's roughly $6 billion (claim 10304) is named and sized but rests on press language rather than a cited order form. The Anthropic agreement (claims 10731, 9318) is named but has no disclosed value. The January 2026 NVIDIA 'collaboration framework' toward more than 5 GW by 2030 (claims 3892, 27948) is framework language, not a purchase commitment. The same applies to the Exemplar Cloud designation (claim 10332), the IBM and CrowdStrike 'partnerships' (claims 4971, 11768), the Aston Martin sponsorship (claim 25576), Genesis Mission participation (claim 11778) and the Core Scientific collaboration (claim 20313). The logo lists (claims 25575, 11767) name counterparties without amounts. The DDTL 4.0 lenders (claim 11225) are financiers, not customers. Three references stay anonymous: a 'new hyperscaler customer' (claim 25574), the DDTL 3.0 'customer contract' (claim 4595), which was later identified as OpenAI (claim 26687), and the 'leading AI enterprise' financed by DDTL 4.0 (claim 11224). Concentration has shifted rather than eased. In early 2025, OpenAI's $11.2 billion was roughly 43% of the $25.9 billion backlog (claims 4969, 4970). By year-end, OpenAI's and Meta's ceilings together equal about 60% of the $60.7 billion in RPO (claim 3555). Meta's March 2026 expansion now makes it the largest named commitment. The arithmetic broadly holds, with caveats. The OpenAI deal was announced at $11.9 billion but booked as $11.2 billion of backlog, a $0.7 billion gap that is never explained. RPO dipped from $15.1 billion to $14.7 billion between December 2024 and March 2025 even though the OpenAI deal was signed in that window (claims 3572, 4970). This suggests the OpenAI deal largely sat in the non-RPO portion of backlog. Starting from the $30.1 billion mid-2025 backlog (claim 25586) and adding the OpenAI, Meta and NVIDIA deals gives about $57 billion, close to the $60.7 billion figure, but that comparison mixes backlog and RPO definitions. Several claim values (claims 10729, 10730, 9317) are stored as '$21,000' or '$6,000' and should be read as billions.

CoreWeave's record splits into a small number of signed, quantified order forms and a much larger set of partnership and logo-list language. The signed agreements are with OpenAI and Meta. OpenAI's are a $11.9 billion agreement (claim 19820), a $4 billion expansion (25573), and an up-to $6.5 billion add-on reaching up to $22.4 billion in total (20198), papered as an order form under an existing MSA (4296). Meta's are a $14.2 billion order form through 2031 (4282) and a roughly $21 billion order form dated March 31, 2026 (2123, 10340). Jane Street's roughly $6 billion commitment (10304) is also quantified. The Anthropic agreement is described as signed and multi-year, but no value is disclosed (9318). Repeated filings restate the same Meta and Jane Street deals (10729, 10694, 10695) and should not be counted twice. NVIDIA is a special case. There is a real order form (4352) and a $6.3 billion collaboration (20209), but the January 2026 material is explicitly a 'collaboration framework' (3892) built around testing, validation and 'work towards' language (17598). IBM (4971), CrowdStrike (11768), Aston Martin (25576), Core Scientific (20313) and the Genesis Mission (11778) are partnerships with no disclosed economics. The named-customer lists (25575, 11767, 10338, 9319) give no deal values. Several material counterparties go unnamed: the 'new hyperscaler customer' (25574) and the 'leading AI enterprise' behind DDTL 4.0 (11224). The DDTL 3.0 customer (4595) is identified only through a separate press statement (26687). Concentration has broadened somewhat but remains heavy. At March 31, 2025, OpenAI's $11.2 billion made up about 43% of the $25.9 billion backlog (4969, 4970). The 4969 figure also differs from the $11.9 billion headline (19820), a gap that is never explained. By late 2025, OpenAI and Meta together accounted for up to $36.6 billion. On revenue arithmetic, the move from $25.9 billion to $30.1 billion of backlog (25586) roughly matches the $4 billion OpenAI expansion. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), an increase of about $45.6 billion. Named 2025 deals sum to roughly $42.9 billion including NVIDIA, before subtracting revenue recognized during the year. The gap therefore has to be filled by the unnamed hyperscaler and smaller customers. Many headline figures are 'up to' ceilings, so the reconciliation is plausible but not verifiable from the record. The 2026 Meta and Jane Street commitments, about $27 billion combined, have no subsequent RPO print on file against which to test them.

CoreWeave's announcements fall into two tiers. The first is a small set of signed, priced order forms under existing Master Services Agreements: OpenAI (claims 19820, 4296), Meta (4282, 2123) and NVIDIA's $6.3 billion capacity purchase (4352, 20209). The second is a larger body of partnership, framework and logo-list language that carries no disclosed value. The OpenAI arithmetic holds internally. The $11.9 billion original deal, the $4 billion expansion and the up-to-$6.5 billion expansion sum to the stated $22.4 billion (25573, 20198). One small gap remains: the first quarter credited only $11.2 billion of that original deal to backlog (4969). Meta's initial $14.2 billion runs through December 2031 (4282). The March 2026 order form adds roughly $21 billion through 2032 (10729, 10340). Some secondary phrasing reads the $21 billion as a total rather than an increment (17654), but the filing language supports roughly $35 billion in aggregate. Jane Street's roughly $6 billion (10304) is named and priced. Anthropic's 'multi-year agreement' (10731, 9318) is signed but carries no figure. Several counterparties have never been named. A 'new hyperscaler customer' (25574) has no identity or value. The DDTL 4.0 financing serves a 'leading AI enterprise' (11224) that is not named. DDTL 3.0's unnamed 'customer contract' (4595) was later tied to OpenAI (26687). The 2026 NVIDIA items are framework language and do not add revenue: the 'collaboration framework' (3892), the more-than-5 GW buildout ambition (17596, 27948) and the Exemplar Cloud designations (10332, 9324). The same applies to the IBM (4971) and CrowdStrike (11768) partnerships, the Genesis Mission (11778), Core Scientific (20313), and the logo lists (25575, 11767, 9320, 9319). On concentration, OpenAI was effectively the whole growth story in early 2025. By December 2025, OpenAI ($22.4 billion) and Meta ($14.2 billion) together accounted for about $37 billion of $60.7 billion in RPO (3555), roughly 60 percent. NVIDIA, which is also CoreWeave's key supplier, added $6.3 billion, so the three named counterparties cover about 70 percent. In 2026, Meta's expansion moves Meta toward parity with OpenAI rather than diluting the concentration. On reconciliation, $15.1 billion of opening RPO (3572) plus about $42.9 billion of named 2025 deals comes to about $58 billion before 2025 revenue recognition. That leaves a residual of several billion dollars that only the unnamed hyperscaler and the long tail can explain. The residual is plausible, but it is unverifiable from the record. The earlier backlog figures ($25.9 billion and $30.1 billion; 4970, 25586) also include non-RPO amounts that depend on delivery, so they are not directly comparable to the year-end RPO figure.

CoreWeave's signed-contract base is anchored by three counterparties. Each is documented through order forms under existing master services agreements. OpenAI's commitments stack from an initial $11.9 billion (claim 19820) through a $4 billion expansion (claim 25573) and an up-to-$6.5 billion order (claims 4296, 20198), for up to roughly $22.4 billion. That total matches the three tranches, but the first deal added only $11.2 billion to backlog (claim 4969), a $0.7 billion gap the filings do not explain. Meta's original order is likewise framed as up to $14.2 billion through 2031, with an expansion option (claims 4282, 4283). The March 31, 2026 order form for about $21 billion (claims 2123, 10340) converts that option into a signed commitment. NVIDIA's $6.3 billion capacity order (claims 4352, 20209) is contractual, but NVIDIA is also CoreWeave's supplier and strategic collaborator, so this revenue is partly circular. Jane Street's roughly $6 billion (claim 10304) is signed and named. Anthropic's multi-year agreement (claims 9318, 10731) is named but carries no disclosed value.

The rest of the record is softer. The NVIDIA collaboration framework and the more-than-5 GW buildout target (claims 3892, 17596, 27948) are framework language, not purchase obligations. The IBM, CrowdStrike, Core Scientific and Aston Martin items (claims 4971, 11768, 20313, 25576) are partnerships or sponsorships. Genesis Mission participation (claim 11778) is not commercial. The recurring logo lists (claims 25575, 11767, 9320, 9319) name customers but attach no dollar values. Two financing-linked counterparties were initially unnamed: the DDTL 3.0 customer (claim 4595) was later identified as OpenAI (claim 26687). The 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' behind DDTL 4.0 (claim 11224) remain unidentified.

Customer concentration has shifted from OpenAI alone to a duopoly. In early 2025, OpenAI's $11.2 billion accounted for most of the $14.7 billion of RPO (claim 4970). By December 31, 2025, RPO had reached $60.7 billion, up from $15.1 billion a year earlier (claims 3555, 3572). The up-to totals for OpenAI, Meta and NVIDIA sum to about $42.9 billion. Added to the year-earlier base and net of revenue burned off, that roughly reconciles to reported RPO. It also implies these three counterparties account for about 70 percent of RPO. That reconciliation depends on 'up to' ceilings being treated as firm. The 2026 Meta and Jane Street deals add about $27 billion on top, which deepens the Meta concentration rather than diversifying it. One data caveat: claim 10729's '$21,000' figure and its March 1 date appear to be extraction artifacts.

CoreWeave's record contains a core of signed, dollar-denominated customer contracts, surrounded by a larger body of partnership, framework and logo-list language that should not be read as backlog. The contracted core is OpenAI, Meta, NVIDIA and Jane Street. OpenAI's tranches reconcile: $11.9 billion (claim 19820), plus $4 billion (25573), plus up to $6.5 billion (20198), equals the stated cumulative "up to approximately $22.4 billion." There is one small gap. The first deal was described as adding $11.2 billion of backlog (4969) against an $11.9 billion headline. The September 2025 order form under the existing MSA (4296) confirms the deals are binding. Meta's initial order form commits "up to approximately $14.2 billion" through 2031 (4282). A March 31, 2026 order form adds roughly $21 billion through 2032 (2123, 10340). Some secondary sources date the expansion to March 1 (10729), which predates the filing, so that date looks like an ingestion artifact. Jane Street's roughly $6 billion (10304) and NVIDIA's $6.3 billion capacity order (4352, 20209) complete the dollar-valued set. Anthropic's "multi-year agreement" (10731, 9318) is signed but has no disclosed value. Several counterparties are left unnamed or identified only later. The DDTL 3.0 "customer contract" (4595) was later tied to OpenAI (26687). The "new hyperscaler customer" (25574) and the "leading AI enterprise" behind DDTL 4.0 (11224) remain anonymous despite apparent materiality. Much of the remaining record is non-contractual. This includes the NVIDIA "collaboration framework" (3892), the ">5 GW" buildout ambition (17596, 27948), and reference-architecture work the company is still "working towards" (17598). The IBM (4971), CrowdStrike (11768) and Core Scientific (20313) relationships carry no stated economics. Aston Martin (25576) is a sponsorship and the Genesis Mission (11778) is an initiative. Exemplar Cloud status is described as "first" (10332) in one source and "among the first" (9324) in another. Logo lists (25575, 11767, 9320) carry no values. Concentration has shifted from near-total OpenAI dependence to an OpenAI-Meta duopoly. Meta's potential roughly $35 billion now exceeds OpenAI's $22.4 billion, if the $21 billion is additive, which claims 17654 and 17655 leave ambiguous. NVIDIA also sits on both sides as supplier and customer. On the revenue arithmetic, RPO rose from $15.1 billion (3572) to $60.7 billion (3555), an increase of about $45.6 billion. The 2025-signed OpenAI, Meta and NVIDIA deals total about $42.9 billion. After netting out 2025 recognized revenue, which is not on this record, the remainder plausibly falls to the hyperscaler and smaller wins. The reconciliation is directionally sound but rests on "up to" ceilings rather than floors.

CoreWeave's signed, named contracts are concentrated in a small set of order forms under existing master services agreements. These are OpenAI (claims 19820, 4296), Meta (4282, 2123), and NVIDIA as a capacity purchaser (4352). Jane Street's roughly $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) carry signed-agreement language, but Anthropic's value is undisclosed. Much of the remaining record is partnership vocabulary with no economics attached. That includes IBM's Granite capacity (4971), the Aston Martin sponsorship (25576), CrowdStrike's "global partnership" (11768, 10330), continued collaboration with Core Scientific (20313), Genesis Mission participation (11778), and the January 2026 NVIDIA "collaboration framework" (3892, 27948). The recurring logo lists (25575, 11767, 9320) name counterparties but attach no contract values.

Several material counterparties go unidentified. The "new hyperscaler customer" signed in Q2 2025 is never named (25574). DDTL 3.0 finances "a customer contract" (4595), which a separate disclosure ties to OpenAI (26687). DDTL 4.0 serves a "leading AI enterprise" whose name is withheld (11224). Readers cannot attribute the backlog these facilities support.

Concentration has moved from a single-tenant story to a duopoly. Through 2025, OpenAI's commitments rose from $11.9 billion to up to $22.4 billion (19820, 25573, 20198), and those three figures sum cleanly. Meta's up-to $14.2 billion order (20197) added a second anchor. The March 2026 Meta expansion of about $21 billion (10729, 10340) appears incremental, which would put Meta near $35 billion and ahead of OpenAI. Claim 17654 could also be read as a $21 billion total, and that ambiguity should be resolved against the order form (2123). Jane Street and Anthropic diversify the book only modestly.

The revenue arithmetic broadly reconciles but has seams. OpenAI's initial deal was announced at $11.9 billion yet entered backlog at $11.2 billion (4969). RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555). Named 2025 additions total about $31 billion: OpenAI $10.5 billion, Meta $14.2 billion, and NVIDIA $6.3 billion (20209). The first OpenAI deal accounts for roughly $11 billion more. Together these approach the increase before recognized revenue is subtracted. Several of these amounts are "up to" ceilings, however, and backlog (4970) includes amounts beyond RPO that depend on delivery conditions. The record also has no post-March-2026 RPO figure against which to test the Meta, Jane Street, and Anthropic additions.

CoreWeave's record splits cleanly into contracts backed by order forms and relationships described in partnership vocabulary. The firm, signed commitments are OpenAI and Meta. OpenAI's $11.9 billion agreement (claim 19820) was followed by a $4 billion expansion (25573), a September 2025 order form under the existing MSA (4296), and an 'up to' $6.5 billion add-on lifting the total to 'up to approximately $22.4 billion' (20198). Meta's order form commits 'up to approximately $14.2 billion' through December 2031 (4282). A second Meta order form dated March 31, 2026 (2123) supports the roughly $21 billion expansion through 2032 (10340, 17654). Claim 10729 is dated 2026-03-01 but describes a March event, so that attribution should be treated as a date artifact. NVIDIA's $6.3 billion is a capacity-purchase order form (4352, 20209), which makes it a supplier acting as a buyer of last resort rather than an arm's-length end customer. The January 2026 NVIDIA 'collaboration framework' (3892, 17596, 27948), with its aspirational 'work towards' software inclusion (17598), carries no disclosed dollar value. The same is true of IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), Aston Martin (25576) and the Genesis Mission (11778). Jane Street's roughly $6 billion (10304) is named and sized. Anthropic is described as a signed multi-year agreement with no value attached (9318, 10731).

Several counterparties are not identified. The 'new hyperscaler customer' (25574) and the 'leading AI enterprise' behind DDTL 4.0 (11224) are both unnamed. The DDTL 3.0 'customer contract' (4595) is identified as OpenAI only in a separate release (26687). The logo lists (25575, 11767, 9320) name customers but attach no economics.

Concentration has risen. OpenAI's $11.2 billion was about 43% of $25.9 billion backlog at March 2025 (4969, 4970). By December 2025, OpenAI's $22.4 billion plus Meta's $14.2 billion came to roughly 60% of $60.7 billion RPO (3560), and about 71% once NVIDIA is included.

The arithmetic broadly reconciles, but it rests on ceilings. RPO rose $45.6 billion during 2025 (3555, 3572), against about $42.9 billion of named 2025 deals that are largely framed as 'up to' amounts. That leaves little room for the logo-list customers once 2025 revenue burn is considered. OpenAI's announced $11.9 billion against $11.2 billion of added backlog (19820, 4969) is a small unexplained gap. Backlog also includes non-RPO amounts that depend on delivery conditions (4970). The 2026 Meta, Jane Street and Anthropic additions have no updated RPO on record to test them against.

CoreWeave's signed contracts are concentrated in a few counterparties, and most of its other announcements carry partnership language rather than contract language. The agreements with the firmest evidence are filed order forms under existing master services agreements. These cover OpenAI (claim 4296), NVIDIA (claim 4352), and Meta in September 2025 and again in March 2026 (claims 4282, 2123). The OpenAI dollar figures reconcile across releases: the original $11.9 billion (claim 19820), the $4 billion expansion (claim 25573) and the up to $6.5 billion add-on (claim 20198) sum to the stated up to $22.4 billion. The first-quarter 2025 release, however, credited only $11.2 billion to backlog from the initial deal (claim 4969). That gap is unexplained. Meta's first commitment is likewise capped, "up to approximately $14.2 billion" (claim 4282), with expansion only as an option (claim 4283). The later $21 billion order (claims 10340, 9317) makes Meta, at roughly $35 billion combined, the largest disclosed counterparty, ahead of OpenAI. NVIDIA is harder to classify. The $6.3 billion capacity commitment (claim 20209) is contractual, but the January 2026 expansion is described as a "collaboration framework" (claims 3892, 17596) with no stated value. Below this tier the evidence thins. Anthropic is named as a signed multi-year agreement with no value attached (claims 9318, 10731). Jane Street's roughly $6 billion (claim 10304) is the only sizable non-hyperscaler figure. The logo lists of AI labs and enterprises (claims 25575, 11767) carry no economics. IBM (claim 4971), CrowdStrike (claim 11768), Aston Martin (claim 25576), the Genesis Mission (claim 11778) and Core Scientific (claim 20313) are partnership or sponsorship language and should not be counted as backlog. Several material counterparties are also unnamed. These include the "new hyperscaler customer" (claim 25574), the "customer contract" behind DDTL 3.0 (claim 4595), which a later release ties to OpenAI (claim 26687), and the "leading AI enterprise" financed by DDTL 4.0 (claim 11224). On the arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3555). The named 2025 additions are roughly $11.2 billion, $4 billion and $6.5 billion from OpenAI, $14.2 billion from Meta and $6.3 billion from NVIDIA. Together they come to about $42 billion, or about $57 billion with opening RPO, before deducting revenue recognized during the year. That leaves several billion dollars of RPO attributable to undisclosed counterparties, most plausibly the unnamed hyperscaler. The pattern also suggests that the "up to" ceilings are being booked at or near full value. The disclosures do not let a reader verify either point.

CoreWeave's record splits into a small set of signed, priced order forms and a much larger layer of partnership language. The signed core is concrete. OpenAI's commitments build from $11.9 billion (claim 19820) through a $4 billion expansion (claim 25573) to an up-to $6.5 billion order form (claims 4296, 20198), and the stated cumulative of up to $22.4 billion reconciles exactly. One wrinkle is that the Q1 2025 release booked only $11.2 billion of the initial deal into backlog (claim 4969). Meta's up-to $14.2 billion order form (claims 4282, 20197) was followed by a roughly $21 billion order form executed March 31, 2026 (claims 2123, 10340, 9317). Jane Street committed about $6 billion (claim 10304). NVIDIA's role is dual: it signed a capacity order form (claim 4352) and a $6.3 billion collaboration (claim 20209), but the January 2026 expansion is explicitly a 'collaboration framework' (claims 3892, 27948) with no stated dollar value. The Anthropic agreement (claim 10731) is named but unpriced.

Several counterparties are described without being named. These include a 'new hyperscaler customer' (claim 25574), a 'leading AI enterprise' behind DDTL 4.0 (claim 11224), and the DDTL 3.0 'customer contract' (claim 4595), although claim 26687 later identifies that last one as OpenAI. IBM, CrowdStrike, Aston Martin, Core Scientific and the Genesis Mission (claims 4971, 11768, 25576, 20313, 11778) are partnership or sponsorship language with no disclosed commitments. The long logo lists (claims 25575, 11767, 9320) establish usage, not contract scale.

Concentration has shifted rather than eased. Through 2025, OpenAI made up about $22.4 billion of roughly $42.9 billion in priced named commitments. Meta's March 2026 order form brings its cumulative total to about $35 billion, making it the largest named counterparty. Jane Street is the only new eight-figure-plus name with a disclosed value.

The revenue arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), an increase of about $45.6 billion. The named OpenAI, Meta and NVIDIA deals account for about $42.9 billion of that. The residual plausibly reflects the unnamed hyperscaler and smaller logos, net of revenue recognized. Several caveats limit this reading. Much of the headline value is phrased as 'up to' amounts (claims 4282, 20198), which are ceilings rather than floors. 'Backlog' (claims 4970, 25586) includes amounts beyond RPO that depend on delivery conditions. Some claim values appear unit-garbled ($21,000 in claim 10729, $61 in claim 3555). The weighted five-year duration (claim 3556) also means only a fraction of this converts to near-term revenue.

CoreWeave's record separates more cleanly than most AI-infrastructure names into contracts executed under master services agreements and a larger set of partnership or logo claims that carry no disclosed economics. The firm contracts are the OpenAI order forms, totalling up to roughly $22.4 billion across $11.9 billion, $4 billion and $6.5 billion tranches (19820, 25573, 4296, 20198), and the Meta order forms, an initial up to $14.2 billion through December 2031 (4282, 17655) followed by a March 31, 2026 order form (2123) valued at about $21 billion (10340). The NVIDIA September 2025 order form (4352) and Jane Street's roughly $6 billion commitment (10304) also appear to be signed. Anthropic is described as a signed multi-year agreement (9318), but no value is given. By contrast, the January 2026 NVIDIA 'collaboration framework' (3892, 17596) is aspirational, as its 'test and validate' and 'work towards' language shows (17598). IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), the Aston Martin sponsorship (25576) and the Genesis Mission (11778) are also partnerships rather than disclosed revenue contracts, and the rotating logo lists (25575, 11767, 9320) attach no dollar figures to any name. Several material counterparties have been left anonymous. These include the 'new hyperscaler customer' (25574) and the 'leading AI enterprise' financed by DDTL 4.0 (11224). The 'customer contract' behind DDTL 3.0 (4595) was identified as OpenAI only in later materials (26687). Concentration remains acute. The named OpenAI, Meta and NVIDIA commitments of about $42.9 billion amount to roughly 70% of the $60.7 billion of December 2025 RPO (3555), before netting any 2025 recognition. The 2026 additions from Meta, Jane Street and Anthropic deepen exposure to the top two as much as they diversify it. The revenue arithmetic does not fully reconcile. OpenAI's initial deal was announced at $11.9 billion (19820) but booked as $11.2 billion of backlog (4969). March 2025 RPO of $14.7 billion (4970) sat below December 2024's $15.1 billion (3572), which implies the OpenAI amount initially sat largely in the conditional, delivery-dependent backlog bucket rather than in RPO. Backlog rose only $4.2 billion in Q2 (25586), roughly the size of the OpenAI expansion alone, which leaves the hyperscaler win with little visible contribution. Sources also conflict on whether Meta's $21 billion is incremental (10729) or a total commitment (17654), and the pervasive 'up to' phrasing means the headline figures are ceilings, not floors.

CoreWeave's record contains more executed paper than most AI-infrastructure names, but the size of each deal is usually stated as a ceiling. OpenAI is backed by an order form under an existing MSA (4296), building from $11.9 billion (19820) through a $4 billion expansion (25573) and an "up to" $6.5 billion tranche to "up to approximately $22.4 billion" (20198). Those tranches sum exactly. The only seam is that the first quarterly report booked $11.2 billion of backlog against the $11.9 billion headline (4969). Meta is likewise documented by order forms, first "up to approximately $14.2 billion" through 2031 (4282, 20197) and then a March 31, 2026 order form (2123) worth about $21 billion through 2032 (10340). Secondary coverage disagrees on whether the $21 billion is incremental (10729) or the new total (17654). Claim 10729 also carries a March 1 date that precedes the order form it describes, so it is probably mis-dated. NVIDIA's $6.3 billion capacity purchase is an order form (4352, 20209). Its January 2026 "collaboration framework" (3892, 17596) and reference-architecture work (17598), however, are framework language with no stated consideration.

The softer layer is large. Jane Street's roughly $6 billion (10304, 10730) and the Anthropic "multi-year agreement" (9318, 10731) are disclosed only through press releases and summaries, and Anthropic carries no value. IBM (4971), CrowdStrike (11768), Aston Martin (25576), Core Scientific (20313) and the Genesis Mission (11778) are partnership, sponsorship or program-participation language rather than revenue contracts. The logo lists (25575, 11767, 9320, 9319) name customers without any economics. Some counterparties remain unidentified. The "new hyperscaler customer" (25574) has never been named, and the "leading AI enterprise" behind DDTL 4.0 (11224) is anonymous. The DDTL 3.0 "customer contract" (4595) was later attributed to OpenAI (26687).

Concentration has diversified in form but not much in substance. In Q1 2025, OpenAI made up about 43% of the $25.9 billion backlog (4969, 4970). By December 2025, OpenAI's $22.4 billion and Meta's $14.2 billion together amounted to roughly 60% of $60.7 billion of RPO (3555), before any revenue already recognized against them is subtracted. Meta's expansion likely makes it the largest single exposure.

The revenue arithmetic broadly holds. RPO rose $45.6 billion in 2025 (3572, 3555). Named 2025 additions total about $42.9 billion, which leaves room for the unnamed hyperscaler and smaller logos after 2025 revenue is netted out. Two caveats apply. Backlog includes amounts beyond RPO that depend on delivery (4970). Disclosed figures are also "up to" ceilings with a roughly five-year weighted duration (3556), so they are not committed minimums. No post-December RPO figure is on record yet to confirm the 2026 Meta, Jane Street and Anthropic additions.

CoreWeave's record has a few large, documented contracts surrounded by many unquantified partnership announcements. The deals that rest on executed instruments are OpenAI, Meta and NVIDIA. OpenAI began with an $11.9 billion agreement (claim 19820), although the Q1 letter booked only $11.2 billion of it as backlog (claim 4969). It then grew by $4 billion (claim 25573) and up to $6.5 billion more under an order form against the existing MSA (claims 4296, 20198), for a stated total of up to $22.4 billion. Meta's initial commitment of up to $14.2 billion through December 2031 (claim 4282) was followed by a March 31, 2026 order form worth roughly $21 billion through 2032 (claims 2123, 10340). The record is unclear whether that $21 billion is incremental or a restated total: claim 10729 says Meta expanded by about $21 billion, while claim 17654 says Meta committed $21 billion in total. NVIDIA's $6.3 billion capacity order (claims 4352, 20209) is a signed order form, but it comes from CoreWeave's key supplier, which makes it partly circular. NVIDIA's 2026 'collaboration framework' (claims 3892, 17596) is framework language and carries no dollar figure.

Several other counterparties are named but not quantified. Anthropic is described as signed and multi-year with no value attached (claims 9318, 10731). Jane Street's roughly $6 billion (claim 10304) is the only mid-sized deal with a disclosed figure. IBM, CrowdStrike, Aston Martin, Core Scientific and the Genesis Mission (claims 4971, 10330, 25576, 20313, 10331) read as partnerships rather than revenue contracts. The logo lists (claims 25575, 11767, 9320) have no economics attached. Some customers are not identified at all. The 'new hyperscaler customer' (claim 25574) is unnamed, and so is the 'leading AI enterprise' behind DDTL 4.0 (claim 11224). DDTL 3.0's anonymous 'customer contract' (claim 4595) is identified as OpenAI elsewhere (claim 26687).

Customer concentration has moved from OpenAI-dominant in 2025 to an OpenAI–Meta duopoly. On these figures the two account for roughly $36 billion of disclosed commitments by year-end 2025, and Meta's 2026 order may have pushed it ahead of OpenAI. Jane Street and Anthropic diversify the base only at the margin.

The revenue arithmetic loosely reconciles. RPO went from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), an increase of about $45.6 billion. The 2025 named additions of OpenAI's $15.6 billion incremental after Q1 plus its earlier $11.2 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion sum to about $47 billion before revenue recognized during the year, so the unnamed hyperscaler and smaller logos must cover the gap. Two caveats apply. Much of this headline value is 'up to' language, which sets ceilings rather than floors. And backlog of $25.9 billion was reported against RPO of $14.7 billion (claim 4970), which shows that management's backlog figure includes amounts that are not yet contractually enforceable. Note also that the claim record supplied was truncated after claim 9324.

CoreWeave's headline commitments rest mostly on executed paper. OpenAI's original $11.9 billion agreement (claim 19820) was followed by a $4 billion expansion (claim 25573) and an order form under an existing master services agreement (claim 4296), which management sized at up to $6.5 billion for a cumulative up to $22.4 billion (claim 20198). Meta's $14.2 billion order form (claim 4282) and its March 31, 2026 order form, reported at roughly $21 billion (claims 2123, 10340), are likewise 8-K-documented. NVIDIA's $6.3 billion capacity order (claims 4352, 20209) is a contract. The January 2026 'collaboration framework' (claims 3892, 17596, 17598) is not one. Anthropic is described as signed but carries no value (claim 9318). Jane Street's roughly $6 billion appears only in press language (claims 10304, 10730). IBM (claim 4971), CrowdStrike (claim 11768), Core Scientific (claim 20313), Aston Martin (claim 25576) and the Genesis Mission (claim 11778) are partnership or sponsorship language with no disclosed economics. The logo rosters (claims 25575, 11767, 9320) name customers but not commitments. Several counterparties stay anonymous: the 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' behind DDTL 4.0 (claim 11224). The DDTL 3.0 'customer contract' (claim 4595) is resolved as OpenAI by claim 26687.

Concentration has shifted from one name to two. At March 31, 2025, OpenAI's $11.2 billion (claim 4969) was about 43% of $25.9 billion of backlog (claim 4970). By December 31, 2025, OpenAI, Meta and NVIDIA together account for about $42.9 billion of announced value against $60.7 billion of RPO (claim 3560), roughly 70%. Adding Meta's March 2026 tranche makes Meta, at about $35 billion cumulative, the largest single exposure.

The arithmetic broadly reconciles, with caveats. RPO of $14.7 billion plus the $11.2 billion OpenAI figure equals the $25.9 billion backlog exactly (claim 4970). That implies the OpenAI deal initially sat outside RPO, pending delivery conditions. The $11.9 billion versus $11.2 billion gap between claims 19820 and 4969 goes unexplained. RPO rose $45.6 billion during 2025 (claims 3555, 3572). Named additions cover most of that, but full-year revenue recognized is not on record, so the residual attributable to unnamed customers cannot be isolated. Readers should note that the OpenAI and Meta figures are framed as 'up to' ceilings (claims 20197, 20198). Claim 10729 is dated March 1, 2026 yet describes a March expansion, which suggests a misdated record rather than a separate deal.

CoreWeave's record separates fairly cleanly into contracts with executed paper and relationship language with no stated economics. The signed tier rests on order forms under existing master services agreements: NVIDIA (claim 4352), OpenAI (claim 4296), and Meta (claims 4282 and 2123). The headline OpenAI and Meta figures trace to these instruments, and the OpenAI arithmetic reconciles. The $11.9 billion initial deal (claim 19820), the $4 billion expansion (claim 25573) and the roughly $6.5 billion add-on (claim 20198) sum to the disclosed $22.4 billion. One wrinkle is that the first-quarter backlog credited OpenAI with only $11.2 billion (claim 4969) against the $11.9 billion announced. Meta is less tidy. The original commitment was up to $14.2 billion through 2031 (claim 4282). The March 2026 order form is described as an expansion of about $21 billion (claims 10729 and 2123), but one secondary source calls $21 billion the total commitment (claim 17654), so it is unclear whether Meta's exposure is about $21 billion or about $35 billion. Jane Street's roughly $6 billion (claim 10304) is named and quantified, though claims 10695 and 10730 restate it and should not be counted again. Anthropic's multi-year agreement (claim 10731) has no stated value. Everything else is softer. The NVIDIA collaboration framework (claims 3892 and 17596), the reference-architecture work it would only 'work towards' (claim 17598), the IBM partnership (claim 4971), continued collaboration with Core Scientific (claim 20313), the CrowdStrike 'partnership' (claims 11768 and 10330), the Genesis Mission (claim 11778), the Aston Martin sponsorship (claim 25576) and the Exemplar Cloud designations (claims 10332 and 9324) carry no disclosed commitments. The logo lists (claims 25575, 11767 and 9320) are names without dollars. Several counterparties are never identified: the 'new hyperscaler customer' (claim 25574) and the 'leading AI enterprise' financed by DDTL 4.0 (claim 11224). The DDTL 3.0 'customer contract' (claim 4595) was later disclosed as OpenAI (claim 26687). Microsoft, historically the dominant customer, does not appear by name anywhere in this record. Concentration has broadened but remains high. Year-end 2025 RPO was $60.7 billion, up from $15.1 billion (claims 3555 and 3572). Named commitments by then from OpenAI ($22.4 billion), Meta ($14.2 billion) and NVIDIA ($6.3 billion, claim 20209) total about $42.9 billion, roughly 71% of RPO. That leaves about $18 billion attributable to unnamed or legacy contracts. NVIDIA appears as both supplier and customer, and that circularity deserves scrutiny. The record supplied was truncated at claim 9317, so later disclosures are not reflected here.

CoreWeave's record is unusually rich in executed paper. The OpenAI relationship is the clearest case: an $11.9 billion agreement (claim 19820), a $4 billion expansion (25573), and an up to roughly $6.5 billion order form under the existing MSA (4296, 20198). These sum exactly to the stated $22.4 billion total. One wrinkle remains. The first-quarter release credited only $11.2 billion of that initial deal to backlog (4969), so roughly $0.7 billion of the headline never reached reported backlog. Meta is similarly documented. An up to $14.2 billion order form runs through December 2031 (4282), followed by a March 31, 2026 order form of about $21 billion through 2032 (2123, 10340). Both are framed as expansions, which implies about $35 billion cumulative. Some press phrasing (17654, 17655) blurs whether the $21 billion is incremental. NVIDIA's $6.3 billion capacity order (4352, 20209) is likewise a signed order form, but the counterparty is also CoreWeave's GPU supplier, which makes that revenue circular. By contrast, the January 2026 NVIDIA arrangement is explicitly a "collaboration framework" (3892, 17596, 17598) and should not be read as revenue. Jane Street's roughly $6 billion commitment (10304, 10730) and the Anthropic multi-year agreement (10731, no value) are named, but no order form appears in this record. Several items are softer. IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), Aston Martin (25576), the Genesis Mission (11778), and Exemplar Cloud status (10332) are partnerships or designations. The logo lists (25575, 11767) carry no contract values. Some counterparties remain unidentified: the "new hyperscaler customer" (25574) and the "leading AI enterprise" behind DDTL 4.0 (11224). The DDTL 3.0 "customer contract" (4595) was later attributed to OpenAI (26687). Concentration has evolved from near-total OpenAI dependence in early 2025 to an OpenAI–Meta duopoly. Together the two account for about $57.6 billion of committed value. The backlog arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), an increase of $45.6 billion. Named 2025 deals total about $42.9 billion: OpenAI incremental, Meta $14.2 billion, and NVIDIA $6.3 billion. The unnamed hyperscaler, smaller logos, and revenue recognized during the year (not disclosed here) must fill the remaining gap. That gap is not trivial given the "up to" ceilings, which RPO may not fully count. Earlier backlog figures of $25.9 billion and $30.1 billion (4970, 25586) included non-RPO amounts contingent on delivery, so they are not like-for-like with year-end RPO.

CoreWeave's record separates into executed order forms, priced commitments without a filed instrument, and partnership language that carries no disclosed economics. The firmest evidence is the 8-K order forms under existing master services agreements. These cover OpenAI (claim 4296), NVIDIA (claim 4352), and Meta in two tranches: the initial commitment of up to $14.2 billion through December 2031 (claim 4282) and the March 31, 2026 order form of roughly $21 billion through 2032 (claims 2123, 10340). The OpenAI build-up does reconcile: $11.9 billion (claim 19820), plus $4 billion (claim 25573), plus $6.5 billion equals the stated $22.4 billion (claim 20198). Two caveats apply. The Meta and later OpenAI figures are framed as 'up to' amounts. And the January 2026 NVIDIA arrangement is explicitly a 'collaboration framework' (claim 3892) rather than a purchase commitment, so it should not be treated as additive to the $6.3 billion order (claim 20209).

Jane Street's roughly $6 billion (claims 10304, 10730) is a named, priced commitment, but no order-form filing appears in this record. Anthropic (claim 10731) is named without any value. Several material contracts remain anonymous: the 'new hyperscaler customer' (claim 25574), the customer whose contract DDTL 3.0 finances (claim 4595), and the 'leading AI enterprise' behind DDTL 4.0 (claim 11224). The logo rosters (claims 25575, 11767, 10338) also carry no terms. The same is true of the IBM, CrowdStrike, Core Scientific, Aston Martin and Genesis Mission items (claims 4971, 11768, 20313, 25576, 11778), which read as partnership or affiliation language rather than revenue contracts.

The revenue arithmetic needs care. At March 31, 2025, backlog was $25.9 billion, of which only $14.7 billion was RPO (claim 4970). The $11.2 billion difference exactly matches the OpenAI backlog addition (claim 4969). That implies the OpenAI deal initially sat entirely in the conditional, delivery-dependent bucket, and it differs from the $11.9 billion headline. RPO then rose from $15.1 billion to $60.7 billion during 2025 (claim 3555). The named OpenAI, Meta and NVIDIA commitments sum to about $42.9 billion, roughly 70 percent of year-end RPO before any recognition. Concentration therefore moved from a single dominant lab toward two or three counterparties, with Meta's 2026 expansion likely pushing it higher. Several extracted values (claims 10729, 9317, 10730) carry unit errors and should be read in billions. The record is also truncated at claim 9320.

CoreWeave's record separates cleanly into executed order forms and softer partnership language. The firmest disclosures are 8-K order forms under existing master services agreements: OpenAI (4296), Meta's $14.2 billion commitment through December 2031 (4282) and its roughly $21 billion March 2026 expansion through 2032 (2123, 10340), and NVIDIA's capacity purchase (4352). Jane Street's approximately $6 billion commitment (10304) is also described as a commitment. Anthropic's multi-year agreement (10731) is named but carries no disclosed value. By contrast, the January 2026 NVIDIA 'collaboration framework' (3892, 17596, 17598) is ecosystem and buildout language, not a revenue contract. The same applies to IBM (4971), Aston Martin (25576), CrowdStrike (11768, 10330), Core Scientific (20313), the Genesis Mission (11778) and Exemplar Cloud status (10332). The logo rosters (25575, 11767, 9320) name counterparties but attach no economics. The MUFG, Morgan Stanley, Goldman and JPMorgan references (11225) describe lenders, not customers.

Several material counterparties are never identified. These include the 'customer contract' that DDTL 3.0 was raised to serve (4595), the 'new hyperscaler customer' signed in Q2 2025 (25574), and the 'leading AI enterprise' behind DDTL 4.0 (11224). Readers cannot tell whether these map to Meta or OpenAI or to someone else.

Concentration has broadened in name more than in dollars. OpenAI anchored 2025 at a cumulative $22.4 billion: $11.9 billion (19820), plus $4 billion (25573), plus $6.5 billion (20198). That sum reconciles. Meta's $14.2 billion and $21 billion orders then made it at least as large an exposure. Jane Street and Anthropic are the first sizeable non-hyperscaler, non-lab diversifiers. Those two names plus OpenAI and Meta still plausibly account for most contracted value.

The arithmetic only partly reconciles. The $11.2 billion of OpenAI backlog cited in 4969 sits about $0.7 billion below the $11.9 billion contract headline (19820), and the gap is unexplained. RPO rose from $15.1 billion (3572) to $60.7 billion (3555) during 2025. Named 2025 deals of $10.5 billion from OpenAI after Q1 plus $14.2 billion from Meta leave roughly $20 billion or more of net additions attributed to no named customer. That remainder is presumably the unnamed hyperscaler and the DDTL-financed contracts. Finally, 17654 and 17655 leave it ambiguous whether Meta's $21 billion supersedes or adds to the $14.2 billion, and 10729 records the figure as '$21,000', a unit error to correct. The claim list supplied ends mid-entry at 9318, so the latest May 2026 disclosure could not be assessed.

CoreWeave's record splits cleanly into binding order forms and softer partnership language. The binding layer is narrow and large. OpenAI's commitments rest on a $11.9 billion agreement (claim 19820), a $4 billion expansion (25573), a September 2025 order form under the existing MSA (4296), and a further deal of up to $6.5 billion, which the company says brings OpenAI to up to $22.4 billion (20198). Meta signed a $14.2 billion order form through December 2031 with an expansion option (4282, 4283), then a March 31, 2026 order form of about $21 billion through 2032 (2123, 10340). NVIDIA's $6.3 billion capacity purchase is also an order form (4352, 20209). Jane Street's roughly $6 billion commitment (10304) is named and sized. The Anthropic agreement (10731) is named only through secondary coverage, with no value disclosed.

Everything else is softer. The IBM partnership (4971), the Aston Martin sponsorship (25576), continued collaboration with Core Scientific (20313), the CrowdStrike partnership (11768), Genesis Mission participation (11778), and Exemplar Cloud status (10332) carry no disclosed dollars. The January 2026 NVIDIA expansion is explicitly a "collaboration framework" (3892) whose software component is to "work towards" inclusion in reference architectures (17598). The logo lists in 25575, 11767 and 10338 name counterparties but attach no economics to them. Three material counterparties remain unidentified: the "new hyperscaler customer" (25574), the customer behind DDTL 3.0 (4595), and the "leading AI enterprise" behind DDTL 4.0 (11224).

Concentration has shifted from one counterparty to two, not to many. In March 2025, $25.9 billion of backlog equalled $14.7 billion of RPO plus $11.2 billion of OpenAI-linked other amounts (4970, 4969), so OpenAI was about 43% of backlog. That $11.2 billion also does not match the $11.9 billion headline. Meta's 2026 order alone is now about the same size as OpenAI's cumulative $22.4 billion. Whether Meta's $21 billion adds to the $14.2 billion or supersedes it is unclear across 10340, 17654 and 17655, and that answer decides which customer is larger.

The arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), an increase of $45.6 billion. Named 2025 deals sum to about $42.9 billion, but most are "up to" ceilings, and 2025 revenue burn-off is not in the record. The residual is therefore probably larger than $2.7 billion and attributable to unnamed or unsized customers. The supplied record ends mid-entry at claim 9324, so later disclosures were not assessed.

CoreWeave's signed, counterparty-named commitments sit with four companies: OpenAI, Meta, NVIDIA and Jane Street. OpenAI is documented as order forms under an existing MSA (claims 19820, 4296), with tranches of $11.9 billion, $4 billion and up to $6.5 billion summing to the stated "up to approximately $22.4 billion" (25573, 20198). There is one wrinkle: the same first deal is booked as only $11.2 billion of backlog (4969). Meta's order form commits "up to approximately $14.2 billion" through 2031 (4282), and a March 31, 2026 order form adds about $21 billion through 2032 (2123, 10340). The NVIDIA capacity order form (4352) corresponds to the $6.3 billion "strategic collaboration" (20209). Jane Street's roughly $6 billion (10304) is quantified, but the record supplies no filing language for it. Anthropic is named with no value or terms, sourced only from secondary coverage (10731).

Everything else is softer. The IBM partnership (4971), the Aston Martin sponsorship (25576), the Core Scientific collaboration (20313), Genesis Mission participation (11778) and the January 2026 NVIDIA "collaboration framework" (3892, 17596) are not revenue contracts, and the last says only that CoreWeave will "work towards" inclusion (17598). CrowdStrike appears both as a "partnership" (11768) and in a customer logo list (11767). Those logo lists, including the later Cohere, Mistral and Jane Street "expansions," carry no dollar figures (25575, 9319). Three material counterparties are never named: the "new hyperscaler customer" (25574), the customer behind DDTL 3.0 (4595), and the "leading AI enterprise" behind DDTL 4.0 (11224).

Concentration has narrowed rather than diversified. In early 2025 OpenAI alone accounted for about 43% of a $25.9 billion backlog (4970). By December 2025, named OpenAI, Meta and NVIDIA ceilings of roughly $42.9 billion equal about 71% of $60.7 billion RPO (3555), before Meta's $21 billion expansion.

The revenue arithmetic only loosely reconciles. Adding the second-half 2025 deals ($14.2B, $6.5B, $6.3B) to the $30.1 billion June backlog (25586) gives about $57 billion before revenue burn. That is close to $60.7 billion, but the comparison mixes "backlog" and "RPO," and "up to" ceilings may not convert one-for-one. The March 2025 RPO of $14.7 billion was below the $15.1 billion of December 2024 (3572) despite the OpenAI signing, which suggests deals enter RPO with a lag. Two items reference events later than their own record dates (10729, 10730), and the claim record is cut off at 9324.

CoreWeave's record separates into two tiers. The first is executed paper, meaning order forms filed under existing master services agreements. Three counterparties sit in this tier. OpenAI's order is dated September 23, 2025 (claim 4296). NVIDIA's capacity purchase order is dated September 2025 (claim 4352). Meta has two orders: the initial one committing up to approximately $14.2 billion through December 2031 (claim 4282), and a March 31, 2026 order (claim 2123) that expanded the relationship by about $21 billion through 2032 (claim 10340). Jane Street's roughly $6 billion commitment (claim 10304) is also stated as a commitment. Everything else is softer. The IBM tie-up (claim 4971), Aston Martin (claim 25576), CrowdStrike (claim 11768), Genesis Mission (claim 11778), the continuing Core Scientific relationship (claim 20313) and the NVIDIA collaboration framework (claim 3892) are partnership or framework language. None carries a dollar value. The Anthropic multi-year agreement (claim 10731) is named but unpriced and comes from secondary reporting. The rotating logo lists (claims 25575, 11767, 9320) name customers without disclosing any economics.

Three material counterparties remain unidentified. One is the 'new hyperscaler customer' (claim 25574). The other two are the customers whose contracts justified the DDTL 3.0 and DDTL 4.0 facilities (claims 4595, 11224), the latter described only as a 'leading AI enterprise.' These are financing-grade contracts with unnamed obligors.

Concentration has broadened in name but not in dollars. OpenAI commitments reached up to $22.4 billion (claim 20198), which reconciles to $11.9 billion plus $4 billion plus $6.5 billion (claims 19820, 25573). One gap needs explaining: the original $11.9 billion deal contributed only $11.2 billion to backlog (claim 4969). Meta's $14.2 billion plus $21 billion totals about $35 billion if additive. Claim 17654 can be read either way. Together these two customers dwarf every other disclosed relationship.

The arithmetic broadly holds. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), an increase of $45.6 billion. Named 2025 additions are OpenAI ($21.7 billion booked), Meta ($14.2 billion) and NVIDIA ($6.3 billion, claim 20209). These total roughly $42 billion, which leaves little room for the hyperscaler and the logo customers once in-year revenue recognition is netted out. That suggests either small contributions from those customers or 'up to' figures booked at less than face value. Two caveats apply. NVIDIA appears as both supplier and customer. Backlog, which was $25.9 billion and then $30.1 billion (claims 4970, 25586), is not the same measure as RPO. The record supplied was truncated at claim 9318, so this read may be incomplete.

CoreWeave's signed, priced customer commitments come from a small set of counterparties. The firmest disclosures are order forms executed under existing Master Services Agreements: OpenAI (claims 19820, 4296), Meta (4282, 2123) and NVIDIA (4352). OpenAI's commitments are stated at up to approximately $22.4 billion (20198), and Meta's at $14.2 billion through 2031 (4282) plus a roughly $21 billion order signed March 31, 2026 (2123, 10340). Jane Street's approximately $6 billion (10304) is also described as a commitment. Anthropic's multi-year agreement (10731) appears only in secondary reporting and carries no figure.

Much of the rest is softer language. The NVIDIA relationship moved from a priced order (4352, 20209) to a 'collaboration framework' (3892) whose software terms only 'work towards' inclusion in reference architectures (17598). IBM (4971), CrowdStrike (11768, 10330), the Aston Martin sponsorship (25576), Genesis Mission participation (11778), and the post-merger Core Scientific language (20313) carry no disclosed economics. The logo lists of 'partner of choice' customers (25575, 11767, 9320) name counterparties but give no contract values.

Three material counterparties remain unidentified. These are the 'new hyperscaler customer' (25574), the customer contract financed by DDTL 3.0 (4595), and the 'leading AI enterprise' behind DDTL 4.0 (11224). They may overlap with named parties, but the record does not say so.

Concentration has deepened rather than eased. At March 31, 2025, OpenAI's $11.2 billion accounted for about 43% of a $25.9 billion backlog (4969, 4970). The $11.2 billion figure also sits $0.7 billion below the $11.9 billion headline (19820). By late 2025, OpenAI, Meta and NVIDIA together represented up to about $42.9 billion against $60.7 billion of RPO (3555), roughly 71%. Including NVIDIA here is notable because it is also CoreWeave's key supplier. The 2026 additions of Jane Street and Anthropic broaden the roster, but Meta's roughly $35 billion combined exposure, if the $21 billion is incremental, makes a single customer even more dominant.

The revenue arithmetic broadly reconciles, but with little room to spare. RPO rose $45.6 billion during 2025 (3572, 3555), and named 2025 deals total up to $42.9 billion. Revenue recognized during the year would have consumed some backlog, so the named deals plausibly explain nearly all gross additions. That leaves little visible contribution from the logo-list customers. The 'up to' qualifiers and the separate backlog definition, which adds non-RPO amounts subject to delivery conditions (4970), weaken any tight tie-out.

It is also unclear whether Meta's $21 billion is additive to the $14.2 billion. Claim 10729 calls it an expansion, while 9317 calls it a new commitment. Claim 10729 also carries a March 1 date that precedes the March 31 signing it describes. The claim record supplied was truncated at claim 9319.

CoreWeave's announced pipeline rests on a short list of signed order forms under existing master services agreements. Its partnership, framework and logo language deserves a heavier discount. The firmest disclosures are the 8-K order forms. These include OpenAI's September 2025 order form (claim 4296), Meta's initial order form of up to about $14.2 billion through December 2031 (claim 4282), Meta's March 31, 2026 order form (claim 2123) and the NVIDIA capacity order form (claim 4352). OpenAI's commitments run from $11.9 billion (claim 19820) through a $4 billion expansion (claim 25573) to an up-to-$6.5 billion deal, which brings the total to up to about $22.4 billion (claim 20198). The words 'up to' make these ceilings rather than floors. Meta's roughly $21 billion expansion is cited in claims 10729, 10694, 10340, 17654 and 9317, but it is one contract and should be counted once. Jane Street's approximately $6 billion (claims 10304, 10695) is a named commitment without an order-form citation in the record. Anthropic's multi-year agreement carries no value (claim 10731). The softer category includes IBM (claim 4971), CrowdStrike (claims 11768, 10330), the Aston Martin F1 sponsorship (claim 25576), the Genesis Mission (claims 11778, 10331) and NVIDIA's January 2026 collaboration framework (claims 3892, 17596, 17598). Logo rosters in claims 25575, 11767 and 9320 name counterparties but give no economics. Three material counterparties remain unidentified. The first is the customer financed by DDTL 3.0 (claim 4595). The second is the 'new hyperscaler customer' (claim 25574). The third is the 'leading AI enterprise' behind DDTL 4.0 (claim 11224). Concentration has deepened rather than diversified. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 31, 2025 (claims 4969, 4970). By year-end, OpenAI's $22.4 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion (claim 20209) totaled $42.9 billion, about 71% of the $60.7 billion in RPO (claim 3555). The arithmetic broadly reconciles. RPO rose $45.6 billion from $15.1 billion (claim 3572), close to the $42.9 billion of named 2025 deals, and the record does not show the 2025 revenue recognized against RPO. Two items do not tie. OpenAI's original deal is $11.9 billion in claim 19820 but $11.2 billion in the backlog in claim 4969. Backlog also mixes RPO with amounts that depend on delivery conditions (claim 4970), so it overstates firm obligations. Claim 9319 is truncated in the record and is not relied on here.

CoreWeave's record separates cleanly into contracts evidenced by filed order forms and a larger body of partnership language that carries no disclosed economics. The signed core is narrow and well named. OpenAI order forms under an existing MSA (claims 19820, 4296), Meta's initial order form of up to $14.2 billion through December 2031 (claim 4282), and its March 2026 order form (claim 2123) are documented contracts, as is NVIDIA's capacity order form (claim 4352). Jane Street's roughly $6 billion commitment (claim 10304) and the multi-year Anthropic agreement (claim 9318) are named and described as signed, but Anthropic carries no dollar figure, and the record cites no filed order form for either. Weaker language sits around this core. NVIDIA's 'collaboration framework' (claim 3892) and $6.3 billion 'strategic collaboration' (claim 20209) are framed as partnership rather than purchase. The IBM, CrowdStrike, Core Scientific, Genesis Mission and Aston Martin items (claims 4971, 11768, 20313, 11778, 25576) are relationships without disclosed consideration. The logo lists (claims 25575, 11767) name customers but attach no values. Three material counterparties remain unidentified: the 'new hyperscaler customer' (claim 25574), the customer contract behind DDTL 3.0 (claim 4595), and the 'leading AI enterprise' backing DDTL 4.0 (claim 11224). Concentration has deepened even as the logo count grew. OpenAI's stated commitments total up to $22.4 billion (claim 20198), and Meta's initial $14.2 billion brings the two to about $36.6 billion against $60.7 billion of RPO at year-end 2025 (claim 3555), roughly 60 percent. Adding NVIDIA's $6.3 billion takes that above 70 percent. Meta's $21 billion expansion (claim 10340) makes it the largest single exposure in 2026, so the Jane Street and Anthropic wins dilute concentration only modestly. The arithmetic broadly reconciles, with caveats. The OpenAI sequence sums correctly: $11.9 billion, plus $4 billion, plus $6.5 billion equals $22.4 billion (claims 19820, 25573, 20198). However, the initial deal was booked as $11.2 billion of backlog (claim 4969), a $0.7 billion gap that is never explained. RPO rose about $45.6 billion during 2025 (claims 3572, 3555), against roughly $42.9 billion of named OpenAI, Meta and NVIDIA deals. That leaves a few billion, before burn-down of recognized revenue, for the unnamed hyperscaler and the logo customers. Most figures are 'up to' ceilings. Secondary coverage also blurs whether Meta's $21 billion is incremental or a total commitment (claims 17654, 10729), which affects 2026 backlog math.

CoreWeave's record mixes executed order forms with softer language. The hardest evidence is a set of order forms filed under master services agreements. These cover OpenAI, where the original $11.9 billion agreement (claim 19820) was followed by a September 2025 order form (4296). They cover Meta, with an order form for up to $14.2 billion through December 2031 (4282) and a second one dated March 31, 2026 for about $21 billion through 2032 (2123, 10340). They also cover NVIDIA, through a September 2025 capacity order form (4352) later described as a $6.3 billion collaboration (20209). Jane Street's roughly $6 billion commitment (10304) is named and sized. Anthropic's multi-year agreement (10731) is named but has no stated value. The IBM partnership (4971), the CrowdStrike 'global partnership' (11768, 10330), the NVIDIA 'collaboration framework' and reference-architecture work (3892, 17598), the Genesis Mission (11778), Aston Martin F1 (25576), and the Core Scientific language (20313) are framework, sponsorship, or relationship disclosures, not quantified contracts. The logo lists (25575, 11767, 9320, 9319) name counterparties but attach no dollars. Several material counterparties remain unidentified: a 'new hyperscaler customer' (25574), the 'customer contract' financed by DDTL 3.0 (4595), and the 'leading AI enterprise' behind DDTL 4.0 (11224).

Concentration has shifted from one dominant customer to two. OpenAI's commitments grew from $11.9 billion to $15.9 billion (25573) and then to 'up to' $22.4 billion (20198). That total reconciles arithmetically, although the first backlog addition was booked at $11.2 billion rather than $11.9 billion (4969). Meta then rose to rival OpenAI. If the $21 billion order form adds to the $14.2 billion form rather than replacing it, Meta's combined exposure is about $35 billion, which would exceed OpenAI's. Secondary press items (17654, 10729) blur this distinction.

The revenue arithmetic broadly holds but cannot be fully closed. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3571), a net increase of about $45.6 billion. Named 2025 signings cover most of it: roughly $22.4 billion from OpenAI, $14.2 billion from Meta, and $6.3 billion from NVIDIA, about $43 billion in total. The residual is plausibly the unnamed hyperscaler and the logo customers. However, revenue recognized in 2025 is not on record, so gross additions cannot be derived. The 'up to' framing may also exceed what RPO actually counts. Backlog, the earlier metric (4970, 25586), also differs from RPO by definition. The 2026 Meta and Jane Street deals fall after the December 31, 2025 RPO date and should appear only in later disclosures.

CoreWeave's signed commitments are concentrated in a few order forms executed under master services agreements. The rest of its counterparty disclosure is partnership, framework or logo-list language. The hard paper starts with OpenAI: the March 2025 agreement (19820) and the September 2025 order form (4296), which brought OpenAI commitments to "up to approximately $22.4 billion" (20198). Meta follows with the $14.2 billion order form running through December 2031 (4282) and a second order form signed March 31, 2026 (2123) for about $21 billion through 2032 (10340, 9317). NVIDIA's September 2025 order form (4352) is valued at $6.3 billion (20209), and NVIDIA is also CoreWeave's key supplier. Jane Street's roughly $6 billion commitment (10304) is the only other counterparty with a dollar figure. Anthropic is described as a signed multi-year agreement (9318), but no value is disclosed. IBM (4971), CrowdStrike (11768, 10330), the Genesis Mission (11778) and Core Scientific (20313) use partnership or collaboration language, not contract language. The January 2026 NVIDIA expansion is explicitly a "collaboration framework" (3892, 17596). The "partner of choice" rosters (11767, 9320, 10338) are customer logos without terms. Some financed contracts are not named at all. DDTL 3.0 funds "a customer contract" (4595), and DDTL 4.0 serves a "leading AI enterprise" (11224). Claim 9119 also cites two large customer contracts, but its text is truncated in the record. Concentration has moved from one customer to two. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 31, 2025 (4969, 4970). By year-end 2025, the OpenAI, Meta and NVIDIA headline amounts total about $42.9 billion against $60.7 billion of RPO (3555), roughly 71% if fully counted. Meta's 2026 expansion could lift it toward $35 billion and make it the largest single exposure. The revenue arithmetic broadly reconciles, with gaps. RPO rose $45.6 billion during 2025 (3555), close to the named additions plus revenue recognized during the year. However, the headline figures are "up to" ceilings, not floors. The OpenAI totals also do not add up as disclosed. The deal was announced at $11.9 billion (19820) but booked as $11.2 billion of backlog (4969), and $11.9 billion plus $6.5 billion falls about $4 billion short of the $22.4 billion cumulative figure (20198). That implies a further OpenAI tranche that is not in this record. Backlog and RPO are also different measures, since the March 2025 backlog included amounts outside RPO that depend on delivery conditions (4970). Some extracted values are mis-scaled and should not be relied on without the source text, such as $21,000 for Meta (10729) and $6,000 for Jane Street (10730).

CoreWeave's record splits cleanly into contracts with executed paperwork and a larger layer of partnership language. The signed core is a small set of order forms under existing master services agreements. These are OpenAI in September 2025 (claim 4296), Meta in September 2025 for up to $14.2 billion (4282) and again on March 31, 2026 for about $21 billion (2123, 10340), and NVIDIA's capacity purchase (4352), later sized at $6.3 billion (20209). The March 2025 OpenAI agreement (19820) and the Jane Street commitment of about $6 billion (10304) are named and dollarized. The Anthropic deal is described as signed and multi-year but carries no value (9318). Most of the remaining claims are softer. The NVIDIA January 2026 'collaboration framework' (3892, 17596) and its 'work towards' reference-architecture language (17598) are frameworks, not bookings. The same applies to the IBM (4971), CrowdStrike (11768, 10330) and Core Scientific (20313) partnerships, the Genesis Mission (11778), the Exemplar Cloud designations (10332, 9324) and the 'partner of choice' logo lists (11767, 9320). Expanded relationships with Cohere and Mistral (9319) carry no terms. Several financings point to counterparties that are not named: 'a customer contract' behind DDTL 3.0 (4595) and a 'leading AI enterprise' behind DDTL 4.0 (11224). Claim 9119, on two large customer contracts, is truncated in the record and cannot be assessed. Concentration has moved from a single dominant counterparty to two. OpenAI's stated total reached up to $22.4 billion (20198), while Meta's two order forms together total roughly $35 billion. Repeated citations of the Meta and Jane Street figures (10729, 10694, 17654, 9317; 10730, 10695) describe one deal each and should not be counted twice. The revenue arithmetic only partly reconciles. OpenAI's initial deal appears as $11.9 billion (19820) but added $11.2 billion to backlog (4969). Adding the $6.5 billion expansion gives $18.4 billion, not $22.4 billion, so roughly $4 billion of OpenAI commitments is missing from this record. RPO fell from $15.1 billion to $14.7 billion in the first quarter of 2025 (3572, 4970), even though the OpenAI deal was signed in that quarter. This suggests much of that deal initially sat in conditional backlog rather than RPO. RPO of $60.7 billion at year-end 2025 (3555) compares with about $42.9 billion of named 2025 OpenAI, Meta and NVIDIA commitments. Read that way, three counterparties account for roughly 70 percent of RPO before revenue burn-down, and the remainder is undisclosed. Many of the headline figures are 'up to' amounts, which weakens a direct reconciliation.

CoreWeave's record contains a small core of documented, signed customer contracts and a much larger layer of partnership and logo language. Most of the disclosed dollars sit in the first category. The executed agreements take the form of order forms under existing master services agreements. They cover OpenAI (claim 4296, following the original $11.9 billion agreement in claim 19820), Meta at up to $14.2 billion through December 2031 (claim 4282) and a further roughly $21 billion through December 2032 (claim 2123, claim 10340, claim 9317), and NVIDIA (claim 4352), whose $6.3 billion commitment (claim 20209) makes it both supplier and customer. Jane Street's roughly $6 billion (claim 10304) is described as a commitment, and Anthropic's deal is called signed (claim 9318), but neither has a cited order form, and Anthropic's value is undisclosed. Several items fall outside signed-contract language. These are the IBM partnership (claim 4971), the NVIDIA collaboration framework (claim 3892, claim 17596), whose software-integration goal is phrased as working towards (claim 17598), the CrowdStrike global partnership (claim 11768), the Core Scientific collaboration (claim 20313), Genesis Mission participation (claim 11778), Exemplar Cloud status (claim 9324), and the partner-of-choice logo lists (claim 11767, claim 9320). The expansions with Cohere and Mistral are also unquantified (claim 9319). Two financing disclosures cite contracts without naming the counterparty: a customer contract behind DDTL 3.0 (claim 4595) and a leading AI enterprise behind DDTL 4.0 (claim 11224). Concentration has moved from OpenAI to Meta. In May 2025, OpenAI's $11.2 billion was about 43 percent of $25.9 billion in backlog (claim 4969, claim 4970). By April 2026, Meta's cumulative headline value of about $35.2 billion exceeds OpenAI's up to $22.4 billion (claim 20198). The revenue arithmetic only partly reconciles. OpenAI's original deal is $11.9 billion in claim 19820 but $11.2 billion in claim 4969. Adding $6.5 billion gives about $18.4 billion, so the $22.4 billion total implies a roughly $4 billion tranche that this record does not show. RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3572, claim 3555), an increase of about $45.6 billion. Named 2025 deals total about $42.9 billion, but those are up-to figures, and 2025 recognized revenue is not in the record. The match is therefore plausible but unverified. The roughly $27 billion of 2026 Meta and Jane Street commitments postdates that RPO figure. The record is truncated at claim 9121.

CoreWeave's signed-contract record is unusually well documented for its largest counterparties, but the remainder of its customer narrative is mostly partnership language. The binding instruments are order forms under existing Master Services Agreements. These include OpenAI (claims 19820, 4296, and 20198, which bring commitments to up to about $22.4 billion), Meta (4282 at up to $14.2 billion through 2031, and 2123 on March 31, 2026, sized at about $21 billion in 10340), and NVIDIA (4352, sized at $6.3 billion in 20209). Jane Street's roughly $6 billion commitment (10304) also reads as contractual, and Anthropic's multi-year agreement is described as signed (9318) but carries no disclosed value. Several claims restate the same Meta $21 billion deal (10729, 10694, 17654, 9317) and should not be counted more than once. Other items are frameworks or affiliations rather than revenue contracts: the IBM partnership (4971), the NVIDIA 'collaboration framework' and reference-architecture work (3892, 17596, 17598), the CrowdStrike 'global partnership' (11768), Genesis Mission participation (11778), and the Core Scientific collaboration (20313). The 'partner of choice' lists (11767, 9320) name logos such as Cursor, Midjourney, Perplexity and Hudson River Trading, but they attach no terms. Some financing disclosures go the other way, citing contracts without naming the counterparty. DDTL 3.0 funds 'a customer contract' (4595), DDTL 4.0 serves a 'leading AI enterprise' (11224), and a May 2026 item references 'two large customer contracts' (9119). Concentration has shifted from a single dominant buyer to a duopoly. In early 2025, OpenAI's roughly $11.9 billion was close to half of the $25.9 billion backlog (4970). By year-end, OpenAI's $22.4 billion and Meta's $14.2 billion together came to about 60% of the $60.7 billion RPO (3555), or about 71% once NVIDIA is included. Meta's March 2026 expansion likely makes it the largest exposure. On arithmetic, RPO rose $45.6 billion during 2025, from $15.1 billion (3572) to $60.7 billion. The named OpenAI, Meta and NVIDIA deals total about $42.9 billion, which broadly reconciles. That fit holds only if the 'up to' ceilings were booked at full value, and only if revenue recognized during the year was offset by smaller unnamed contracts. Two gaps remain unexplained. OpenAI's initial deal appears as $11.9 billion in one disclosure (19820) and $11.2 billion of added backlog in another (4969), a $0.7 billion difference. The record also does not itemize roughly $4 billion of the $22.4 billion OpenAI total.

CoreWeave's record splits cleanly into executed order forms and softer relationship language. The signed tier rests on order forms under master services agreements with OpenAI (19820, 4296, 20198), Meta (4282, 2123) and NVIDIA (4352). It also includes a roughly $6 billion Jane Street commitment (10304) and a multi-year Anthropic agreement (9318) whose value was never disclosed. The rest reads differently. IBM is a "partnership" (4971) and CrowdStrike a "global partnership" (11768). NVIDIA's January 2026 arrangement is a "collaboration framework" that commits the parties only to "work towards" software inclusion (3892, 17598). Core Scientific pledges to "collaborate on shared opportunities" (20313), and Cohere and Mistral appear only as "expanded relationships" (9319). The "partner of choice" rosters (11767, 9320), Exemplar Cloud status (10332) and Genesis Mission (11778) name organizations without attaching any amount or term.

Several material obligations carry no named counterparty. DDTL 3.0 finances "a customer contract" (4595), and DDTL 4.0 funds services for a "leading AI enterprise" (11224). A May 2026 item references two large customer contracts (9119), but that claim is truncated in the record supplied. None of these can be tied to a named deal from the disclosures alone.

Concentration has shifted without diversifying much. In March 2025, OpenAI's $11.9 billion contract (19820) sat against a $25.9 billion backlog (4970). By year-end 2025, three named commitments totaled about $42.9 billion: OpenAI up to $22.4 billion (20198), Meta up to $14.2 billion (20197) and NVIDIA at $6.3 billion (20209). That is roughly 71% of the $60.7 billion RPO (3555). The March 31, 2026 Meta order form for about $21 billion (2123, 10340) lifts Meta to roughly $35 billion nominal, overtaking OpenAI as the largest single exposure.

The arithmetic only partly reconciles. OpenAI's initial deal is $11.9 billion in the announcement (19820) but $11.2 billion of added backlog in the quarterly release (4969), and the $0.7 billion gap is unexplained. The $22.4 billion cumulative OpenAI figure (20198) also exceeds the $11.9 billion and $6.5 billion tranches by about $4 billion, which implies an intermediate expansion absent from this record. The Meta and OpenAI totals are framed as "up to," making them ceilings rather than floors. The Q1 2025 backlog also counted $11.2 billion beyond RPO only "subject to" delivery conditions (4970). Finally, about $17.8 billion of year-end RPO is not attributable to any named counterparty.

CoreWeave's record contains three counterparties backed by executed paper: OpenAI, Meta and NVIDIA. Each is documented as an order form under an existing Master Services Agreement. OpenAI's initial agreement (claim 19820, $11.9 billion) was followed by a September 2025 order form (4296) and an expansion to a cumulative "up to approximately $22.4 billion" (20198). Meta's $14.2 billion order form runs through December 2031 (4282). A second Meta order form dated March 31, 2026 (2123) carries roughly $21 billion (10340, 9317). NVIDIA's September 2025 order form (4352) corresponds to the $6.3 billion figure (20209). Jane Street's roughly $6 billion commitment (10304) is named and sized, but the record cites only press language, not a filing.

Everything else needs a discount. The January 2026 NVIDIA expansion is explicitly a "collaboration framework" (3892, 17596), and reference-architecture validation (17598) is not revenue. IBM (4971), CrowdStrike (11768), Core Scientific (20313) and the DOE Genesis Mission (11778) are described as partnerships, not contracts. The logo lists (11767, 9320) and the Cohere/Mistral "expanded relationships" (9319) carry no values. Anthropic is a signed multi-year agreement (9318), but it is unsized. Several financings point to unnamed counterparties: "a customer contract" behind DDTL 3.0 (4595), a "leading AI enterprise" behind DDTL 4.0 (11224), and "two large customer contracts" (9119, which is truncated in the record).

Concentration has broadened at the margin but remains heavy. In early 2025 OpenAI alone was nearly half of the $25.9 billion backlog (4970). By year-end 2025, the named OpenAI, Meta and NVIDIA amounts total about $42.9 billion, roughly 70% of the $60.7 billion RPO (3555). If the March 2026 Meta deal is additive, as "expanded… by about $21 billion" (10729) implies, Meta becomes the largest counterparty at about $35 billion.

The arithmetic roughly reconciles, with caveats. RPO rose $45.6 billion during 2025 (3572, 3555), close to the $42.9 billion of named 2025 deals. Recognized revenue, which is not on record, would widen that gap. The OpenAI figures do not tie internally: $11.9 billion headline versus $11.2 billion of backlog (4969). In addition, $11.9 billion plus $6.5 billion leaves about $4 billion of the $22.4 billion total unexplained in this record. The pervasive "up to" phrasing, and backlog that depends on "delivery and availability" (4970), mean headline values are ceilings rather than floors. No 2026 RPO figure is on record to test the Meta, Jane Street and Anthropic additions.

CoreWeave's record separates cleanly into executed order forms and softer partnership language. The signed, dollar-denominated agreements all name their counterparty. OpenAI's original contract was $11.9 billion (claim 19820), although the May 2025 backlog bridge credits only $11.2 billion to it (4969), and that gap is never explained. OpenAI then signed a September 2025 order form under its existing MSA (4296), described as up to $6.5 billion. NVIDIA signed a $6.3 billion capacity order form (4352, 20209). Meta signed a $14.2 billion order form running to December 2031 (4282) and a second, roughly $21 billion order form dated March 31, 2026 (2123, 10340). Jane Street's roughly $6 billion commitment (10304) is sourced to an announcement rather than a filed order form. The $21 billion Meta figure recurs in five claims (10729, 10694, 10340, 17654, 9317), so it should be counted once.

Several items are framework or relationship language rather than priced contracts. These include the January 2026 NVIDIA "collaboration framework" (3892, 17596, 17598), the IBM Granite partnership (4971), the CrowdStrike "global partnership" (11768), Genesis Mission participation (11778), and the post-merger Core Scientific collaboration (20313). The Anthropic "multi-year agreement" (9318, 10731) is described as signed but carries no value. The "partner of choice" logo rosters (11767, 9320) and the "expanded relationships" with Cohere and Mistral (9319) carry no contractual detail. Two financings explicitly back contracts with unnamed parties: DDTL 3.0 references "a customer contract" (4595), and DDTL 4.0 references a "leading AI enterprise" (11224). Readers should not assume either is a new counterparty.

Concentration has shifted from one buyer to two. Through mid-2025, OpenAI was effectively the entire named book. By late 2025, OpenAI's stated cumulative total of up to $22.4 billion (20198) was a little over half of roughly $43 billion in named commitments. After March 2026, Meta's combined $35.2 billion overtook OpenAI.

The arithmetic only partly reconciles. The record's two OpenAI contracts, $11.9 billion plus $6.5 billion, sum to $18.4 billion, not $22.4 billion. That implies roughly $4 billion in undisclosed or unrecorded OpenAI expansion. Over 2025, RPO rose from $15.1 billion to $60.7 billion (3572, 3555), an increase of $45.6 billion before netting out revenue recognized during the year. The named 2025 deals cover about $43 billion of that increase, and much of it is framed as "up to" ceilings rather than fixed obligations. The residual, plus whatever sits behind the unnamed DDTL customers, is not attributed to any named counterparty in the record.

CoreWeave's record separates cleanly into executed order forms and softer partnership language. The binding instruments are order forms under existing Master Services Agreements: OpenAI (19820, 4296), NVIDIA (4352), and Meta (4282, then a new order form dated March 31, 2026 in 2123, sized at roughly $21 billion in 10340). Jane Street's roughly $6 billion commitment (10304) and the multi-year Anthropic agreement (9318) are described as signed, though Anthropic carries no disclosed value. Several dollar figures are ceilings, not floors. Meta's $14.2 billion is "up to" (4282, 20197), and OpenAI's cumulative $22.4 billion is likewise "up to" (20198). Everything else is weaker. The IBM "partnership" (4971), the CrowdStrike "global partnership" (11768, 10330), Genesis Mission participation (11778), the Core Scientific collaboration (20313), and the January 2026 NVIDIA "collaboration framework" (3892, 17598) describe relationships, not quantified commitments. The "partner of choice" logo lists (11767, 9320, 10338) name companies but disclose no contract terms.

Three counterparties appear only anonymously. One is the "customer contract" financed by DDTL 3.0 (4595). Another is the "leading AI enterprise" behind DDTL 4.0 (11224). The third is the "two large customer contracts" in 9119, a claim that is truncated in the record supplied. Given the timing and scale, these likely map to OpenAI or Meta, but the filings do not say so, and readers should not assume it.

Concentration was acute and has eased only partly. At March 31, 2025, OpenAI's $11.2 billion (4969) was about 43% of the $25.9 billion backlog (4970). By December 31, 2025, OpenAI's up-to $22.4 billion plus Meta's up-to $14.2 billion equal roughly 60% of the $60.7 billion RPO (3560). Adding NVIDIA's $6.3 billion (20209) brings the share to about 70%. The 2026 additions from Meta, Jane Street and Anthropic broaden the named roster, but Meta's $21 billion alone keeps the top two dominant.

The arithmetic reconciles only directionally. RPO rose $45.6 billion during 2025 (3555). Named 2025 deals total about $38.9 billion before burn-down: OpenAI $11.9 billion and $6.5 billion, Meta $14.2 billion, and NVIDIA $6.3 billion. The OpenAI figures also conflict internally: $11.9 billion (19820) versus $11.2 billion (4969). In addition, $11.9 billion plus $6.5 billion falls about $4 billion short of the stated $22.4 billion cumulative total, implying an undisclosed tranche. Because "up to" values may exceed what RPO recognizes, the residual cannot be attributed to named customers from this record. Note also that some claim values (10729, 9317) are mis-scaled as "$21,000" and should be read as $21 billion.

CoreWeave's record splits into a small set of signed, dollar-denominated order forms and a larger layer of collaboration language. The order forms are the hard core. OpenAI's March 2025 agreement (claim 19820, $11.9 billion) and September 2025 order form under an existing MSA (claim 4296) are both signed. So are Meta's September 2025 order form, up to $14.2 billion through December 2031 (claims 4282, 4197), and its March 31, 2026 order form of about $21 billion through 2032 (claims 2123, 10340), and NVIDIA's September 2025 capacity order form (claims 4352, 20209). Jane Street's roughly $6 billion commitment (claims 10304, 10730) and the Anthropic multi-year agreement (claim 9318) are described as signed, but Anthropic carries no disclosed value. The softer tier is substantial. It includes the IBM partnership (claim 4971), the Core Scientific collaboration (claim 20313), the NVIDIA collaboration framework that 'works towards' reference-architecture inclusion (claims 3892, 17598), the CrowdStrike 'global partnership' (claim 11768), Genesis Mission participation (claim 11778), and logo lists of 'AI pioneers' (claims 11767, 9320). None of these carries a disclosed commitment. Three financing disclosures cite large contracts with unnamed counterparties: a 'customer contract' (claim 4595), a 'leading AI enterprise' (claim 11224), and 'two large customer contracts' (claim 9119). These are among the largest capital-deployment events on record and cannot be attributed. Concentration remains heavy despite the 2026 broadening. Named, valued commitments total roughly $70 billion, and OpenAI (up to $22.4 billion, claim 20198) plus Meta (about $35 billion combined) account for over 80% of it. NVIDIA appears simultaneously as supplier, collaborator and $6.3 billion customer, which complicates any read of arm's-length demand. The arithmetic only loosely reconciles. OpenAI's $11.9 billion headline became $11.2 billion of backlog (claim 4969), and the jump from $11.9 billion plus $6.5 billion to a $22.4 billion total implies an undisclosed roughly $4 billion September tranche. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3571). Year-end 2024 RPO plus OpenAI, Meta's first order and NVIDIA sums to about $58 billion, before netting 2025 revenue recognized. That leaves a gap plausibly filled by the unnamed financed contracts. The 'up to' ceilings also make it unclear how much sits in RPO rather than in optional capacity. Meta's $21 billion, Jane Street and Anthropic postdate the $60.7 billion figure and should not be read into it.

CoreWeave's record splits cleanly between executed order forms and softer relationship language. The binding agreements are order forms under existing master services agreements: OpenAI's March 2025 contract (claim 19820) and its September 2025 order form (claim 4296), Meta's September 2025 order form of up to $14.2 billion (claim 4282) and its March 31, 2026 order form of about $21 billion (claims 2123, 10340), and NVIDIA's September 2025 capacity order form (claim 4352). Jane Street's roughly $6 billion commitment (claim 10304) and the signed multi-year Anthropic agreement (claim 9318) are described as firm, though Anthropic carries no disclosed value. The rest is framing rather than contract. The IBM partnership (claim 4971), Core Scientific collaboration (claim 20313), NVIDIA's collaboration framework, which only commits to 'work towards' including CoreWeave software (claims 3892, 17598), the CrowdStrike partnership (claim 11768), Genesis Mission participation (claim 11778), and the 'partner of choice' rosters (claims 11767, 9320) carry no disclosed dollar commitments. Several financed obligations are never named: the contract behind DDTL 3.0 (claim 4595), the 'leading AI enterprise' behind DDTL 4.0 (claim 11224), and 'two large customer contracts' (claim 9119).

Concentration was extreme at the outset. OpenAI's $11.2 billion was about 43 percent of the $25.9 billion backlog reported for March 2025 (claims 4969, 4970). By year-end, OpenAI's up-to $22.4 billion (claim 20198), Meta's $14.2 billion and NVIDIA's $6.3 billion (claim 20209) totaled about $42.9 billion. That equals roughly 71 percent of the $60.7 billion RPO (claim 3571). The 2026 additions shift weight toward Meta, whose cumulative commitments approach $35 billion, rather than diversifying away from hyperscale buyers. NVIDIA's dual role as supplier and capacity buyer also warrants scrutiny.

The arithmetic broadly reconciles but has seams. RPO rose $45.6 billion during 2025, from $15.1 billion (claim 3572) to $60.7 billion, against about $42.9 billion of named additions. That is plausible once revenue recognized during the year is netted out, but the named figures are 'up to' ceilings, so they are not strictly comparable with RPO. OpenAI's initial deal appears as both $11.9 billion (claim 19820) and $11.2 billion (claim 4969). Separately, $22.4 billion less $11.9 billion and $6.5 billion leaves about $4 billion of OpenAI expansion that does not appear in this record. Claim 17654 also describes Meta's $21 billion as a fresh spending commitment, whereas the filings treat it as an expansion layered on the earlier order form.

CoreWeave's record is unusually rich in executed paper, though most of it sits with a few counterparties. The strongest evidence is a set of order forms under existing master services agreements. These cover OpenAI (claim 4296), Meta's initial up-to-$14.2 billion commitment through December 2031 (claim 4282), Meta's roughly $21 billion March 2026 order form running through December 2032 (claims 2123, 10340, 9317), and an NVIDIA capacity purchase (claim 4352). Jane Street's approximately $6 billion commitment (claim 10304) and the Anthropic agreement (claim 9318) are described as committed or signed. No dollar value is given for Anthropic. Softer language should be read differently. The IBM partnership (claim 4971), continued collaboration with Core Scientific (claim 20313), the NVIDIA collaboration framework that will only work towards reference-architecture inclusion (claims 3892, 17598), the CrowdStrike global partnership (claim 11768), Genesis Mission participation (claim 11778), and Exemplar Cloud designations (claims 10332, 9324) are relationship or status claims, not quantified contracts. The partner-of-choice logo lists (claims 11767, 10338, 9320) and the expanded relationships with Cohere and Mistral (claim 9319) name counterparties but attach no economics. Two financing disclosures go the other way, citing material contracts whose counterparty is not named. DDTL 3.0 funds an unidentified customer contract (claim 4595), and DDTL 4.0 serves a leading AI enterprise (claim 11224). Investors are asked to underwrite debt against a customer they cannot see. Concentration has been high and remains so. OpenAI's $11.2 billion was roughly 43 percent of the $25.9 billion March 2025 backlog (claims 4969, 4970). By December 2025, OpenAI's up-to-$22.4 billion (claim 20198) and Meta's $14.2 billion together equal about 60 percent of $60.7 billion in RPO (claim 3555). Meta's March 2026 expansion takes its disclosed commitments past $35 billion, so Jane Street and Anthropic diversify at the margin rather than structurally. The arithmetic only loosely reconciles. RPO rose $45.6 billion during 2025 (claim 3555). The named 2025 deals, OpenAI's $11.9 billion (claim 19820), its $6.5 billion expansion, Meta's $14.2 billion, and NVIDIA's $6.3 billion (claim 20209), sum to about $38.9 billion. The roughly $4 billion OpenAI tranche implied by the $22.4 billion total does not appear on record. Revenue recognized during the year would widen the unexplained residual further. The record also gives two figures for the original OpenAI deal, $11.9 billion versus $11.2 billion (claims 19820, 4969). Repeated up-to phrasing means some RPO may reflect ceilings rather than floors.

CoreWeave's signed-contract core is narrower than its partnership language suggests, but it is real. Three counterparties are documented through order forms under existing Master Services Agreements: OpenAI (claim 4296), Meta (claims 4282, 2123) and NVIDIA (claim 4352). Those filings are the evidentiary anchor. Anthropic is described as a signed multi-year agreement (claim 9318), but no value is disclosed. Jane Street's roughly $6 billion commitment (claim 10304) is named and sized, though the record rests on announcement and press language rather than a filed order form (claims 10730, 10695). Everything else is softer. The NVIDIA collaboration framework (claim 3892) and reference-architecture work (claim 17598) are cooperation, not revenue. The same applies to the IBM partnership (claim 4971), the CrowdStrike "global partnership" (claim 11768), Genesis Mission participation (claim 11778) and continued collaboration with Core Scientific (claim 20313). The "partner of choice" logo lists (claims 11767, 10338, 9320) name companies without contract terms. Two financing disclosures point to material contracts with unidentified parties: the "customer contract" behind DDTL 3.0 (claim 4595) and the "leading AI enterprise" behind DDTL 4.0 (claim 11224). Neither names the customer. Concentration has broadened from a near-single-customer book to a small club. In March 2025 OpenAI dominated (claim 19820). Meta arrived in September 2025 (claim 4282). By year-end, RPO had grown from $15.1 billion to $60.7 billion (claims 3572, 3555). OpenAI at up to $22.4 billion (claim 20198), Meta at up to $14.2 billion and NVIDIA at $6.3 billion (claim 20209) sum to about $42.9 billion, roughly 71% of that RPO. The March 2026 Meta order for about $21 billion (claims 2123, 10340), described as new rather than a restatement (claim 9317), would lift Meta to about $35 billion and make it the largest named exposure. The revenue arithmetic does not fully reconcile. OpenAI's initial deal is $11.9 billion in one disclosure (claim 19820) and $11.2 billion of backlog in another (claim 4969). Neither figure plus the $6.5 billion expansion reaches the stated $22.4 billion total, which leaves about $4 billion unattributed in this record. Backlog and RPO are also different measures. At March 2025, backlog was $25.9 billion against RPO of $14.7 billion (claim 4970), with the gap contingent on delivery and availability requirements. Repeated "up to" phrasing implies headline figures are ceilings rather than floors. The Meta, Jane Street and Anthropic deals post-date the December 2025 RPO and should appear in 2026 disclosures. The supplied record is truncated at claim 9121, so any later reconciliation is not reflected here.

CoreWeave's signed-contract record is narrower than its partnership vocabulary. The commitments documented as executed order forms under existing master services agreements come from four counterparties. OpenAI signed an initial deal recorded at $11.9 billion in the press release but at $11.2 billion of backlog in the filing (19820, 4969), then a September 2025 order form of up to roughly $6.5 billion, which brought commitments to up to $22.4 billion (4296, 20198). Meta signed up to $14.2 billion through December 2031 (4282, 20197) and a further order form of about $21 billion through 2032 on March 31, 2026 (2123, 10340, 9317). NVIDIA's $6.3 billion capacity purchase is an order form (4352, 20209). Jane Street committed about $6 billion (10304). Anthropic's agreement is described as signed but carries no disclosed value (9318), and the Cohere and Mistral expansions are likewise unquantified (9319). Everything else is softer. The January 2026 NVIDIA arrangement is a 'collaboration framework' with test-and-validate objectives (3892, 17598). IBM, CrowdStrike, Genesis Mission and Core Scientific are partnerships or participations with no economics attached (4971, 11768, 11778, 20313). The rotating 'partner of choice' logo lists name customers without commitments (11767, 9320). Several financings reference a material contract with an unnamed counterparty, either 'a customer contract' or a 'leading AI enterprise' (4595, 11224, 9119), so the largest debt-backed obligations are not tied to a disclosed name.

Concentration has shifted from one counterparty to two without diluting. At March 31, 2025, OpenAI's $11.2 billion was about 43% of $25.9 billion of backlog. Backlog less the OpenAI deal equals RPO exactly ($14.7 billion), which suggests the OpenAI deal sat outside RPO pending delivery conditions (4970). By December 31, 2025, RPO reached $60.7 billion against $15.1 billion a year earlier (3555, 3572). OpenAI and Meta ceilings alone total $36.6 billion, about 60% of that figure, and about 71% with NVIDIA included. Meta's March 2026 expansion pushes Meta's cumulative ceilings to about $35 billion.

The arithmetic broadly reconciles, but only if the 'up to' ceilings are booked near full value. Named deals leave roughly $18 billion of year-end RPO for all other customers, a plausible extension of the prior base. However, the OpenAI $11.9 versus $11.2 billion gap is unexplained. Several Meta and Jane Street figures rest on secondary press restatements (10729, 10730, 10694) rather than company text.

CoreWeave's record is unusually heavy with executed paper, but it mixes signed contracts with lighter partnership language. The firmest disclosures are order forms under existing master services agreements. These cover OpenAI (claim 4296), Meta (claims 4282 and 2123) and NVIDIA (claim 4352), and each states either a dollar ceiling or an execution date. Jane Street's roughly $6 billion commitment (claims 10304 and 10730) and the multi-year Anthropic agreement (claim 9318) are described as signed, though no value is disclosed for Anthropic. Weaker language covers the IBM partnership (claim 4971), Core Scientific collaboration (claim 20313), the NVIDIA collaboration framework that only 'works towards' inclusion in reference architectures (claims 3892 and 17598), the CrowdStrike 'global partnership' (claims 11768 and 10330), Genesis Mission participation (claim 11778) and the Exemplar Cloud designation (claims 10332 and 9324). None of these carries disclosed economics. The 'partner of choice' rosters (claims 11767, 10338 and 9320) and the 'expanded relationships' with Cohere and Mistral (claim 9319) name counterparties without contract terms. Two financing disclosures cite an unnamed 'customer contract' (claim 4595) and a 'leading AI enterprise' (claim 11224), so the counterparties that actually anchor the DDTL 3.0 and 4.0 debt are never identified. Concentration has become more extreme while also widening. In March 2025 OpenAI was effectively the only named anchor. By November 2025 the company cited OpenAI at up to $22.4 billion (claim 20198), Meta at up to $14.2 billion (claim 20197) and NVIDIA at $6.3 billion (claim 20209), about $42.9 billion in total. That equals roughly 71% of the $60.7 billion RPO at year-end 2025 (claim 3555), if those 'up to' ceilings sit fully in RPO. Meta's further $21 billion order, signed March 31, 2026 (claims 2123 and 9317), would take its total to about $35 billion, so Meta now overtakes OpenAI as the largest single exposure. The revenue arithmetic does not reconcile cleanly. The OpenAI deal was announced at $11.9 billion (claim 19820) but booked as $11.2 billion of backlog (claim 4969). RPO actually fell from $15.1 billion (claim 3572) to $14.7 billion (claim 4970) in the quarter the OpenAI deal closed. That suggests the deal initially sat in the conditional 'other amounts' portion of backlog rather than in RPO. The 'up to' phrasing throughout means headline totals are maximum figures, not guaranteed revenue. No post-March 2026 RPO figure is on record, so the Meta expansion, the Jane Street commitment and the Anthropic deal cannot yet be tested against reported obligations.

CoreWeave's record separates into two groups. One is contracts executed under master services agreements. The other is partnership and logo language that carries no disclosed economics. The firmest evidence is the order forms filed with the SEC: OpenAI (4296), Meta's initial order form for up to $14.2 billion through December 2031 (4282), Meta's March 31, 2026 order form (2123) valued at roughly $21 billion (10340, 9317), and an NVIDIA capacity order form (4352). Jane Street's roughly $6 billion commitment (10304) and the Anthropic agreement (9318) are described as signed, but Anthropic's value is never disclosed. NVIDIA's later items read differently. The $6.3 billion "strategic collaboration" (20209) and the January 2026 "collaboration framework" (3892, 17596, 17598) use framework language, and the latter commits to testing and validation rather than spend.

The softer layer is large. IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), the Genesis Mission (11778, 10331) and the Exemplar Cloud designation (10332, 9324) are partnerships, memberships or badges. The recurring "partner of choice" logo lists (11767, 10338, 9320) and the Cohere/Mistral "expanded relationships" (9319) name counterparties but give no terms. Two financing disclosures fund contracts with counterparties that go unnamed: "a customer contract" behind DDTL 3.0 (4595) and a "leading AI enterprise" behind DDTL 4.0 (11224). These are material obligations whose counterparty the reader has to infer.

Concentration has moved from one anchor to two. In early 2025, OpenAI alone accounted for most of the incremental backlog (4969). By late 2025, OpenAI had reached up to about $22.4 billion (20198) and Meta $14.2 billion. After March 2026, Meta's stated commitments total about $35 billion, overtaking OpenAI. Most headline values are "up to" figures, which caps them rather than guaranteeing them.

The arithmetic broadly reconciles, with caveats. RPO of $60.7 billion at December 31, 2025 (3555) compares with about $43 billion from OpenAI, Meta's first order form and NVIDIA. That leaves about $18 billion for the legacy book and smaller customers, which is plausible given $15.1 billion of RPO a year earlier (3572). The OpenAI figure itself is inconsistent: $11.9 billion in one disclosure (19820) and $11.2 billion of backlog in another (4969). March 2025 backlog of $25.9 billion included only $14.7 billion of RPO (4970), so headline deal values can sit outside RPO until delivery conditions are met. Claim 10729 is dated before the March 2026 Meta expansion it describes, and several values (10729, 3555) carry unit-scaling errors in the record.

CoreWeave's signed-contract record rests on three counterparties: OpenAI, Meta, and NVIDIA. Each is documented through order forms under existing Master Services Agreements. OpenAI's are claims 4296 and 19820, Meta's are 4282 and 2123, and NVIDIA's is 4352. The company calls OpenAI's total commitment "up to approximately $22.4 billion" (claim 20198). That figure does not cleanly reconcile with the record. The original contract is booked at $11.9 billion (claim 19820) but appears as $11.2 billion of added backlog (claim 4969). The $6.5 billion expansion leaves roughly $4 billion of OpenAI commitments with no matching claim. Meta's first order form, $14.2 billion through December 2031 (claim 4282), was followed by a roughly $21 billion order dated March 31, 2026 (claims 2123, 10340), putting Meta near $35 billion. NVIDIA's $6.3 billion "strategic collaboration" (claim 20209) and later "collaboration framework" (claim 3892) are framework-style language. That relationship is circular because NVIDIA is also CoreWeave's supplier, so it should be weighed differently from arm's-length demand. Below that tier, Jane Street's roughly $6 billion commitment (claim 10304) is named and sized. Anthropic's "multi-year agreement" is described as signed (claim 9318) but no dollar figure is attached. IBM (claim 4971), CrowdStrike (claim 11768), and Core Scientific (claim 20313) appear only as "partnerships." The logo lists of "partner of choice" names (claims 11767, 9320) and the Genesis Mission and Exemplar Cloud designations (claims 11778, 9324) are not contract disclosures. Several financings point to customers left unnamed: "a customer contract" for DDTL 3.0 (claim 4595), a "leading AI enterprise" for DDTL 4.0 (claim 11224), and "two large customer contracts" (claim 9119, truncated). Concentration has moved from one name to two. In March 2025, OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog (claims 4969, 4970). Remaining performance obligations then grew from $15.1 billion to $60.7 billion by December 31, 2025 (claims 3555, 3572). Even counting Meta's first order and OpenAI's full ceiling, OpenAI and Meta account for about $36.6 billion, roughly 60% of that total. Meta's $21 billion and Jane Street's $6 billion came after year-end and will add further weight at the top. The arithmetic only partly reconciles. Much of the headline figures are "up to" ceilings, while RPO reflects contracted amounts, so treating the ceilings as RPO probably overstates named-customer coverage. That leaves about $24 billion of year-end RPO either attributed to unnamed counterparties or explained by the gap between ceilings and recognized obligations.

CoreWeave's record contains a small set of signed, dollar-denominated order forms and a much larger set of looser relationship language. The signed tier is clear. OpenAI's agreements were filed as order forms under an existing MSA (19820, 4296), and the company now cites commitments of up to about $22.4 billion (20198). Meta signed an up-to-$14.2 billion order form running through December 2031 (4282) and a roughly $21 billion order form dated March 31, 2026, running through 2032 (2123, 10340). NVIDIA signed a $6.3 billion capacity order form under its own MSA (4352, 20209), which makes CoreWeave's key supplier also a material customer. Jane Street committed about $6 billion (10304). Anthropic signed a multi-year agreement with no disclosed value (9318).

Everything else is softer. These include the NVIDIA 'collaboration framework' and reference-architecture work (3892, 17596, 17598), the IBM and CrowdStrike 'partnerships' (4971, 11768), Genesis Mission participation (11778), and continued 'collaboration' with Core Scientific (20313). None of them carries a value or confirms a contract. The 'partner of choice' logo lists (11767, 9320) and the 'expanded relationships' with Cohere and Mistral (9319) do not establish contract size. Two financings are tied to unnamed counterparties: DDTL 3.0 funds 'a customer contract' (4595), and DDTL 4.0 serves a 'leading AI enterprise' (11224). Readers should not assume which counterparty either refers to.

Concentration has shifted, but it remains high. Through 2025, OpenAI, Meta and NVIDIA account for roughly $42.9 billion of named ceilings, about 71% of the $60.7 billion in year-end RPO (3555). In 2026, Meta's expansion lifts it to about $35.2 billion and makes it the largest named counterparty. Jane Street and Anthropic widen the base only modestly.

The arithmetic only partly reconciles. The initial OpenAI figure appears as $11.9 billion (19820) and also as $11.2 billion of added backlog (4969). Adding the $6.5 billion expansion to either figure falls short of $22.4 billion, which implies roughly $4 billion of OpenAI commitments that are not in this record. RPO rose from $15.1 billion to $60.7 billion (3572, 3555). Named 2025 deals explain most of that $45.6 billion increase, but only if each 'up to' ceiling is booked in full. That would leave little room for other customers or for revenue recognized during the year. Earlier backlog figures also mixed RPO with amounts contingent on delivery (4970). Some secondary coverage merges the $14.2 billion and $21 billion Meta deals into one figure (17654, 17655), but the filings show two separate order forms.

CoreWeave's announced deal flow rests on a small number of order forms signed under existing master services agreements, surrounded by a wider ring of partnership language that carries no disclosed economics.

The signed contracts are with OpenAI, Meta, NVIDIA and Jane Street. OpenAI's March 2025 agreement was valued at $11.9 billion (claim 19820) but booked as $11.2 billion of backlog (claim 4969), a gap the record does not explain. It was followed by a September 2025 order form (claim 4296) and an expansion of up to $6.5 billion, which brought total OpenAI commitments to up to $22.4 billion (claim 20198). Meta signed an order form for up to $14.2 billion through 2031 (claim 4282) and a second, roughly $21 billion order form on March 31, 2026 (claims 2123, 9317). Several later items (claims 10729, 10694, 10340, 17654, 17655) restate those same two Meta contracts and should not be counted again. NVIDIA's September 2025 capacity order form (claim 4352) appears to be the $6.3 billion collaboration (claim 20209). Jane Street committed about $6 billion (claim 10304). The Anthropic agreement is described as signed (claim 9318), but its value is undisclosed.

The softer material is a different category. It includes the January 2026 NVIDIA "collaboration framework" (claim 3892), the IBM and CrowdStrike "partnerships" (claims 4971, 11768), Genesis Mission participation (claim 11778), continued collaboration with Core Scientific (claim 20313), and logo lists of customers with no contract values (claims 11767, 9320, 9319). Three financing-linked disclosures leave the counterparty unnamed: "a customer contract" (claim 4595), a "leading AI enterprise" (claim 11224) and "two large customer contracts" (claim 9119).

Concentration remains heavy even as it shifts. In early 2025 the book was effectively OpenAI. By the end of 2025, OpenAI's up-to $22.4 billion amounted to roughly 37% of $60.7 billion in RPO (claim 3555), and Meta's $14.2 billion to about 23%. The March 2026 Meta order lifts Meta's cumulative commitments to roughly $35 billion, likely making it the largest counterparty.

The arithmetic broadly reconciles, with one caveat. RPO rose by $45.6 billion in 2025 (claims 3572, 3555), while the named OpenAI, Meta and NVIDIA signings that year total about $42.9 billion. That leaves little room for every other customer, especially after netting out revenue recognized during the year. Those deal values are also expressed as "up to" figures. Separately, RPO slipped to $14.7 billion in March 2025 even as backlog reached $25.9 billion (claim 4970). That pattern suggests the first OpenAI deal initially sat in delivery-contingent backlog rather than in RPO.

CoreWeave's record contains real signed contracts, but they sit among partnership, framework and logo-list language that should not be counted the same way. The executed agreements are order forms under existing master services agreements with OpenAI (claims 19820, 4296), Meta (4282, 2123) and NVIDIA (4352). The amounts are disclosed: OpenAI up to roughly $22.4 billion cumulative (20198), Meta up to $14.2 billion followed by a roughly $21 billion March 2026 order (10340, 9317), and NVIDIA $6.3 billion (20209). Jane Street's roughly $6 billion commitment (10304) is also described as committed. The Anthropic deal is called a signed multi-year agreement (9318), but no value is given. Everything else is softer. The IBM tie-up is a "partnership" (4971). CrowdStrike is a "global partnership" (11768, 10330). The January 2026 NVIDIA expansion is explicitly a "collaboration framework" (3892), and part of it only commits the parties to "work towards" reference-architecture inclusion (17598). Genesis Mission is participation in an initiative (11778). The "partner of choice" lists naming Cursor, Midjourney, Perplexity and others (11767, 9320) carry no contract terms. Several material obligations are tied to counterparties the company does not name. DDTL 3.0 financed "a customer contract" (4595), DDTL 4.0 serves a "leading AI enterprise" (11224), and a later filing cites "two large customer contracts" (9119). In each case, readers cannot tell whether these are the named deals above or additional ones.

Concentration was extreme early on and has moved toward two names rather than spreading out. At March 31, 2025, the OpenAI deal alone was about 43% of $25.9 billion of backlog (4969, 4970). By December 31, 2025, RPO had reached $60.7 billion (3555). The OpenAI, Meta and NVIDIA ceilings sum to about $42.9 billion of that, roughly 70%. That share is overstated to the extent the deals are "up to" figures, and RPO is a narrower measure than backlog. After the March 2026 Meta order, Meta's cumulative commitments of about $35 billion likely surpass OpenAI's.

The arithmetic holds together only loosely. The first OpenAI deal appears as $11.9 billion in one source (19820) and $11.2 billion of backlog in another (4969). The $22.4 billion cumulative figure implies the September 2025 order form (4296) was worth around $4 billion, but that number is never stated. Secondary coverage blurs the original $14.2 billion Meta order with the separate $21 billion expansion (17655, 17654). One claim also dates the March 2026 Meta expansion to March 1 (10729), while the order form itself is dated March 31 (2123). Finally, about $18 billion of December 2025 RPO cannot be tied to any named counterparty in the record.

CoreWeave's disclosed signed commitments come from a small number of named counterparties, and most are documented as order forms under existing master services agreements. OpenAI is the clearest case. It has an initial agreement described at $11.9 billion in one filing (19820) and $11.2 billion of backlog in another (4969), a September 2025 order form (4296), and a further expansion of up to $6.5 billion that brings its total to up to $22.4 billion (20198). Meta signed an order form worth up to $14.2 billion through December 2031 (4282, 17655) and a separate March 31, 2026 order form of about $21 billion running through 2032 (2123, 10340, 9317). NVIDIA's $6.3 billion order form (4352, 20209) is contractual. Its January 2026 expansion, however, is a collaboration framework (3892) with work-towards language on reference architectures (17598). NVIDIA is also CoreWeave's GPU supplier, so its role as a customer deserves separate scrutiny. Jane Street's roughly $6 billion commitment (10304, 10730) and the Anthropic agreement (9318, 10731) are described as signed, but Anthropic's value is undisclosed. Softer language covers IBM (4971), CrowdStrike (11768, 10330), Core Scientific (20313), the Genesis Mission (11778), and the logo lists and expanded relationships with Cohere, Mistral and others (11767, 9320, 9319), none of which carries a dollar figure. Two financing disclosures tie debt to unnamed counterparties: a customer contract for DDTL 3.0 (4595) and a leading AI enterprise for DDTL 4.0 (11224). Readers cannot tell whether these are OpenAI, Meta, or someone else. Concentration has intensified rather than diversified. In March 2025, OpenAI's $11.2 billion was about 43% of $25.9 billion of backlog (4969, 4970). By December 2025, OpenAI's up-to $22.4 billion, Meta's up-to $14.2 billion and NVIDIA's $6.3 billion total about $42.9 billion, roughly 71% of the $60.7 billion RPO (3555, 3572). The arithmetic reconciles only loosely. The up-to amounts are ceilings, and backlog has included amounts beyond RPO that depend on delivery conditions (4970). Some OpenAI value has presumably already been recognized, and the $11.9 billion versus $11.2 billion gap is unexplained. The 2026 additions of Meta's $21 billion, Jane Street's $6 billion and Anthropic come after the latest RPO figure and have not yet been tested against a reported balance. On headline figures, Meta and OpenAI together account for about $57.6 billion of named commitments. Meta's further expansion option (4283) remains optional and should not be counted.

CoreWeave's record separates cleanly into signed order forms with named, sized counterparties and softer language that carries no disclosed value. The contracts with the strongest evidence are OpenAI, Meta and NVIDIA, each documented as order forms under existing master services agreements (claims 4296, 2123, 4352). OpenAI's commitments reach up to approximately $22.4 billion (20198). Meta committed up to $14.2 billion through 2031 (4282), then signed a new $21 billion order form in March 2026 (9317, 10340). NVIDIA's $6.3 billion capacity purchase (4352, 20209) is contracted, but the January 2026 NVIDIA "collaboration framework" (3892, 17596) and the pledge to "work towards" reference-architecture inclusion (17598) are frameworks, not revenue. Jane Street's roughly $6 billion commitment (10304) is named and sized. Anthropic's agreement is described as signed and multi-year (9318), but no value is given.

The weaker layer is larger than it looks. The IBM partnership (4971), Core Scientific collaboration (20313), CrowdStrike "global partnership" (11768), Genesis Mission participation (11778), and the "partner of choice" logo lists naming Cognition, Cursor, Perplexity and others (11767, 9320) have no contract values attached. Cohere and Mistral "expanded relationships" (9319) are also unsized. Two financing disclosures cite an unidentified "customer contract" (4595) and a "leading AI enterprise" (11224). These are material enough to justify dedicated delayed-draw term loans, yet the counterparties are never named. The DDTL banks (11225) are lenders, not customers.

Concentration has shifted from OpenAI-dominant in early 2025 to a Meta-OpenAI duopoly. If the March 2026 $21 billion is incremental, as the "new commitment" wording suggests (9317), Meta's roughly $35 billion now exceeds OpenAI's. Some coverage frames it as an expansion "by about $21 billion" (10729), which leaves that reading slightly ambiguous.

The arithmetic mostly holds, with two frictions. OpenAI's initial deal was reported at $11.9 billion (19820) but booked as $11.2 billion of backlog (4969). Adding $6.5 billion to $11.9 billion gives $18.4 billion, not $22.4 billion, so about $4 billion of intermediate OpenAI expansion is not itemized here. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3571), an increase of $45.6 billion. Named 2025 deals total about $42.9 billion, and 2025 revenue also drew down backlog, so unnamed contracts must account for a meaningful share of bookings. Most headline figures are "up to" ceilings, so realized value depends on delivery conditions (4970).

CoreWeave's contract record separates cleanly into two tiers. The first tier is signed order forms under existing master services agreements, disclosed in 8-Ks. These include OpenAI (claims 19820, 4296), Meta (claims 4282, 2123) and NVIDIA's capacity purchase (claim 4352). They carry dollar figures and end dates. Meta's initial commitment runs to about $14.2 billion through December 14, 2031 (claim 4282), and a March 31, 2026 order form adds roughly $21 billion through December 2032 (claims 2123, 10340). Jane Street's approximately $6 billion commitment (claim 10304) also reads as a signed agreement. Anthropic's multi-year agreement is described as signed (claim 9318) but has no disclosed value.

The second tier is softer. It includes the NVIDIA 'collaboration framework' (claim 3892), which is separate from the $6.3 billion order (claim 20209). It also covers the IBM and CrowdStrike 'partnerships' (claims 4971, 11768), continued collaboration with Core Scientific (claim 20313), Genesis Mission participation (claim 11778), and the Exemplar Cloud designation (claim 10332). Logo lists of 'partners of choice' also belong here, such as Cursor, Midjourney, Perplexity and Hudson River Trading (claims 11767, 9320). None of these carries a disclosed amount, so none should be read as backlog.

Several material contracts are left unnamed. DDTL 3.0 financed 'a customer contract' (claim 4595). DDTL 4.0 funds services for a 'leading AI enterprise' (claim 11224). A later filing references 'two large customer contracts' (claim 9119). Whether these map to OpenAI, Meta or someone else is not stated.

Concentration has shifted from OpenAI to Meta without spreading out much. OpenAI totals reach up to about $22.4 billion (claim 20198), and Meta now stands near $35.2 billion combined. The OpenAI arithmetic also has gaps. The $11.9 billion announcement (claim 19820) becomes $11.2 billion of backlog (claim 4969). Adding the $6.5 billion expansion to $11.9 billion gives only $18.4 billion, which implies an interim tranche of about $4 billion that is not on this record.

RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3560). The named 2025 deals are OpenAI up to $22.4 billion, Meta $14.2 billion and NVIDIA $6.3 billion, about $42.9 billion together. Adding those to opening RPO roughly bridges to the year-end figure before subtracting recognized revenue. Two caveats keep this directional only. Many of the figures are 'up to' ceilings, and 2025 revenue burn is not in the record. Even on these numbers, three counterparties account for about 71% of year-end RPO.

CoreWeave's record separates cleanly into executed order forms and softer partnership language, and the two should not be weighted equally. The executed contracts are order forms under existing Master Services Agreements with three counterparties. OpenAI signed one in March 2025 for $11.9 billion (19820) and another on September 23, 2025 (4296). Meta signed one for up to about $14.2 billion through December 2031 (4282) and another on March 31, 2026 (2123), described as about $21 billion through 2032 (10340). NVIDIA signed a capacity-purchase order form (4352). Jane Street's roughly $6 billion commitment (10304) is named and sized, but the record contains no order-form filing for it. Anthropic's multi-year agreement is described as signed (9318) but carries no dollar value. Everything else uses softer language. NVIDIA's January 2026 arrangement is explicitly a "collaboration framework" (3892, 17596). IBM (4971) and CrowdStrike (11768) are "partnerships" with no value attached. Core Scientific is framed as continued collaboration (20313). The Genesis Mission is a government initiative, not a customer (11778). The logo lists naming Cognition, Cursor, Perplexity and others (11767, 9320, 9319) disclose no terms. Several of the most consequential contracts are unnamed: a "customer contract" behind DDTL 3.0 (4595), a "leading AI enterprise" behind DDTL 4.0 (11224), and "two large customer contracts" (9119; the record is truncated at that claim). Concentration has shifted. Meta has overtaken OpenAI, with roughly $35 billion in cumulative commitments against OpenAI's stated $22.4 billion (20198). Jane Street and Anthropic add some breadth, but the book remains dominated by two or three hyperscale buyers. The revenue arithmetic does not fully reconcile. OpenAI's initial deal appears as $11.9 billion in one disclosure (19820) and as $11.2 billion of backlog in another (4969). The later $6.5 billion expansion is said to bring the total to $22.4 billion (20198), which leaves roughly $4 billion unaccounted for in the claims on record. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of about $45.6 billion. The named 2025 deals (OpenAI's $22.4 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion, per 20209) total about $43 billion before netting revenue recognized during the year. That leaves a gap of several billion dollars, most plausibly filled by the unnamed financed contracts. The earlier backlog figure of $25.9 billion also exceeded RPO of $14.7 billion (4970) because it included amounts conditional on delivery, so headline deal values should not be read as firm RPO.

CoreWeave's disclosed book rests on a small number of executed order forms under existing Master Services Agreements. The rest of its partnership vocabulary should be read as something softer. The hardest evidence is the 8-K language on three counterparties. OpenAI signed an order form on September 23, 2025 (claim 4296), following the March agreement (claims 19820, 4969). Meta signed order forms carrying an up-to $14.2 billion commitment through December 2031 (claim 4282) and a further roughly $21 billion commitment on March 31, 2026 (claims 2123, 9317, 10340). NVIDIA's September 2025 order form is a capacity purchase by a supplier-investor rather than an arm's-length end customer (claim 4352). Jane Street's approximately $6 billion commitment (claim 10304) and the signed Anthropic agreement (claim 9318) are named and described as committed. The Anthropic deal carries no disclosed value. Against these, the NVIDIA 'collaboration framework' and reference-architecture work toward inclusion (claims 3892, 17596, 17598) are framework language. So are the IBM, CrowdStrike and Core Scientific 'partnerships' (claims 4971, 11768, 20313), Genesis Mission participation (claim 11778), and the rotating 'partner of choice' logo lists (claims 11767, 9320). None discloses a dollar commitment. Cohere and Mistral appear only as 'expanded relationships' (claim 9319). Several of the most capital-intensive obligations are unnamed. DDTL 3.0 finances 'a customer contract' (claim 4595), and DDTL 4.0 serves a 'leading AI enterprise' (claim 11224). Another disclosure references 'two large customer contracts' (claim 9119), and the record for that claim is truncated. Concentration has broadened in name but not obviously in weight. In early 2025, OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog (claims 4969, 4970). By late 2025, OpenAI's up-to $22.4 billion (claim 20198) plus Meta's $14.2 billion equal about 60% of the $60.7 billion year-end RPO (claim 3555). Adding NVIDIA's $6.3 billion (claim 20209) brings that to roughly 70%. Meta's 2026 expansion makes it the single largest disclosed counterparty. The arithmetic only loosely reconciles. RPO rose $45.6 billion during 2025 (claims 3572, 3555). The identified 2025 deals sum to roughly $39 billion, which leaves a gap, before revenue burn, that the named list does not explain. OpenAI's own figures diverge: $11.9 billion versus $11.2 billion for the initial deal (claims 19820, 4969). Its $22.4 billion cumulative total also implies an unitemized tranche of about $4 billion. Many headline figures are 'up to' ceilings. Backlog also includes amounts contingent on delivery, beyond contractual RPO (claim 4970). Finally, the record does not settle whether Meta's $21 billion adds to the $14.2 billion or partly overlaps it (claims 17654, 17655).

CoreWeave's record splits into a small set of contracts backed by filed order forms and a much larger ring of partnership and logo language. Only a few are documented as executed order forms under existing master services agreements. These are OpenAI (claim 4296), Meta in September 2025 at up to about $14.2 billion through December 2031 (claims 4282, 17655) and again on March 31, 2026 at about $21 billion through December 2032 (claims 2123, 10340), and NVIDIA (claim 4352). Several secondary items repeat the same Meta $21 billion tranche (claims 10729, 10694, 17654, 9317) and should not be counted twice. Jane Street's roughly $6 billion commitment is named and sized (claim 10304). Anthropic is named as a signed multi-year agreement but with no value (claims 9318, 10731).

The rest is softer. The IBM Granite partnership (claim 4971), the CrowdStrike global partnership (claims 11768, 10330), the Genesis Mission (claims 11778, 10331) and the Core Scientific collaboration (claim 20313) carry no dollar terms. The January 2026 NVIDIA arrangement is explicitly a collaboration framework (claim 3892) and only commits to work towards reference-architecture inclusion (claim 17598). The "partner of choice" rosters (claims 11767, 9320, 9319) are logo lists, not contract disclosures. The counterparties that most directly drive capital spending stay unnamed. Examples include the customer financed by DDTL 3.0 (claim 4595), the "leading AI enterprise" behind DDTL 4.0 (claim 11224) and two large contracts in claim 9119, which is truncated in this record.

Concentration has shifted from OpenAI-led in early 2025 to a Meta-and-OpenAI duopoly. Meta's named commitments total about $35.2 billion, against up to $22.4 billion for OpenAI (claim 20198). NVIDIA's $6.3 billion (claim 20209) is a supplier also acting as a buyer of capacity, which weakens it as independent demand.

The arithmetic only partly reconciles. OpenAI is $11.9 billion in claim 19820 but $11.2 billion in claim 4969. Adding $6.5 billion to $11.9 billion gives $18.4 billion, not $22.4 billion, so an intermediate tranche of about $4 billion is missing from this record. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3555, 3572), a $45.6 billion increase. The named 2025 deals cover roughly $42.9 billion before subtracting revenue recognized during the year, which this record does not provide, so part of the increase likely comes from unnamed contracts. The $25.9 billion backlog figure (claim 4970) also includes non-RPO amounts that depend on delivery conditions. "Up to" wording across the Meta and OpenAI deals means the headline figures are ceilings rather than guaranteed revenue.

CoreWeave's signed-contract base is narrower than its partnership roster suggests. Only four counterparties carry both a name and a dollar figure backed by an executed order form under an existing master services agreement: OpenAI (claims 19820, 4296, 20198), Meta (4282, 2123, 10340), NVIDIA (4352, 20209) and Jane Street (10304). Anthropic is described as a signed multi-year agreement (9318) but no value is disclosed. The rest of the record uses softer language. IBM (4971), CrowdStrike (11768, 10330) and Core Scientific (20313) are 'partnerships' or 'collaboration.' The January 2026 NVIDIA arrangement is explicitly a 'collaboration framework' whose software commitments are to 'work towards' inclusion (3892, 17598). Genesis Mission and Exemplar Cloud are program memberships or designations, not revenue contracts (11778, 10332, 9324). The logo lists naming Cursor, Perplexity, Mistral and others carry no amounts (11767, 9320, 9319). Three material contracts go unnamed: the customer financed by DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), and 'two large customer contracts' (9119, which is truncated in the record).

Concentration has changed hands rather than eased. In March 2025, OpenAI's $11.2 billion accounted for roughly 43% of $25.9 billion in backlog (4969, 4970). By year-end, OpenAI's up-to-$22.4 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion summed to about $42.9 billion, roughly 71% of the $60.7 billion RPO (3555). NVIDIA is also CoreWeave's GPU supplier, which makes it a circular counterparty. The March 2026 Meta order form for about $21 billion (9317, 10340) makes Meta, not OpenAI, the largest named exposure, at roughly $35 billion if the new form is additive to the 2025 one. The record implies it is additive but never states a cumulative figure (17654, 17655).

The arithmetic reconciles only loosely. OpenAI's initial deal appears as $11.9 billion in one filing (19820) and $11.2 billion of backlog in another (4969). Adding $6.5 billion to $11.9 billion gives $18.4 billion, not the stated $22.4 billion total (20198), so an increment of about $4 billion is missing from the record. RPO grew $45.6 billion during 2025 (3572, 3571), which is broadly consistent with the named 2025 deals net of recognized revenue. However, most of those figures are 'up to' ceilings, and the backlog metric counts conditional amounts that sit outside RPO (4970). No post-December-2025 RPO appears in the record, so the Meta and Jane Street additions cannot yet be tested against disclosed obligations.

CoreWeave's signed commitments are concentrated in a small number of order forms executed under existing Master Services Agreements. The record ties those to four named counterparties. OpenAI has a March 2025 agreement (claim 19820) and a September 2025 order form (claim 4296). Meta has a September 2025 order form (claims 4282, 4283) and a March 31, 2026 order form (claims 2123, 10340). NVIDIA has a September 2025 capacity order form (claim 4352), which the company later sized at $6.3 billion (claim 20209). Jane Street's roughly $6 billion commitment (claim 10304) is the fourth. Everything else carries softer language. The January 2026 NVIDIA relationship is a "collaboration framework" whose stated deliverables are testing and validation work (claims 3892, 17596, 17598). IBM (claim 4971), CrowdStrike (claims 11768, 10330), Core Scientific (claim 20313) and the DOE Genesis Mission (claims 11778, 10331) are described as partnerships or participation, with no dollar value or term. Anthropic's "multi-year agreement" (claims 9318, 10731) is signed but unpriced. The logo lists naming Cursor, Perplexity, Hudson River Trading and others as organizations for which CoreWeave is "partner of choice" (claims 11767, 10338, 9320) are marketing, not contract disclosures. Two contracts that justified debt financing are left unnamed. DDTL 3.0 financed "a customer contract" (claim 4595), and DDTL 4.0 serves a "leading AI enterprise" (claim 11224), so a reader cannot attribute that leverage to a specific counterparty. Concentration shifted from OpenAI toward Meta but did not broaden. OpenAI's commitments reached up to $22.4 billion (claim 20198). Meta's two tranches sum to roughly $35.2 billion. Against the $60.7 billion of RPO at December 31, 2025 (claim 3555), OpenAI plus the first Meta tranche plus NVIDIA come to about $42.9 billion, roughly 70 percent. That figure is directionally useful but not exact, because the OpenAI and Meta amounts are "up to" ceilings. The OpenAI arithmetic does not fully reconcile on the record. The $11.9 billion initial contract (claim 19820) and the $6.5 billion expansion (claim 20198) sum to $18.4 billion, which implies an undisclosed tranche of about $4 billion, plausibly the September 2025 order form (claim 4296). The company also booked $11.2 billion, not $11.9 billion, into backlog (claim 4969). RPO rose by $45.6 billion during 2025 (claims 3555, 3572), which is consistent with the named OpenAI, Meta and NVIDIA signings. The Meta $21 billion, Jane Street and Anthropic deals came after year end, and no later RPO figure in this record lets them be tested. Some secondary-source claims are dated before the Meta order form they describe (claim 10729). The claim list is also truncated at claim 9121, so any later disclosures are not reflected here.

CoreWeave's signed-contract record is thick and specific, but several of its headline 'partnerships' are not customer agreements. The OpenAI, Meta and NVIDIA relationships are documented as order forms under existing Master Services Agreements (claims 4296, 4282, 4352, 2123). These are the hard contracts. OpenAI's disclosed commitment grew from $11.9 billion in March 2025 (19820), booked as $11.2 billion of backlog (4969), to up to $22.4 billion after a further up to $6.5 billion (20198). That arithmetic implies roughly $4 billion from the September 2025 order form, which is never separately quantified. Meta's $14.2 billion order form (4282) was followed by a roughly $21 billion order form signed March 31, 2026 (2123, 10340, 9317). The 'option to materially expand' (4283) was not itself a commitment. Jane Street's roughly $6 billion (10304) and Anthropic's unpriced multi-year agreement (9318) are also described as signed.

Softer language covers much of the remaining record. IBM (4971), CrowdStrike (11768), Core Scientific (20313) and the Genesis Mission (11778) are presented as partnerships or collaborations with no contract value. The January 2026 NVIDIA 'collaboration framework' (3892, 17596, 17598) is explicitly framework language and should not be confused with the $6.3 billion order form (20209). The 'partner of choice' logo lists (11767, 9320) and the 'expanded relationships' with Cohere and Mistral (9319) establish usage, not committed dollars. Two financing-linked contracts remain unidentified. One is the 'customer contract' behind DDTL 3.0 (4595). The other is the 'leading AI enterprise' behind DDTL 4.0 (11224).

Concentration has shifted rather than diluted. In March 2025, OpenAI alone was about 43% of the $25.9 billion backlog (4970, 4969). By December 2025, RPO had risen to $60.7 billion from $15.1 billion (3555), an increase of $45.6 billion. The named 2025 deals, OpenAI's $22.4 billion, Meta's $14.2 billion and NVIDIA's $6.3 billion, total about $42.9 billion. That is roughly 70% of year-end RPO. Many of these amounts are 'up to' figures, so recognized RPO may be lower. In 2026, Meta's cumulative roughly $35 billion overtook OpenAI as the largest disclosed counterparty. Jane Street's $6 billion only modestly broadens the base.

The revenue arithmetic broadly reconciles with the disclosed deals. $60.7 billion over a roughly five-year weighted duration (3561) implies about $12 billion of annual contracted revenue, assuming delivery. Two caveats remain. The gap between the $11.9 billion and $11.2 billion OpenAI figures is unexplained. Claim 10729 is dated March 1, 2026, yet it describes the Meta expansion signed March 31, which suggests a source-dating error.

CoreWeave's signed-contract record is unusually well documented, because most of its large deals appear as order forms filed under existing master services agreements. The OpenAI relationship began with an $11.9 billion agreement (claim 19820), which was booked as $11.2 billion of added backlog (claim 4969); the gap between those two figures has not been explained. A second OpenAI order form followed (claim 4296), and total OpenAI commitments reached up to about $22.4 billion (claim 20198). Meta signed an order form for up to $14.2 billion through 2031 (claims 4282, 17655), then a further roughly $21 billion order form on March 31, 2026 (claims 2123, 10340, 9317). NVIDIA's $6.3 billion capacity purchase was also executed as an order form (claims 4352, 20209). Jane Street committed about $6 billion (claim 10304). Anthropic signed a multi-year agreement with no disclosed value (claim 9318).

A second tier of announcements is softer. The January 2026 NVIDIA expansion is a collaboration framework (claim 3892), and its software commitments only 'work towards' inclusion in NVIDIA reference architectures (claim 17598). IBM (claim 4971), CrowdStrike (claim 11768), Core Scientific (claim 20313) and the Genesis Mission (claim 11778) are described as partnerships or initiatives, with no contract value attached. The logo lists (claims 11767, 9320) and the 'expanded relationships' with Cohere and Mistral (claim 9319) name counterparties but not terms. Going the other way, two debt facilities were raised against contracts whose counterparty is not named: 'a customer contract' (claim 4595) and a 'leading AI enterprise' (claim 11224).

Concentration has moved from OpenAI dominance in early 2025 to an OpenAI–Meta duopoly. By March 2026, Meta's cumulative named commitments of roughly $35 billion exceed OpenAI's $22.4 billion. Jane Street and Anthropic diversify the book only at the margin.

The 2025 arithmetic broadly reconciles. RPO rose from $15.1 billion to $60.7 billion (claims 3555, 3572), an increase of about $45.6 billion. The OpenAI, Meta and NVIDIA deals account for about $42.9 billion of that. The remaining gap is small enough to be covered by other customers once 2025 revenue burn-off is considered. The caveat is that the OpenAI and Meta figures are 'up to' amounts, and the Q1 2025 backlog of $25.9 billion was only partly RPO (claim 4970), so some announced value may sit outside RPO. The 2026 Meta, Jane Street and Anthropic deals cannot yet be tested, because no post-December 2025 RPO figure is on record here.

CoreWeave's signed-customer record is anchored by a small number of order forms executed under existing master services agreements. These are the strongest form of disclosure in the file. They cover OpenAI on September 23, 2025 (claim 4296), Meta in September 2025 (claim 4282) and again on March 31, 2026 (claim 2123), and NVIDIA as a capacity purchaser in September 2025 (claim 4352). The original OpenAI agreement is described at $11.9 billion in one source (claim 19820) and $11.2 billion of backlog in another (claim 4969). OpenAI's commitments grew to up to about $22.4 billion after a $6.5 billion expansion (claim 20198). Meta added roughly $21 billion on top of its initial $14.2 billion (claims 4282, 10340, 9317). Jane Street's roughly $6 billion (claim 10304) is a named, sized commitment. The Anthropic agreement is described as signed and multi-year but carries no value (claim 9318). Much of the rest is softer. The January 2026 NVIDIA arrangement is expressly a collaboration framework with test-and-validate language (claims 3892, 17598). IBM, CrowdStrike and the Genesis Mission are framed as partnerships or participation, not priced contracts (claims 4971, 10330, 11778). The logo lists (claims 11767, 9320) say 'partner of choice' without stating commitments. The financing disclosures refer only to unnamed counterparties: 'a customer contract' (claim 4595), a 'leading AI enterprise' (claim 11224), and 'two large, non-investment grade customers' (claim 9119). Those cannot be mapped to named deals from this record. On concentration, the book has moved from OpenAI-dominant in early 2025 toward a Meta–OpenAI duopoly. Meta's roughly $35 billion now edges out OpenAI's up to $22.4 billion, and Jane Street and Anthropic add a thinner second tier. Much of the headline value carries 'up to' qualifiers. On arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3555). The named 2025 deals, OpenAI, Meta and NVIDIA's $6.3 billion (claim 20209), total roughly $43 billion. Starting RPO plus these deals, less revenue recognized in the year, lands in the same range as $60.7 billion. That is a directional fit, not a precise tie-out. The fit is loosened by the Q1 2025 split of $25.9 billion backlog against $14.7 billion RPO (claim 4970), which suggests some OpenAI value initially sat outside RPO. The Meta $21 billion appears in at least five claims and should be counted once. Note that the input record is truncated after claim 9121.

CoreWeave's disclosed pipeline sorts into three tiers. The first is executed order forms under existing master services agreements, disclosed in 8-K language. These are the OpenAI contract announced in March 2025 at $11.9 billion (claim 19820), a further OpenAI order form in September 2025 (4296), Meta's order form committing up to roughly $14.2 billion through December 2031 (4282), and Meta's March 31, 2026 order form (2123), which the company values at about $21 billion through 2032 (10340, 9317). Jane Street's roughly $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) are also described as signed, though Anthropic's value is never quantified. The second tier is softer language. The NVIDIA 'collaboration framework' (3892, 17596), the IBM Granite 'partnership' (4971), the CrowdStrike 'global partnership' (11768), Genesis Mission participation (11778), Exemplar Cloud status (10332, 9324) and the rotating 'partner of choice' logo lists (11767, 9320) carry no disclosed dollar commitments and should not be read as backlog. The third tier is financing disclosures that reference customers without naming them: 'a customer contract' behind DDTL 3.0 (4595), a 'leading AI enterprise' behind DDTL 4.0 (11224), and 'two large, non-investment grade customers' behind DDTL 5.0 (9119). The credit description in the last of these is itself a material concentration flag.

Concentration was acute through 2025. OpenAI and Meta account for essentially all quantified named value, and Meta alone now totals roughly $35 billion across two order forms, with a further expansion option (4283). The 2026 additions of Jane Street, Anthropic, and the relationship expansions with Cohere and Mistral (9319) broaden the roster. However, the quantified dollars remain dominated by the same two hyperscale-scale buyers. The Hudson River Trading entry (10732) is truncated in the record and cannot be assessed.

The arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of about $45.6 billion. The quantified named 2025 deals total roughly $26.1 billion: OpenAI at $11.9 billion and Meta at $14.2 billion. That leaves close to $20 billion attributable to the unquantified September OpenAI and NVIDIA order forms (4296, 4352) and to unnamed counterparties. The OpenAI figure is also inconsistent across sources, at $11.9 billion in the press release against $11.2 billion of added backlog (4969). In addition, the March 2025 'revenue backlog' of $25.9 billion blended $14.7 billion of RPO with amounts contingent on delivery requirements (4970). Investors should anchor on RPO rather than backlog, and treat the unnamed DDTL counterparties as the key diligence gap.

CoreWeave's record is strongest where it cites order forms under existing master services agreements. Those are the OpenAI contract of March 2025 (claim 19820) and its September 2025 follow-on (4296), the Meta order forms of September 2025 (4282) and March 31, 2026 (2123, 10340), and NVIDIA's capacity-purchase order form (4352). Jane Street's roughly $6 billion commitment (10304) and the Anthropic agreement, described as 'signed' (9318), also read as executed, although the Anthropic deal carries no disclosed value. Much of the remaining partnership flow is not contract language. The IBM 'partnership' (4971) and the CrowdStrike 'global partnership' (11768, 10330) are not signed agreements. The NVIDIA 'collaboration framework' (3892, 17596) commits the parties only to 'work towards' reference-architecture inclusion (17598). Genesis Mission participation (11778) and Exemplar Cloud designations (10332, 9324) are affiliations and certifications, not revenue. The 'partner of choice' rosters (11767, 10338, 9320) and 'expanded relationships' with Cohere and Mistral (9319) name logos but attach no terms. The Hudson River Trading entry (10732) is truncated and cannot be assessed.

Several financing disclosures refer to counterparties that are never named. DDTL 3.0 funds 'a customer contract' (4595), DDTL 4.0 serves a 'leading AI enterprise' (11224), and DDTL 5.0 backs 'two large, non-investment grade customers' (9119). Because Meta is investment grade, the DDTL 5.0 counterparties are probably not Meta, but that is an inference and the company does not identify them.

Counterparty concentration has shifted from a single anchor to a pair of anchors. In early 2025 OpenAI alone accounted for most of the disclosed contract value. By early 2026 Meta's named commitments of about $14.2 billion (4282) plus about $21 billion (10340) exceeded the disclosed OpenAI figures. Readers should not double-count restatements of the same deals (17654, 17655, 10694, 9317). The $14.2 billion figure is also a ceiling ('up to'), not a floor.

The revenue arithmetic reconciles only partly. OpenAI's deal appears as $11.9 billion (19820) but as $11.2 billion of backlog (4969), and the gap is unexplained. Backlog of $25.9 billion (4970) includes amounts outside RPO, so the two measures are not interchangeable. RPO grew from $15.1 billion (3572) to $60.7 billion (3555) over 2025. The named additions (OpenAI's first tranche and Meta's $14.2 billion) explain only about $26 billion of that roughly $45.6 billion increase, before netting out revenue recognized during the year. The value of the September OpenAI order form is absent from the record, so roughly $20 billion of RPO growth cannot be traced to named counterparties. The March 2026 Meta expansion and the Jane Street deal come after the December 31, 2025 measurement date and are not part of that figure.

CoreWeave's record separates cleanly into two tiers. The first is contracts documented as executed order forms under existing master services agreements. These are OpenAI (claims 4969, 4296), NVIDIA's capacity purchase (4352), Meta's initial commitment of up to $14.2 billion through 2031 (4282) and Meta's March 2026 order form (2123), which the company later sized at about $21 billion (10340, 9317). Anthropic is described as a "signed multi-year agreement" (9318), but no value is disclosed. Jane Street's roughly $6 billion (10304) is framed as a commitment, and Hudson River Trading's arrangement appears only in secondary reporting (10732). The second tier is softer. It includes the NVIDIA "collaboration framework" (3892) and the reference-architecture work toward possible inclusion (17598), the IBM and CrowdStrike "partnerships" (4971, 11768), and Genesis Mission participation (11778). None of these carries a disclosed dollar commitment. The recurring "partner of choice" logo lists (11767, 10338, 9320) name counterparties but are marketing rosters, not contract disclosures. Meta's expansion option through 2032 (4283) was likewise contingent until the March order form.

Several material obligations remain unidentified. DDTL 3.0 funds "a customer contract" (4595). DDTL 4.0 serves a "leading AI enterprise" (11224). DDTL 5.0 supports "two large, non-investment grade customers" (9119). The Indonesia buildout cites unnamed "existing customer demand" (8762). Because debt facilities are secured against these contracts, the anonymity matters more here than in the logo lists.

Concentration has shifted rather than diluted. In May 2025, OpenAI's $11.2 billion equaled about 43% of the $25.9 billion backlog (4969, 4970). Meta has since become the largest named counterparty, at about $35 billion if the $21 billion is additive, as "new $21 billion commitment" (9317) implies. The March 2026 deals are also partly double-reported across claims 10729, 10694, 17654 and 10340, so they should not be summed twice.

The arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3571). Named 2025 additions with disclosed values total about $25.4 billion (OpenAI $11.2 billion plus Meta $14.2 billion), and the September OpenAI order form's value is not on record here. That leaves roughly $20 billion or more of net RPO growth, before revenue burn-off, that cannot be tied to a named, sized counterparty. The 2026 Meta and Jane Street figures postdate the year-end RPO and should appear in later balances. Backlog (4970) and RPO also differ by definition, so the two series should not be compared directly. On a five-year weighted duration (3584), $60.7 billion implies roughly $12 billion of annual recognition, and that figure is heavily weighted to two or three hyperscale and lab counterparties.

CoreWeave's record of counterparty claims mixes a small set of signed, valued contracts with a larger set of partnership, framework and logo-list language. Only a few disclosures are clearly binding. These are order forms executed under existing Master Services Agreements with NVIDIA (claim 4352), OpenAI (claim 4296) and Meta (claim 2123). Alongside them sit the valued commitments from Meta at up to $14.2 billion through December 2031 (claim 4282) and about $21 billion through December 2032 (claims 10340, 9317), and from Jane Street at about $6 billion (claim 10304). The OpenAI deal added $11.2 billion of backlog (claim 4969), but the September 2025 order form's value is not in the quoted text. Anthropic (claim 9318) and Hudson River Trading (claim 10732) are described as signed multi-year agreements, but no values are disclosed. Other items should not be read as contracted revenue. These include IBM's partnership (claim 4971), the NVIDIA collaboration framework (claim 3892), the CrowdStrike global partnership (claim 11768), Genesis Mission participation (claim 11778), NVIDIA Exemplar Cloud status (claim 10332), and the logo lists naming Cursor, Perplexity and others (claims 11767, 9320). Meta's expansion option through 2032 (claim 4283) is also an option, not a commitment. Several financing disclosures reference counterparties that are never named: 'a customer contract' behind DDTL 3.0 (claim 4595), a 'leading AI enterprise' behind DDTL 4.0 (claim 11224), and 'two large, non-investment grade customers' behind DDTL 5.0 (claim 9119). These are debt-backed obligations whose credit exposure cannot be attributed from the record. Concentration has shifted rather than diluted. In May 2025, OpenAI's $11.2 billion was about 43 percent of $25.9 billion in backlog (claim 4970). By the end of 2025, RPO had reached $60.7 billion, up from $15.1 billion (claims 3555, 3572), with Meta's $14.2 billion now alongside OpenAI. Those two named, valued deals account for roughly $25 billion, or about 42 percent of year-end RPO, and the remainder is not itemized. On revenue arithmetic, the record has backlog and RPO but no recognized revenue figures, so reconciliation is only partial. A five-year weighted duration (claim 3584) implies roughly $12 billion a year of contracted revenue on a straight-line basis. The 2026 Meta and Jane Street deals add about $27 billion after the last disclosed RPO date, so an updated RPO figure is needed to confirm they have been booked. The May 2025 backlog also included non-RPO amounts that depend on delivery conditions (claim 4970), so headline deal values should not be assumed to equal contracted obligations.

CoreWeave's record separates cleanly into contracts documented under order forms and a much larger layer of partnership language. The signed tier consists of OpenAI, disclosed first as an $11.2 billion strategic deal (claim 4969) and then as a September 2025 order form under the existing MSA (claim 4296). Meta follows with a roughly $14.2 billion commitment through December 2031 (claim 4282) and a further order form dated March 31, 2026 (claim 2123) worth about $21 billion through 2032 (claims 10340, 9317). Jane Street committed about $6 billion (claim 10304), and NVIDIA signed a capacity order form (claim 4352). Anthropic is described as a signed multi-year agreement with no value attached (claim 9318). Hudson River Trading appears only through secondary reporting (claim 10732). By contrast, the IBM Granite partnership (claim 4971), the NVIDIA collaboration framework whose software work is phrased as 'work towards' (claims 3892, 17598), the CrowdStrike global partnership (claim 11768), Genesis Mission participation (claim 11778), and the 'partner of choice' logo lists (claims 11767, 9320) carry no disclosed commitment and should not be read as contracted revenue.

A meaningful share of the financing-linked demand is unnamed. The DDTL 3.0 facility funds 'a customer contract' (claim 4595), DDTL 4.0 serves a 'leading AI enterprise' (claim 11224), and DDTL 5.0 backs 'two large, non-investment grade customers' (claim 9119). The last of these is a credit-quality flag the company chose not to attach to names.

Concentration has shifted from OpenAI toward Meta rather than diversified. At March 31, 2025, backlog of $25.9 billion exceeded RPO of $14.7 billion by exactly $11.2 billion (claim 4970), the OpenAI figure, which suggests that deal sat outside RPO pending delivery conditions. By year-end 2025, RPO had reached $60.7 billion, up from $15.1 billion (claim 3555). Named quantified deals through that date (OpenAI $11.2 billion plus Meta $14.2 billion) explain roughly $25 billion. The remaining roughly $35 billion must sit largely in the unquantified September OpenAI order form, the NVIDIA order form, and unnamed contracts. The post-period Meta ($21 billion) and Jane Street ($6 billion) deals add about $27 billion. Whether Meta's $21 billion is incremental to or partly overlaps the $14.2 billion is not stated on the record (claims 17654, 17655). With a roughly five-year weighted duration (claim 3584), the arithmetic is plausible, but it does not reconcile line by line from disclosed deals.

CoreWeave's record separates cleanly into two groups. The first is binding order forms with named counterparties and stated dollar values. The second is a larger set of partnership, framework and logo-list language that carries no disclosed economics.

The signed tier is narrow. It covers OpenAI, which added $11.2 billion of backlog (4969) and signed a further order form under its MSA in September 2025 (4296). It covers Meta, first at up to $14.2 billion through December 2031 (4282), then a separate order form on March 31, 2026 worth about $21 billion through December 2032 (2123, 10340, 9317). Claims 17654 and 17655 describe these as distinct commitments, so they should not be netted. Jane Street's roughly $6 billion commitment also belongs here (10304). NVIDIA's September 2025 order form (4352) is a signed capacity purchase, but the record gives no value. Anthropic is described as a "signed multi-year agreement" (9318) and Hudson River Trading as a multi-year agreement (10732, secondary source). Neither is quantified.

Everything else is softer. The IBM "partnership" (4971), the NVIDIA "collaboration framework" (3892, 17596, 17598), the CrowdStrike "global partnership" (11768) and Genesis Mission participation (11778) describe collaboration, not purchase commitments. The recurring "partner of choice" rosters (11767, 10338, 9320) name Cursor, Midjourney, Perplexity and others without contract terms. The "expanded relationships" with Cohere and Mistral (9319) are equally unquantified. Notably, the debt-financed contracts are the least identified. DDTL 3.0 backs "a customer contract" (4595). DDTL 4.0 serves a "leading AI enterprise" (11224). DDTL 5.0 serves "two large, non-investment grade customers" (9119). None is named, and the last is flagged as sub-investment-grade.

Concentration has broadened in name count but not obviously in dollars. Meta alone now accounts for roughly $35 billion of disclosed commitments, and OpenAI for at least $11.2 billion. The 2026 additions of Jane Street, Anthropic and HRT are either modest or undisclosed.

The arithmetic does not fully reconcile from the record. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), a $45.6 billion increase before netting revenue recognized. The quantified 2025 deals, OpenAI's $11.2 billion and Meta's up-to $14.2 billion, total about $25.4 billion. The residual of roughly $20 billion or more must sit in the unpriced OpenAI and NVIDIA order forms or in unnamed contracts. The $21 billion Meta and $6 billion Jane Street deals postdate December 31, 2025 and are not in that RPO figure. The approximately five-year weighted duration (3584) implies about $12 billion a year of contracted revenue, heavily weighted to two counterparties.

CoreWeave's record has a small core of documented contracts and a much larger layer of partnership and logo language. The firm contracts are order forms under existing master services agreements. These include OpenAI (4296, with $11.2 billion of backlog in 4969), Meta's initial commitment of up to $14.2 billion through December 2031 (4282), Meta's March 31, 2026 order form of about $21 billion (2123, 10340, 9317), and an NVIDIA capacity order (4352). Jane Street's roughly $6 billion commitment (10304) and the signed but unpriced Anthropic agreement (9318) also sit near this tier. Hudson River Trading's multi-year deal (10732) comes only from secondary reporting. Meta's 'option to materially expand' (4283) is not a commitment, and 'up to' in 4282 sets a ceiling. The six Meta $21 billion claims (10729, 10694, 17654, 10340, 9317, 2123) describe a single transaction.

The softer tier includes the IBM 'partnership' (4971), the NVIDIA 'collaboration framework' that commits only to 'work towards' reference-architecture inclusion (3892, 17596, 17598), the CrowdStrike 'global partnership' (11768, 10330), Genesis Mission participation (11778, 10331), Exemplar Cloud designations (10332, 9324), and 'expanded relationships' with Cohere and Mistral (9319). None of these carries a value. Logo rosters (11767, 10338, 9320) show breadth but no dollars. The financing disclosures stay anonymous. They refer to 'a customer contract' (4595), a 'leading AI enterprise' (11224), and 'two large, non-investment grade customers' (9119).

Concentration has moved rather than eased. At March 31, 2025, backlog was $25.9 billion against RPO of $14.7 billion (4970). The $11.2 billion gap equals the OpenAI figure exactly, which suggests that deal sat outside RPO pending delivery conditions. OpenAI was then about 43% of backlog. By December 31, 2025, RPO had reached $60.7 billion, up from $15.1 billion (3555, 3572), with a five-year weighted duration (3584). Named, quantified 2025 deals (OpenAI $11.2 billion and Meta $14.2 billion) explain about $25 billion of the $45.6 billion increase. The rest falls to the unpriced September OpenAI order form, the NVIDIA order, and unnamed contracts. Meta's cumulative named commitments of about $35 billion now make it the largest disclosed counterparty. The unnamed non-investment-grade customers funded under DDTL 5.0 add credit concentration that the logo lists do not show.

The arithmetic reconciles only in part. It holds at the March 2025 level. For year-end 2025, roughly $20 billion of RPO growth has no named, priced counterparty, and none of the 2026 figures can yet be tied to a reported RPO balance.

CoreWeave's record contains a small group of executed agreements. Around them sits a larger body of partnership and logo language that should not be read as contracted revenue. The firmest items are order forms under existing master services agreements. These cover OpenAI (4296), NVIDIA as a capacity purchaser (4352), and Meta: the September 2025 order form of up to about $14.2 billion through December 2031 (4282), and the March 31, 2026 order form of about $21 billion through December 2032 (2123, 10340, 9317). Jane Street's roughly $6 billion commitment (10304) is also quantified. Anthropic is described as a signed multi-year agreement (9318) but carries no value, and Hudson River Trading's multi-year agreement (10732) is likewise unpriced. The IBM partnership (4971), the NVIDIA collaboration framework (3892, 17596, 17598), the CrowdStrike global partnership (11768, 10330), Genesis Mission participation (11778, 10331) and the Exemplar Cloud designations (10332, 9324) are relationship or program language, not purchase commitments. The same caution applies to the rotating 'partner of choice' logo lists (11767, 9320), which name customers but attach no contract terms.

Several financing disclosures refer to counterparties that are never named. These include the customer contract behind DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), and the 'two large, non-investment grade customers' behind DDTL 5.0 (9119). Investors cannot tell whether these map onto the named deals or represent separate exposure. The same is true of the broad 'most of the world's leading AI labs' framing (8926).

Named concentration has broadened in logos but deepened in dollars. In 2025 the book was essentially OpenAI and Meta. Through 2026 Meta's cumulative quantified commitments reached roughly $35 billion (4282, 9317), while Jane Street added a mid-sized named commitment (10304).

The arithmetic reconciles only partially. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of about $45.6 billion. The quantified named deals in that window are OpenAI's $11.2 billion (4969), which was stated as backlog rather than RPO, and Meta's $14.2 billion (4282). Together they total about $25.4 billion. That leaves roughly $20 billion unattributed on this record, and the value of the September 2025 OpenAI order form (4296) is not captured here. The 2026 Meta and Jane Street commitments, together about $27 billion, postdate the $60.7 billion figure, and no later RPO disclosure is on record to confirm conversion. Claim 10729 is dated before the March 31 order form it describes and should be treated as a dating artifact.

CoreWeave's record contains a small core of genuinely signed counterparties and a much larger layer of partnership and logo language. The contractual core is clear. OpenAI signed order forms under an existing MSA, with $11.2 billion added to backlog in 2025 (4969) and a further order form in September 2025 whose value is not in the record (4296). Meta signed an order form for up to about $14.2 billion through December 2031 (4282) and a second order form on March 31, 2026 for about $21 billion through December 2032 (2123, 10340, 9317). NVIDIA signed a capacity order form under its MSA (4352). Anthropic's multi-year agreement is described as signed but carries no disclosed value (9318). Jane Street's roughly $6 billion commitment (10304) and Hudson River Trading's multi-year deal (10732) rest mainly on press paraphrase. Everything else is softer. The IBM partnership (4971), the NVIDIA collaboration framework and its 'work towards' reference-architecture language (3892, 17596, 17598), the CrowdStrike 'global partnership' (11768), the Genesis Mission (11778) and the Exemplar Cloud designations (10332, 9324) do not disclose contract terms. The 'partner of choice' logo lists (11767, 9320) and 'expanded relationships' with Cohere and Mistral (9319) name companies but not commitments. The financing disclosures consistently decline to name their counterparties. They refer to 'a customer contract' (4595), a 'leading AI enterprise' (11224), 'two large, non-investment grade customers' (9119) and 'most of the world's leading AI labs' (8926). Concentration has shifted from one dominant name to two. OpenAI's $11.2 billion was roughly 43% of the $25.9 billion backlog at March 2025 (4970). Meta's stated commitments now total about $35.2 billion across two order forms, which likely makes Meta the largest counterparty. The DDTL 5.0 disclosure adds non-investment-grade credit exposure without naming the customers. The revenue arithmetic reconciles only partially. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), an increase of $45.6 billion. Named 2025 dollar figures for OpenAI and Meta sum to $25.4 billion, so about $20 billion of the increase is attributable only to undisclosed amounts, most plausibly the September OpenAI order form and the NVIDIA order form. The comparison is also imperfect for two reasons. Revenue backlog and RPO are defined differently (4970), and Meta's figure is an 'up to' ceiling. The 2026 Meta $21 billion and Jane Street $6 billion deals postdate the December RPO figure and should appear in later disclosures. Several secondary claims carry inconsistent dates, such as a March expansion dated March 1 (10729), and warrant verification against the filings.

CoreWeave's disclosed pipeline splits into two tiers. The first is executed order forms under master services agreements with named counterparties. The second is partnership, framework and logo-list language that carries no disclosed economics. Executed paper covers OpenAI, with an $11.2 billion backlog addition (4969) and a September 2025 order form (4296) whose value is not captured here. It covers Meta, at up to $14.2 billion through 2031 (4282) and a further roughly $21 billion order form dated March 31, 2026 (2123, 10340, 9317). Jane Street has committed about $6 billion (10304), and NVIDIA placed a capacity order form (4352). Anthropic (9318) and Hudson River Trading (10732) have signed multi-year agreements with no disclosed value. The IBM Granite arrangement (4971), the NVIDIA 'collaboration framework' (3892, 17596), the CrowdStrike 'global partnership' (11768, 10330) and Genesis Mission participation (11778) read as cooperative rather than contracted revenue. The NVIDIA software item says only 'work towards' inclusion (17598). The 'partner of choice' rosters (11767, 9320) name logos without terms. MUFG, Morgan Stanley, Goldman Sachs and JPMorgan (11225, 9121) are lenders, not customers.

Several material counterparties go unnamed. These include the customer behind DDTL 3.0 (4595), the 'leading AI enterprise' financed through DDTL 4.0 (11224), and the 'two large, non-investment grade customers' behind DDTL 5.0 (9119). Because these facilities finance dedicated infrastructure, the unnamed counterparties likely carry significant exposure.

Concentration has shifted from OpenAI to Meta rather than diversified. In May 2025 the OpenAI deal alone equaled roughly 43% of $25.9 billion in backlog (4969, 4970). Meta's cumulative named commitment now stands near $35 billion, the largest single exposure on record. The April 2026 $14.2 billion reference (17655) restates the 2025 deal and is not additive. Repeated citations of the $21 billion (10729, 10694, 17654) and Jane Street (10730, 10695) figures are also duplicates.

On arithmetic, note that the May 2025 backlog figure bundles $14.7 billion of RPO with conditional amounts (4970), so it should not be compared directly with later RPO. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3571). Named, valued deals signed within that year reach at least $25.4 billion, leaving roughly $35 billion unattributed or tied to undisclosed order values. The Meta $21 billion and Jane Street $6 billion commitments postdate December 31, 2025 and should appear only in 2026 RPO. With a roughly five-year weighted duration (3584), $60.7 billion implies about $12 billion of annual contracted run-rate, subject to delivery conditions.

CoreWeave's record separates cleanly into executed order forms and softer relationship language. The firm contracts sit under master services agreements. They include an OpenAI order form (claim 4296) on top of the earlier $11.2 billion OpenAI deal (claim 4969), an NVIDIA capacity order form (claim 4352), and Meta's order form of up to $14.2 billion through December 2031 (claims 4282, 17655). Meta then signed a further order form of about $21 billion on March 31, 2026 (claims 2123, 10340, 9317). Six separate claims describe that one Meta deal, and three describe the single Jane Street commitment of about $6 billion (claims 10304, 10730, 10695), so the headline count overstates the number of distinct deals. Anthropic (claim 9318) and Hudson River Trading (claim 10732) are described as multi-year agreements, but neither has a disclosed value. The rest is weaker language. The IBM partnership (claim 4971), the NVIDIA collaboration framework that only commits to 'work towards' reference-architecture inclusion (claims 3892, 17598), the CrowdStrike 'global partnership' (claims 11768, 10330), Genesis Mission participation (claim 11778), Exemplar Cloud status (claims 10332, 9324), and the 'partner of choice' logo lists (claims 11767, 9320) carry no disclosed consideration. The same is true of the 'expanded relationships' with Cohere and Mistral (claim 9319).

Several material counterparties remain unnamed. The DDTL 3.0 facility funds 'a customer contract' (claim 4595), and DDTL 4.0 serves a 'leading AI enterprise' (claim 11224). DDTL 5.0 backs contracts with 'two large, non-investment grade customers' (claim 9119). These are the contracts that support debt, and a reader cannot tie them to names.

Customer concentration has shifted but not dispersed. In May 2025, OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog (claims 4969, 4970). By the end of 2025, Meta's $14.2 billion was roughly 23% of $60.7 billion in RPO (claims 3555, 4282). Meta's commitments now total about $35 billion. Jane Street and Hudson River Trading add a trading-firm segment, but it is small by comparison.

The revenue figures reconcile only partly. At March 31, 2025, backlog minus RPO was exactly $11.2 billion ($25.9 billion less $14.7 billion, claim 4970). That match suggests the OpenAI deal was first counted in backlog that depended on delivery conditions, not in RPO. RPO rose $45.6 billion during 2025 (claims 3571, 3572). The quantified named deals, OpenAI's $11.2 billion and Meta's 'up to' $14.2 billion, explain only about $25 billion of that increase. The remaining roughly $20 billion must come from the unpriced September OpenAI and NVIDIA order forms and the unnamed DDTL-financed contracts. The March and April 2026 Meta and Jane Street deals fall after that RPO date and should not be read into it.

CoreWeave's disclosure record separates fairly cleanly into two tiers. The first consists of signed order forms under existing master services agreements, filed as material contracts. These are OpenAI (claims 4969, 4296), NVIDIA's capacity purchase (4352), Meta's initial commitment of up to $14.2 billion through December 2031 (4282) and its roughly $21 billion March 2026 expansion through 2032 (2123, 10340, 9317). Jane Street's approximately $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) are described as signed or committed, but the record carries no filed contract terms for either, and no value for Anthropic. Hudson River Trading's agreement (10732) comes only from secondary reporting. The second tier is softer language: the IBM Granite 'partnership' (4971), the NVIDIA 'collaboration framework' and the undertaking to 'work towards' reference-architecture inclusion (3892, 17598), the CrowdStrike 'global partnership' (11768), Genesis Mission participation (11778), Exemplar Cloud designations (10332, 9324), and the rotating 'partner of choice' logo lists (11767, 9320). None of these carries a disclosed dollar commitment, and they should not be read as contracted revenue.

Several of the largest financing-linked obligations are unnamed. DDTL 3.0 finances 'a customer contract' (4595), the March 2026 facility serves a 'leading AI enterprise' (11224), and DDTL 5.0 supports contracts with 'two large, non-investment grade customers' (9119). The company also claims to serve 'most of the world's leading AI labs and hyperscalers' (8926) without identifying them. Readers cannot tell whether these overlap with OpenAI or Meta or represent new counterparties. That matters most for the non-investment-grade disclosure.

Concentration has narrowed rather than diversified at the top. Meta's named commitments alone total about $35.2 billion, and OpenAI's disclosed May tranche adds $11.2 billion. Newer names such as Jane Street, Anthropic and the trading firms broaden the list but not yet the dollar weighting.

The arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of about $45.6 billion. The quantified 2025 named deals, OpenAI's $11.2 billion and Meta's $14.2 billion, explain roughly $25 billion of that, before netting revenue recognized during the year. The remaining $20 billion or more must come from the unquantified September OpenAI and NVIDIA order forms, or from unnamed customers. The May 2025 $25.9 billion 'backlog' figure (4970) also mixed $14.7 billion of RPO with conditional amounts. Headline backlog and RPO should therefore not be treated interchangeably when matching announcements to the balance sheet.

CoreWeave's contract record has two tiers. The top tier is order forms under existing Master Services Agreements, with named counterparties and, in most cases, dollar values. These are OpenAI's $11.2 billion strategic deal (4969) and its September 2025 order form (4296, value not on record); Meta's initial commitment of up to $14.2 billion through December 2031 (4282, 17655), followed by a March 31, 2026 order form (2123) worth about $21 billion through 2032 (10340, 9317); and Jane Street's roughly $6 billion (10304). Anthropic (9318) and Hudson River Trading (10732) are described as signed multi-year agreements, but neither has a disclosed value. The second tier is softer language that should not be read as revenue: the IBM partnership (4971), the NVIDIA "collaboration framework" and reference-architecture work that the company says it will "work towards" (3892, 17596, 17598), the CrowdStrike "global partnership" (11768, 10330), Genesis Mission participation (11778), Exemplar Cloud designations (10332, 9324), and the "partner of choice" logo lists (11767, 9320). None of these carries a disclosed dollar value.

A meaningful share of the financing-linked demand is also anonymous. DDTL 3.0 funds "a customer contract" (4595), DDTL 4.0 serves an unnamed "leading AI enterprise" (11224), and DDTL 5.0 is dedicated to "two large, non-investment grade customers" (9119). The disclosures do not say which named counterparties, if any, these are.

Concentration has broadened from one counterparty to two. In May 2025, OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog (4970). By the end of 2025, OpenAI's and Meta's disclosed values together came to about $25.4 billion, roughly 42% of the $60.7 billion in remaining performance obligations (RPO) (3555). If Meta's $21 billion is incremental, as "new commitment" in 9317 implies, Meta alone now exceeds $35 billion. Jane Street is a distant third.

The arithmetic reconciles only partly. RPO fell from $15.1 billion (3572) to $14.7 billion in Q1 2025 even as backlog reached $25.9 billion. That suggests the OpenAI deal was booked largely outside RPO, conditional on delivery (4970). The $45.6 billion increase in RPO during 2025 exceeds the $25.4 billion in disclosed named deals, which leaves about $20 billion attributable only to OpenAI's unpriced September order form and to unnamed customers. The Meta $14.2 billion is also an "up to" ceiling rather than a floor. Several press-derived claims carry month-start dates that predate their filings (10729, 10731), so they should be dated from the 8-K, not the article.

CoreWeave's record separates cleanly into executed order forms and softer relationship language, and the executed paper is concentrated in very few names. The signed agreements are order forms under existing Master Services Agreements with NVIDIA (claim 4352), OpenAI (claim 4296) and Meta (claims 4282 and 2123). Meta's initial commitment of up to about $14.2 billion through December 2031 (claim 4282) is quantified; the March 2026 order form is reported at roughly $21 billion through 2032 (claims 10340, 9317). Jane Street's approximately $6 billion commitment (claim 10304) is also described as committed. The Anthropic and Hudson River Trading agreements are called multi-year and signed (claims 9318, 10732), but no value is disclosed for either.

The rest of the record should not be read as contracted revenue. The NVIDIA expansion is explicitly a collaboration framework (claims 3892, 17596), and its software element is framed as testing and working toward inclusion (claim 17598). IBM (claim 4971) and CrowdStrike (claims 11768, 10330) are partnership announcements without disclosed terms. The Genesis Mission (claim 11778) is programme participation. The logo lists naming Cognition, Cursor, Mercado Libre, Perplexity and others (claims 11767, 9320) indicate customer relationships of unknown size. Several financing disclosures describe counterparties without naming them: 'a customer contract' (claim 4595), a 'leading AI enterprise' (claim 11224), and 'two large, non-investment grade customers' (claim 9119). These are material enough to secure dedicated debt facilities but cannot be attributed from the record.

Concentration has shifted from OpenAI toward Meta. In May 2025 the $11.2 billion OpenAI deal (claim 4969) was about 43% of the $25.9 billion backlog (claim 4970). Meta's two disclosed commitments now total about $35 billion. The $21 billion figure recurs across many claims (10729, 10694, 17654, 9317, 10340), but these are repeated reports of one transaction, not separate deals.

The arithmetic reconciles only partially. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3571). Meta's $14.2 billion and OpenAI's $11.2 billion together explain roughly $25 billion of that $45.6 billion increase. That leaves about $20 billion attributable to the unpriced September OpenAI and NVIDIA order forms and to unnamed customers. Part of the OpenAI figure may also already have been in the March 2025 backlog, which mixed RPO with conditional amounts (claim 4970). The Meta $21 billion and Jane Street deals were signed in 2026, after the December 2025 RPO date, so they should not be read into that $60.7 billion figure.

CoreWeave's record separates cleanly into a small set of order forms and a much larger layer of partnership language. The binding instruments are the OpenAI order forms (4969, 4296), the Meta order forms of September 2025 and March 31, 2026 (4282, 2123), and an NVIDIA order form under a 2023 MSA (4352). The Anthropic agreement is described as signed but carries no disclosed value (9318). The Jane Street commitment of about $6 billion (10304) and the Hudson River Trading deal (10732) rest mainly on press and secondary sourcing. The NVIDIA 'collaboration framework' (3892, 17596) is aspirational, with reference-architecture inclusion framed as work 'towards' (17598). The IBM (4971), CrowdStrike (11768) and Genesis Mission (11778) items are partnerships rather than disclosed revenue contracts. The 'partner of choice' logo lists (11767, 9320) and the Exemplar Cloud designations (10332, 9324) are marketing and certification claims, not commitments. Several counterparties are left unnamed, most notably the customer behind DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), and the 'two large, non-investment grade customers' financed by DDTL 5.0 (9119). Those undisclosed borrowers are where credit risk sits. Concentration began extreme: OpenAI's $11.2 billion was roughly 43% of the $25.9 billion backlog in Q1 2025 (4969, 4970). Meta's $14.2 billion and $21 billion order forms then made the book a two-anchor structure. The 2026 additions of Anthropic, Jane Street and Hudson River Trading broaden the base at the margin, but they are small beside those two anchors. On arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of $45.6 billion. Named, valued 2025 deals explain only about $25.4 billion of that increase, from OpenAI's $11.2 billion and Meta's $14.2 billion. Even that figure is generous, because backlog is not RPO and Meta's figure is an 'up to' ceiling (4970, 4282). Roughly $20 billion therefore traces to the September OpenAI and NVIDIA order forms, whose values are not on this record. The March 2026 Meta deal and the April 2026 Jane Street deal postdate the RPO measurement and should not be read into it. Readers should also avoid double-counting the $21 billion, which is restated across five claims (10340, 10694, 10729, 17654, 9317). The source list itself was truncated, so later disclosures may be missing.

CoreWeave's signed, named commitments come mainly from order forms filed under existing master services agreements. OpenAI has two: the $11.2 billion strategic deal (claim 4969) and a September 2025 order form (claim 4296). Meta has the $14.2 billion commitment running through December 2031 (claim 4282) and a March 31, 2026 order form of about $21 billion (claims 2123, 10340, 9317). NVIDIA's capacity order form (claim 4352) is a third. Jane Street's roughly $6 billion commitment (claim 10304) is named and quantified, though the May 2026 release calls it an 'expanded relationship' (claim 9319). Anthropic is described as signed but carries no dollar figure (claim 9318). Hudson River Trading's multi-year agreement rests on secondary reporting (claim 10732). Much of the remaining partnership language is softer. The NVIDIA expansion is a 'collaboration framework' (claims 3892, 17596, 17598). IBM (claim 4971), CrowdStrike (claim 11768) and the Genesis Mission (claim 11778) are partnerships or initiatives, not disclosed contracts. The 'partner of choice' logo lists (claims 11767, 9320, 8926) name counterparties but say nothing about contract terms. The most material financing disclosures leave the customer unnamed. DDTL 3.0 funds 'a customer contract' (claim 4595). DDTL 4.0 serves a 'leading AI enterprise' (claim 11224). DDTL 5.0 backs 'two large, non-investment grade customers' (claim 9119). The facilities in claim 8760 come from an unnamed colocation provider. Concentration has moved from OpenAI toward Meta. In March 2025, total backlog of $25.9 billion was $14.7 billion of RPO plus exactly $11.2 billion of conditional amounts (claim 4970), and that figure matches the OpenAI deal. This suggests OpenAI then made up about 43% of backlog and sat largely outside RPO. By 2026, Meta's two order forms total about $35 billion if the $21 billion is additive, which the separate order form implies. Claim 17654 frames that figure ambiguously, so treat additivity as likely rather than confirmed. On the arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3555). Named, quantified 2025 deals explain only part of the $45.6 billion increase: OpenAI's $11.2 billion plus Meta's $14.2 billion. The September OpenAI order form and the DDTL-backed contracts are not quantified in this record. With a weighted-average duration of about five years (claim 3556, repeated in claims 3561, 3573 and 3584), $60.7 billion implies roughly $12 billion of annual recognition. That depends on delivery by counterparties that include non-investment-grade names. The record appears truncated after claim 8760.

The signed-contract set at CoreWeave is narrower than the announcement flow suggests, and it is worth separating the two. Genuine executed agreements with named counterparties are OpenAI (claim 4969, formalized as an order form under a pre-existing MSA in claim 4296), Meta (claims 4282 and 2123, both order forms under the December 2023 MSA), NVIDIA's own capacity purchase (claim 4352, an order form under the April 2023 MSA), Jane Street (claim 10304), and Anthropic (claim 9318, described as signed). Hudson River Trading (claim 10732) is announced but the commitment is undisclosed. Against that, IBM (claim 4971), CrowdStrike (claims 11768 and 10330), and the January 2026 NVIDIA arrangement (claims 3892, 17596, 17598) are partnership and collaboration-framework language, not order forms; the 5-gigawatt buildout and the reference-architecture validation carry no counterparty payment obligation. NVIDIA Exemplar Cloud (claims 10332, 9324) and Genesis Mission participation (claims 11778, 10331) are designations, not revenue. The roster claims (11767, 10338, 9320, 9319) name Cognition, Cursor, Mercado Libre, Midjourney, Runway, Cohere, Mistral, Perplexity and others without a single disclosed dollar value or contract term.

The unnamed counterparties matter more than the named ones, because they appear precisely where the financing sits. The DDTL 3.0 facility was raised against an unidentified customer contract (claim 4595); DDTL 4.0 against a leading AI enterprise (claim 11224); DDTL 5.0 against two large, non-investment-grade customers (claim 9119). Debt capacity is being underwritten against credits the company will not name.

Concentration has shifted rather than diffused. OpenAI's $11.2 billion was roughly forty-three percent of the $25.9 billion March 2025 backlog (claims 4969, 4970), and notably sat outside the $14.7 billion RPO figure, in the conditional other amounts bucket. Meta then became the second pole, at $14.2 billion (claim 4282) and later approximately $21 billion (claims 10340, 9317).

The arithmetic does not close cleanly. RPO went from $15.1 billion to $60.7 billion across 2025 (claims 3572, 3571), a $45.6 billion increase, against roughly $25.4 billion of named 2025 commitments; some twenty billion is unattributed to any disclosed counterparty. Separately, claims 17654 and 17655 place the $14.2 billion and $21 billion Meta figures side by side without establishing whether the second is incremental or inclusive.

The signed-contract set is materially smaller than the announcement set. Only a handful of counterparties appear in documents carrying the structural markers of an executed agreement — a named party, an order form under a pre-existing master services agreement, and a dollar figure: OpenAI (claim 4296), Meta (claims 4282, 2123), NVIDIA (claim 4352) and Jane Street (claim 10304). Everything else sits lower on the commitment ladder. The January 2026 NVIDIA items are framework language rather than purchase language; claim 3892 describes a collaboration framework, and claim 17598 commits only to test and validate CoreWeave software and work towards including it in reference architectures. CrowdStrike is announced as a global partnership (claims 11768, 10330) while simultaneously appearing in the customer roster (claims 11767, 10338), with no disclosure resolving which. Anthropic and Hudson River Trading are characterized as multi-year agreements with no value, term, or filing behind them (claims 9318, 10732). The logo lists — Cognition, Cursor, Mercado Libre, Midjourney, Runway, Perplexity, World Labs, Adaption Labs, Advaita Bio (claims 11767, 10338, 9320) — are named but unquantified, which is the weakest disclosure form present.

The more telling pattern is the unnamed. The company financed capital expenditure against a customer contract it declined to identify (claim 4595), then a leading AI enterprise (claim 11224), then two large, non-investment grade customers (claim 9119). Three successive term loans are underwritten by counterparties investors cannot name, and the non-investment grade characterization implies the concentration has not broadened into credit-diverse hands.

The arithmetic partly reconciles and partly does not. The May 2025 disclosure is internally tidy: backlog of $25.9 billion less RPO of $14.7 billion leaves $11.2 billion (claim 4970), exactly the OpenAI strategic deal (claim 4969) — meaning that entire relationship sat outside RPO. The 2025 step-up is harder. RPO moved from $15.1 billion to $60.7 billion (claim 3555), roughly $45.6 billion added, against which the only quantified 2025 signing is Meta's $14.2 billion (claim 4282). The $21 billion Meta expansion (claims 10340, 9317) and Jane Street's $6 billion post-date the measurement, and claims 17654 and 17655 present $21 billion and $14.2 billion side by side without saying whether the former subsumes the latter. Roughly $30 billion of committed backlog therefore traces to counterparties whose contract values were never separately disclosed.

The record separates into three tiers, and the disclosure language tracks the tier closely. Genuine signed customer instruments — order forms executed under pre-existing master services agreements — exist for only three counterparties: NVIDIA (claim 4352), OpenAI (claim 4296), and Meta (claims 4282, 2123). Jane Street (claim 10304) and Anthropic (claim 9318) are described as committed and signed but without a filed instrument cited. Meta is the only counterparty whose dollar figure sits inside the agreement description itself, at roughly $14.2 billion through December 2031 (claim 4282) and about $21 billion through December 2032 (claim 10340).

The second tier is named but unsized. The logo lists of Cognition, Cursor, Mercado Libre, Midjourney and Runway (claims 11767, 10338), later Adaption Labs, Advaita Bio, Hudson River Trading, Perplexity and World Labs (claim 9320), arrive under partner-of-choice framing rather than as contracts. CrowdStrike appears as a global partnership (claims 11768, 10330), NVIDIA's January 2026 five-gigawatt announcement as a collaboration framework (claims 3892, 17596, 17598), and Genesis Mission (claim 11778) and Exemplar Cloud (claim 10332) as designations. None carries committed revenue. The third tier is the material one and it is anonymous: a customer contract (claim 4595), a leading AI enterprise (claim 11224), two large non-investment-grade customers (claim 9119), an unnamed colocation provider (claim 8760).

Concentration has broadened in name count more than in substance. OpenAI alone was $11.2 billion of the $25.9 billion May 2025 backlog (claims 4969, 4970); by 2026 the book is an OpenAI-plus-Meta core with a largely unsized tail (claims 10731, 10732), and claim 9119 implies the financing stack still rests on very few names.

The arithmetic strains. RPO moved from $15.1 billion to $60.7 billion across 2025 (claim 3555), a $45.6 billion increase, against which the only dollar-denominated named agreement signed in that window is Meta's $14.2 billion. The September OpenAI order form is the presumptive remainder but was never sized, leaving roughly $31 billion attributable to no quantified, named deal. The $21 billion Meta expansion (claim 10729) and $6 billion Jane Street commitment (claim 10730) postdate December 31, 2025 and cannot sit inside that figure; and the May 2025 backlog is a broader measure than the $14.7 billion RPO it contains, so the two should not be compared across periods.

CoreWeave's disclosure record separates cleanly into two tiers, and the gap between them is the story. The genuinely executed tier consists of order forms under pre-existing master services agreements, which is contract language, not intent language: OpenAI on September 23, 2025 (claim 4296), Meta's initial commitment of up to approximately $14.2 billion through December 14, 2031 (claim 4282), the March 31, 2026 Meta order form of roughly $21 billion through December 2032 (claims 2123, 10340), NVIDIA's capacity purchase under the April 2023 MSA (claim 4352), Jane Street at approximately $6 billion (claim 10304) and a multi-year Anthropic agreement (claim 9318). Everything else is softer. The January 2026 NVIDIA collaboration framework (claim 3892) and the 5-gigawatt buildout announcement (claim 17596) create no payment obligation, and claim 17598 concedes the software integration is only something the parties work toward. IBM (claim 4971), CrowdStrike (claims 11768, 10330), Genesis Mission participation (claims 11778, 10331) and the Exemplar Cloud designation (claims 10332, 9324) are positioning, not backlog.

On identification, the named-logo lists have grown while the dollar-bearing counterparties have stayed anonymous. Cognition, Cursor, Mercado Libre, Midjourney and Runway (claims 11767, 10338), then Cohere and Mistral (claim 9319), then Adaption Labs, Advaita Bio, Hudson River Trading, Perplexity and World Labs (claim 9320) arrive without a single contract value. Meanwhile the financings describe their collateral only generically: a customer contract behind DDTL 3.0 (claim 4595), a leading AI enterprise behind DDTL 4.0 (claim 11224), and, most tellingly, two large, non-investment grade customers behind DDTL 5.0 (claim 9119). An unnamed colocation provider appears in claim 8760, and claim 8926 asserts most of the world's leading AI labs without naming them.

The arithmetic only partly reconciles. RPO moved from $15.1 billion to $60.7 billion across 2025 (claims 3572, 3555), a build of roughly $45.6 billion, against named and quantified 2025 additions of $11.2 billion from OpenAI (claim 4969) and $14.2 billion from Meta. Over half the increase therefore attaches to deals whose value was never disclosed, including the September OpenAI order form and the NVIDIA commitment. Whether the $21 billion Meta figure supersedes or stacks on the $14.2 billion is also unresolved on the record (claims 17654, 17655), and at a five-year weighted-average duration (claim 3584) the implied run-rate depends heavily on that answer.

The contractual record separates cleanly along a disclosure-channel line: what arrived via 8-K order forms is signed, and what arrived via press release generally is not. Meta is the only counterparty with a fully documented paper trail — an initial commitment of up to approximately $14.2 billion through December 2031 under an existing MSA dated December 10, 2023 (claim 4282), an expansion option through 2032 (claim 4283), and a second order form executed March 31, 2026 for roughly $21 billion (claims 2123, 10340, 9317, 10729). OpenAI is named in a September 2025 order form (claim 4296) and earlier credited with $11.2 billion of backlog (claim 4969). NVIDIA appears twice in materially different registers: a September 2025 capacity purchase order form under the same MSA structure (claim 4352), which is a binding commitment from a related party, and a January 2026 collaboration framework (claims 3892, 17596) whose operative verb is to test, validate and work toward inclusion in reference architectures (claim 17598) — framework language, not revenue.

Against that, a large named cohort carries no disclosed economics. Anthropic (claims 9318, 10731) and Hudson River Trading (claim 10732) are described as multi-year agreements without value; Jane Street is the rare press-release name with a figure, about $6 billion (claims 10304, 10730). Cognition, Cursor, Mercado Libre, Midjourney, Runway (claims 11767, 10338), Cohere, Mistral (claim 9319), Adaption Labs, Advaita Bio, Perplexity and World Labs (claim 9320) are partner-of-choice listings. CrowdStrike (claims 11768, 10330), IBM (claim 4971), the Genesis Mission (claims 11778, 10331) and the NVIDIA Exemplar Cloud designation (claims 10332, 9324) are positioning, not contracts. Conversely, the largest contracts are precisely the ones left unnamed: the DDTL 3.0 customer contract (claim 4595), DDTL 4.0's leading AI enterprise (claim 11224), and DDTL 5.0's two large, non-investment-grade customers (claim 9119).

The arithmetic is directionally consistent but incompletely bridged. RPO rose from $15.1 billion to $60.7 billion (claims 3572, 3555), a $45.6 billion increase, against roughly $25.4 billion of named, quantified 2025 additions — leaving over 40% attributable to unidentified counterparties, with the unquantified OpenAI and NVIDIA order forms plausibly absorbing much of it. Concentration has broadened in names while the dollars remain lodged with Meta, OpenAI and Microsoft (claim 648).

Only a minority of CoreWeave's announced relationships are documented as executed customer contracts carrying disclosed economics. Four counterparties clear that bar: OpenAI, whose May 2025 strategic deal added $11.2 billion of backlog (claim 4969) and which executed a further order form under the 2023 MSA in September 2025 with no dollar value given in the record (claim 4296); Meta, with an initial commitment of up to approximately $14.2 billion through December 2031 (claim 4282) and a March 2026 order form of approximately $21 billion through December 2032 (claims 2123, 10340, 9317); Jane Street at approximately $6 billion (claims 10304, 10730); and NVIDIA, though its September 2025 order form (claim 4352) is a capacity purchase by NVIDIA, a different animal from third-party demand. Anthropic and Hudson River Trading are named with multi-year agreements but no disclosed value (claims 9318, 10732).

Against that, a substantial tier is named without contractual substance or described without being named at all. IBM (claim 4971), CrowdStrike (claims 11768, 10330) and Leidos (claim 8669) appear as partnerships and collaborations, not priced commitments, and CrowdStrike is simultaneously recited in a partner-of-choice logo list (claims 11767, 10338) that also carries Cognition, Cursor, Mercado Libre, Midjourney, Runway, and later Adaption Labs, Advaita Bio, Perplexity, World Labs, Cohere and Mistral (claims 9320, 9319) — none with attached economics. More material are the anonymized ones: a customer contract behind DDTL 3.0 (claim 4595), a leading AI enterprise behind DDTL 4.0 (claim 11224), and two large, non-investment-grade customers behind DDTL 5.0 (claim 9119). Microsoft, reported as among the biggest customers (claim 648), never appears as the named party to a disclosed order form here.

The arithmetic does not reconcile. Named, priced deals total roughly $52 billion even before asking whether Meta's $21 billion supersedes or supplements the earlier $14.2 billion — an unresolved ambiguity worth pressing. Backlog reached $104.2 billion in Q2 2026 (claim 1302), leaving about half unattributed. Separately, RPO of $60.7 billion at December 2025 (claim 3555) sits far below backlog, the difference being amounts contingent on delivery and availability conditions (claim 4970). Concentration has broadened from an OpenAI-and-Microsoft core toward Meta, trading firms and government-adjacent work, but the DDTL 5.0 language confirms credit concentration persists.

Only a handful of CoreWeave counterparties are backed by executed contract documents with disclosed economics. Meta is the clearest case: an order form under a December 2023 MSA committing up to approximately $14.2 billion through December 2031 (4282), with an option to expand through 2032 (4283), followed by a new order form dated March 31, 2026 (2123) characterized as roughly $21 billion through December 2032 (10340, 9317). Because the 2026 order form appears to be the exercise of that option, the two figures cannot safely be added, and nothing disclosed says whether the later number is incremental or cumulative (10729, 10694). OpenAI is named in an executed order form (4296) and credited with $11.2 billion of backlog (4969); NVIDIA has a signed order form whose capacity commitment is undisclosed (4352). Jane Street's approximately $6 billion (10304, 10730, 10695) and multi-year agreements with Anthropic (9318, 10731) and Hudson River Trading (10732) are described as signed, but only Jane Street carries a number.

Everything else sits in a softer register. The January 2026 NVIDIA item is a collaboration framework (3892), not an order, and the distinction matters precisely because the same counterparty also supplies real order-form volume. IBM (4971), CrowdStrike (11768, 10330) and Leidos (8669) are partnership or collaboration announcements without term, value, or contract reference, while Genesis Mission participation (11778, 10331) and NVIDIA Exemplar Cloud status (10332, 9324) are membership and designation rather than revenue. The partner-of-choice rosters naming Cognition, Cursor, Mercado Libre, Midjourney and Runway (11767, 10338), then Adaption Labs, Advaita Bio, Perplexity and World Labs (9320), plus expanded relationships with Cohere and Mistral (9319), identify counterparties but attach no economics.

Concentration has narrowed in names and widened in dollars. At March 2025, OpenAI alone was roughly 43% of the $25.9 billion backlog (4970). RPO moved from $15.1 billion (3572) to $60.7 billion (3571), and backlog from $99.4 billion to $104.2 billion with more than $25 billion signed early in the following quarter (1302, 1303). The arithmetic does not close: named, valued deals total roughly $27 billion against a far larger increase, and the biggest financings are tied to unidentified parties, namely a customer contract (4595), a leading AI enterprise (11224), and two large, non-investment-grade customers (9119). Backlog also exceeds RPO by construction (4970), so headline growth comparisons mix two measures.

CoreWeave's record separates into three tiers, and only the first carries contractual weight. Genuine signed instruments — order forms executed under pre-existing master services agreements — exist with OpenAI (claim 4296, MSA-based order form dated September 23, 2025), Meta (claim 2123, order form of March 31, 2026 under a December 2023 MSA), and NVIDIA (claim 4352). Anthropic is described as a signed multi-year agreement (claim 9318), and Jane Street as a roughly $6 billion commitment (claim 10304). Hudson River Trading is reported only as an announced multi-year agreement (claim 10732), sourced to press rather than filing.

The second tier is named but uncontracted. IBM (claim 4971) and CrowdStrike (claims 11768, 10330) appear as partnerships with no quantum, term, or committed spend. The partner-of-choice lists — Cognition, Cursor, Mercado Libre, Midjourney, Runway (claims 11767, 10338), then Adaption Labs, Advaita Bio, Perplexity, World Labs (claim 9320) — are logo inventories, not disclosed obligations; Cohere and Mistral appear as expanded relationships (claim 9319) without figures. NVIDIA's 2026 posture softens from order form to collaboration framework (claim 3892), and the Exemplar Cloud designation (claims 10332, 9324) and Genesis Mission participation (claims 11778, 10331) are credentials, not revenue.

The third tier is the one that matters most and names no one. The financings repeatedly describe counterparties by silhouette: a customer contract behind DDTL 3.0 (claim 4595), a leading AI enterprise behind DDTL 4.0 (claim 11224), and two large, non-investment-grade customers behind DDTL 5.0 (claim 9119). The arrangers are named precisely (claims 11225, 9121, 8736) while the obligors are not — non-investment-grade being the operative disclosure.

Concentration has broadened in logo count but not in dollars. OpenAI alone was $11.2 billion of a $25.9 billion backlog in March 2025 (claims 4969, 4970). RPO moved $15.1 billion to $60.7 billion (claims 3572, 3555); backlog reached $104.2 billion by Q2 2026 (claim 1302). Named commitments — OpenAI $11.2 billion, Meta $14.2 billion then $21 billion (claims 4282, 10340), Jane Street $6 billion — total roughly $52 billion, about half the backlog. Whether Meta's $21 billion supersedes or stacks on the $14.2 billion is never clarified, and the $4.8 billion sequential backlog increase (claim 1302) sits awkwardly against $6 billion of Jane Street and $25 billion of subsequent commitments (claims 1303, 3011). The arithmetic does not close.

The contractually hard set is smaller than the announcement flow suggests. Only six counterparties appear with executed instruments and dollar figures: OpenAI, via an order form under the April 2023 MSA (claim 4296) and the earlier $11.2 billion backlog addition (claim 4969); Meta, via the December 2023 MSA order forms totaling roughly $14.2 billion through 2031 (claim 4282) and a further ~$21 billion through 2032 signed March 31, 2026 (claims 2123, 10340, 9317); NVIDIA, under a September 2025 order form (claim 4352); Jane Street at approximately $6 billion (claim 10304); Anthropic, described as a signed multi-year agreement but with no disclosed value (claim 9318); and Hudson River Trading, multi-year and undisclosed (claim 10732). Everything else is softer. The IBM arrangement is a "partnership" (claim 4971), CrowdStrike a "global partnership" (claims 11768, 10330) even while appearing on the customer roster (claims 11767, 10338), and the January 2026 NVIDIA item is explicitly a "collaboration framework" (claim 3892), which should not be double-counted against the September order form. NVIDIA Exemplar Cloud (claims 10332, 9324) and the DOE Genesis Mission (claims 11778, 10331) are designations, not revenue. The lists of Cognition, Cursor, Mercado Libre, Midjourney, Runway, Perplexity, World Labs, Adaption Labs, Advaita Bio, Cohere and Mistral (claims 11767, 10338, 9320, 9319) carry no disclosed economics at all.

The named-versus-unnamed split is most visible in the financings. DDTL 3.0 cites "a customer contract" (claim 4595), DDTL 4.0 a "leading AI enterprise" (claim 11224), DDTL 5.0 "two large, non-investment grade customers" (claim 9119), and the Indonesia build an unnamed colocation lessor and unnamed "existing customer demand" (claims 8760, 8762). Concentration has broadened nominally but not in dollars: OpenAI alone was $11.2 billion of a $25.9 billion backlog in May 2025 (claims 4969, 4970), and Microsoft and OpenAI were still described as the biggest a year later (claim 648).

The arithmetic only partially reconciles. RPO rose $45.6 billion in 2025, from $15.1 billion to $60.7 billion (claim 3555), yet the only 2025 deal with a disclosed value is Meta's $14.2 billion; the OpenAI and NVIDIA order forms were unpriced. Similarly, Q2 2026 backlog rose just $4.8 billion to $104.2 billion (claim 1302) despite the $21 billion Meta expansion, which landed in Q1, while the $25 billion of "new commitments" (claims 1303, 3011) sits outside both. The backlog-to-RPO gap remains the conditional bucket flagged in claim 4970. Note also that several figures repeat across claims (3555/3560/3571; 10729/10694; 10730/10695), and the "$21,000" unit artifacts in claims 10729 and 9317 should not be read as corroborating distinct events.

CoreWeave's record has a small core of signed, order-form-level contracts, surrounded by a much larger volume of partnership and logo-list language. The signed contracts are OpenAI, through the $11.2 billion strategic deal (4969) and a September 2025 order form under an existing MSA (4296). Meta signed a $14.2 billion order form running through 2031 (4282). It then signed a second order form on March 31, 2026 (2123), valued at roughly $21 billion through 2032 (10340, 9317). NVIDIA's September 2025 order form (4352) also sits on the customer side of the ledger. Jane Street's roughly $6 billion commitment (10304) and the Anthropic agreement, described as signed (9318), are firm, although Anthropic's value is undisclosed. The Hudson River Trading agreement (10732) comes only from secondary reporting. By contrast, IBM (4971), CrowdStrike (11768, 10330), the NVIDIA collaboration framework (3892), Exemplar Cloud status (10332, 9324) and the Genesis Mission (11778, 10331) are partnership or designation language with no disclosed dollar commitment. The same is true of the logo lists (11767, 10338, 9320) and the 'expanded relationships' with Cohere and Mistral (9319). The banks named in 11225, 9121 and 8736 are lenders, not customers. Several material counterparties remain anonymous. These include the customer financed by DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), and the 'two large, non-investment grade customers' behind DDTL 5.0 (9119). That last disclosure is the most credit-relevant one on file, and it has no names attached. Concentration has shifted without falling. In 2025 the book was anchored by OpenAI and Microsoft (648). Meta has since become a co-anchor, with about $35 billion in cumulative named commitments. Jane Street is the only newly named counterparty in the multi-billion range. The revenue arithmetic reconciles only partially. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), an increase of $45.6 billion. Named, quantified 2025 deals total about $25.4 billion (OpenAI plus Meta). The OpenAI and NVIDIA September order forms have no value in this record, so roughly $20 billion cannot be traced to a named source here. In Q2 2026, backlog rose only $4.8 billion to $104.2 billion (1302), less than Jane Street's $6 billion alone, which implies meaningful revenue burn-off or differences in booking timing. The more than $25 billion in early-Q3 commitments (1303, 3011) is unattributed; Hudson River Trading may account for part of it. Duplicative duration claims (3556, 3561, 3573, 3584) and the date mismatch in 10729 should be treated as sourcing noise.

CoreWeave's contract record is stronger than most AI-infrastructure peers because much of it rests on executed order forms under existing master services agreements, not on memoranda. The clearest signed agreements are the NVIDIA order form (claim 4352), the OpenAI order form (4296), which follows the $11.2 billion OpenAI backlog addition (4969), and two Meta order forms. The first Meta order form committed up to about $14.2 billion through 2031 (4282), and the second, dated March 31, 2026, added roughly $21 billion through 2032 (2123, 9317). Anthropic is described as 'signed' but without a dollar figure (9318). Jane Street's roughly $6 billion commitment (10304) is named and sized, though the record does not include an order-form filing for it.

A second tier is softer and should not be read as revenue. This includes the NVIDIA 'collaboration framework' (3892), the IBM 'partnership' (4971), the CrowdStrike 'global partnership' (11768, 10330), Genesis Mission participation (11778) and the Exemplar Cloud designations (10332, 9324). The recurring 'partner of choice' logo lists (11767, 9320) and the 'expanded relationships' with Cohere and Mistral (9319) name counterparties but disclose no terms. Hudson River Trading's August agreement (10732) is likewise unsized.

Several material obligations remain unidentified. These include the customer contract behind DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), the 'two large, non-investment grade customers' behind DDTL 5.0 (9119) and the colocation landlord in Indonesia (8760). Financing tied to non-investment-grade counterparties deserves particular attention.

Concentration has shifted rather than diluted. In 2025, OpenAI's $11.2 billion was the anchor against $25.9 billion of backlog (4970, 4969). By 2026, Meta's cumulative commitments of about $35 billion appear to rival or exceed it, and Microsoft and OpenAI are still cited among the largest customers (648).

The arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), which is broadly consistent with the named OpenAI, Meta and NVIDIA orders. Backlog then reached $99.4 billion by March 2026 (1302), a jump of about $39 billion in which the $21 billion Meta order explains only about half. The gap also mixes RPO and a broader backlog definition (4970). The second quarter added only $4.8 billion despite Jane Street's $6 billion, which implies revenue burn-down. The 'more than $25 billion' of new third-quarter commitments (1303, 3011) has no disclosed counterparty. Readers should also note that the August report of Meta 'adding' $21 billion (8668) restates the March deal and is not a new commitment.

CoreWeave's record contains a small group of contracts with filing-level specificity, surrounded by a much larger volume of partnership language. The firmest items are order forms executed under existing master services agreements. These cover OpenAI (4296), with an earlier $11.2 billion backlog addition (4969); Meta, at up to approximately $14.2 billion through 2031 (4282) and then about $21 billion through December 2032 under a March 31, 2026 order form (2123, 10340, 9317); and NVIDIA as a capacity purchaser (4352). The company's own release gives Jane Street's roughly $6 billion commitment (10304), and it states that the Anthropic agreement was signed (9318), though no value is given. The Hudson River Trading agreement (10732), and the dates attached to Anthropic and Jane Street (10731, 10695), come only from secondary press paraphrase. Several items do not evidence contracted revenue: the IBM 'partnership' (4971), the NVIDIA 'collaboration framework' (3892), the CrowdStrike 'global partnership' (11768, 10330), Genesis Mission participation (11778), the Leidos 'collaboration' (8669), and the Exemplar Cloud designations (10332, 9324). Logo rosters (11767, 10338, 9320) and 'expanded relationships' with Cohere and Mistral (9319) name counterparties but give no terms. The largest opacity sits in the financing disclosures. DDTL 3.0 funds 'a customer contract' (4595), DDTL 4.0 serves a 'leading AI enterprise' (11224), and DDTL 5.0 backs 'two large, non-investment grade customers' (9119), none of them identified. The more than $25 billion of commitments signed after the second quarter is also unattributed (1303, 3011). Concentration has broadened in name but perhaps not in dollars. In early 2025, OpenAI alone accounted for $11.2 billion of a $25.9 billion backlog (4969, 4970). Meta's two order forms now total roughly $35 billion, and press still describes Microsoft and OpenAI as among the biggest customers (648). The arithmetic broadly reconciles, but only loosely. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), which is consistent with Meta's $14.2 billion plus the OpenAI tranches. Meta's $21 billion falls on March 31 and so belongs in the first quarter, which helps explain a $99.4 billion backlog (1302), though backlog is a broader measure than RPO (4970). The second quarter rose only $4.8 billion, to $104.2 billion, despite Jane Street's roughly $6 billion and Anthropic's unvalued deal. That gap implies meaningful burn-down, or smaller-than-headline booked values, and the disclosures do not itemize either.

CoreWeave's signed customer record is stronger than most AI-infrastructure peers, but it is carried by a small number of counterparties. The firmest items are order forms executed under existing Master Services Agreements. These are with OpenAI (4296, following the $11.2 billion backlog addition in 4969), NVIDIA (4352), and Meta: $14.2 billion through 2031 (4282), then a roughly $21 billion order form dated March 31, 2026 (2123, 10340, 9317). Jane Street's roughly $6 billion commitment (10304) and the Anthropic agreement described as 'signed' (9318) also read as contracted, although Anthropic carries no disclosed value. Hudson River Trading's multi-year deal is sourced only secondhand (10732). Microsoft appears as a top customer only in press coverage (648). A second tier consists of relationship language rather than commitments. That tier includes the IBM 'partnership' (4971), the NVIDIA 'collaboration framework' (3892), the CrowdStrike 'global partnership' (11768), the Leidos 'collaboration' (8669), Genesis Mission participation (11778, 10331), Exemplar Cloud designations (10332, 9324), and logo lists (11767, 9320). Meta's expansion option (4283) is optionality, not backlog. Several of the most material contracts remain unidentified. These include the customer behind DDTL 3.0 (4595), the 'leading AI enterprise' financed by DDTL 4.0 (11224), and the 'two large, non-investment grade customers' behind DDTL 5.0 (9119). The last of these is a notable credit disclosure.

Concentration has broadened in name count but not clearly in dollars. In early 2025, OpenAI's $11.2 billion was about 43% of $25.9 billion backlog (4969, 4970). By 2026 Meta's cumulative roughly $35 billion arguably makes it the largest disclosed exposure, with OpenAI's September order value not stated in the record.

The arithmetic partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3571), a $45.6 billion increase. The quantified named deals explain only about $25 billion of it, and gross additions would have to exceed $45.6 billion because revenue recognized during the year consumes RPO. The undisclosed OpenAI and NVIDIA order values plausibly fill the gap, but this cannot be verified. Readers should also note that 'backlog' and 'RPO' are different measures: backlog includes amounts contingent on delivery (4970). The move from $60.7 billion RPO to $99.4 billion backlog therefore mixes definitions. Backlog grew only $4.8 billion in Q2 2026 (1302) despite Jane Street's $6 billion and the Anthropic deal. That implies meaningful burn-down or smaller-than-headline bookings. The $25 billion signed after quarter-end (1303, 3011) remains entirely unattributed.

The signed-contract record at CoreWeave is narrower than the partnership language around it. Three counterparties have executed order forms under master services agreements that the company disclosed in formal filings: OpenAI (claim 4296, following the $11.2 billion strategic deal in claim 4969), NVIDIA (claim 4352), and Meta, whose initial commitment of up to about $14.2 billion (claim 4282) was followed by a roughly $21 billion order form dated March 31, 2026 (claim 2123). That single Meta deal recurs across claims 10729, 10694, 10340, 9317 and likely 8668, so it should not be counted more than once. Jane Street's roughly $6 billion commitment (claim 10304) and the Anthropic agreement (claim 9318) are also described as signed, but Anthropic's value is undisclosed, and so are the terms of the Hudson River Trading agreement (claim 10732). A different category covers IBM (claim 4971), the NVIDIA collaboration framework (claim 3892), CrowdStrike (claims 11768 and 10330), Leidos (claim 8669) and the Genesis Mission (claim 11778). These are partnerships or collaborations with no disclosed dollar commitment. The logo lists in claims 11767, 10338 and 9320 are marketing, not contract disclosures.

Several of the largest financing-linked obligations remain anonymous. DDTL 3.0 finances "a customer contract" (claim 4595). DDTL 4.0 serves a "leading AI enterprise" (claim 11224). DDTL 5.0 backs contracts with "two large, non-investment grade customers" (claim 9119). The more than $25 billion in new commitments signed early in Q3 2026 carries no named counterparty (claim 1303).

Concentration has diluted in form but not clearly in substance. OpenAI alone was about 43% of the $25.9 billion backlog at March 2025 (claims 4969 and 4970). By 2026, Meta's roughly $35 billion in combined commitments rivals it, and Microsoft and OpenAI are still described as among the biggest customers (claim 648).

The revenue arithmetic reconciles only loosely. Backlog is not RPO: the $25.9 billion figure included just $14.7 billion of RPO, with the rest subject to delivery conditions (claim 4970). RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572 and 3571), which the named Meta and OpenAI deals can plausibly support. The Q2 2026 backlog, however, grew by only about $4.8 billion, to $104.2 billion (claim 1302), in a quarter that included Jane Street's $6 billion and the Anthropic agreement. That implies substantial burn-down from recognized revenue or modest undisclosed deal values. The value of the NVIDIA order form and the September 2025 OpenAI order form are not in the quoted record, and the record itself is truncated at claim 8668.

CoreWeave's record separates cleanly into executed order forms and softer relationship language. The strongest disclosures are 8-K order forms under existing master services agreements: OpenAI (4296, with $11.2 billion of backlog attributed to the strategic deal in 4969), NVIDIA (4352), and Meta, whose initial commitment of up to about $14.2 billion (4282) was followed by a roughly $21 billion order form dated March 31, 2026 (2123, 10340, 9317). Jane Street's approximately $6 billion commitment (10304) and the 'signed multi-year agreement' with Anthropic (9318) also read as contracts, though Anthropic carries no disclosed value. Hudson River Trading's agreement (10732) is sourced only to secondary press. By contrast, IBM (4971), CrowdStrike (11768, 10330), the NVIDIA 'collaboration framework' (3892), Leidos (8669) and the Genesis Mission (11778) are partnership or framework announcements with no disclosed economics. The 'partner of choice' logo lists (11767, 9320) name companies such as Cursor, Midjourney and Perplexity but do not describe commitments.

The most consequential counterparties are often unnamed. The DDTL 3.0 facility financed 'a customer contract' (4595), DDTL 4.0 serves a 'leading AI enterprise' (11224), DDTL 5.0 funds 'two large, non-investment grade customers' (9119), and the Indonesia build cites unspecified 'existing customer demand' (8762). Concentration has broadened from OpenAI alone to OpenAI and Meta, with Microsoft also named among the largest customers (648). OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog at March 31, 2025 (4969, 4970). Meta's disclosed commitments now total roughly $35 billion. Because the DDTL 5.0 borrowers are described as non-investment grade, a meaningful share of the newest backlog likely sits with weaker credits than the hyperscalers.

The revenue arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3555), a $45.6 billion increase. Named, quantified 2025 deals explain only about $14 billion of that, because the value of the September OpenAI order form is not on record. Backlog then reached $99.4 billion by the first quarter of 2026 and $104.2 billion by the second (1302). Meta's $21 billion explains only about half of the first-quarter step-up. Another $25 billion in post-quarter commitments is unattributed (1303, 3011). Readers should also note that 'backlog' includes amounts beyond RPO that are subject to delivery conditions (4970). RPO was $14.7 billion at March 31, 2025, below the year-end 2024 figure despite the OpenAI deal, so the two measures should not be compared directly. Finally, claim 10729 is dated before the March event it describes.

CoreWeave's record separates cleanly into executed contracts and softer relationship language, and the contracted layer is concentrated in a few names. The strongest evidence is the order forms filed under existing master services agreements: NVIDIA in September 2025 (claim 4352), OpenAI the same month (claim 4296), and Meta in September 2025 and again on March 31, 2026 (claims 4282, 2123). Anthropic is described as a signed multi-year agreement (claim 9318), and Jane Street as a roughly $6 billion commitment (claim 10304), but neither has a filed order form on record. The first Meta figure is a ceiling, 'up to approximately $14.2 billion' (claim 4282), and its expansion right is only an option (claim 4283). Softer items should not be read as revenue. These include the IBM 'partnership' (claim 4971), the NVIDIA 'collaboration framework' (claim 3892), the CrowdStrike 'global partnership' (claims 11768, 10330), Genesis Mission participation (claims 11778, 10331), and the rotating 'partner of choice' name lists (claims 11767, 9320). Several financing disclosures refer to counterparties that are never named: 'a customer contract' behind DDTL 3.0 (claim 4595), a 'leading AI enterprise' behind DDTL 4.0 (claim 11224), and 'two large, non-investment grade customers' behind DDTL 5.0 (claim 9119). The phrase 'leading AI labs and hyperscalers' (claim 8926) is likewise unidentified. Concentration has shifted rather than eased. In May 2025, OpenAI's $11.2 billion accounted for about 43% of the $25.9 billion backlog (claims 4969, 4970). Meta's commitments now total roughly $35 billion in named figures (claims 4282, 9317), and Microsoft and OpenAI are still cited as the largest customers (claim 648). The DDTL 5.0 disclosure adds a credit-quality concern that the named-customer narrative does not address. The revenue arithmetic reconciles only loosely. RPO rose from $15.1 billion to $60.7 billion during 2025 (claim 3555), which is broadly consistent with the OpenAI and Meta order forms, although the September 2025 OpenAI value is not quoted on record. Backlog reached $99.4 billion in Q1 2026 and $104.2 billion in Q2 2026 (claim 1302). The March Meta deal fits within Q1, but backlog is a broader measure than RPO and year-end 2025 backlog is not on record, so the two series cannot be bridged precisely. Q2 backlog grew only $4.8 billion even though the $6 billion Jane Street deal was announced in that quarter, which implies meaningful burn-down or other offsets. Recognized revenue is not on record to confirm this. Readers should treat the August 2026 '$21 billion' Meta reference (claim 8668) as a restatement of the March deal rather than a new award. The '$25 billion in new commitments' (claims 1303, 3011) has no named counterparty.

CoreWeave's signed contracts are mostly concentrated in a handful of order forms executed under existing Master Services Agreements. The clearest are the September 2025 OpenAI order form (claim 4296), the September 2025 NVIDIA capacity order (claim 4352), and two Meta order forms: an initial commitment of up to about $14.2 billion through 2031 (claims 4282, 4283) and a roughly $21 billion order dated March 31, 2026 (claims 2123, 9317, 10340). Anthropic's agreement is described as "signed" (claim 9318) but carries no disclosed value. Jane Street's roughly $6 billion commitment (claims 10304, 10730) is the only other named deal with a figure attached. Hudson River Trading's August agreement (claim 10732) is reported only secondhand. Much of the rest is looser. The IBM Granite "partnership" (claim 4971), the NVIDIA "collaboration framework" (claim 3892), the CrowdStrike "global partnership" (claims 11768, 10330), the Genesis Mission (claims 11778, 10331) and the Exemplar Cloud designations (claims 10332, 9324) are relationship or certification language, not disclosed revenue contracts. The rotating "partner of choice" lists (claims 11767, 10338, 9320) and the "expanded relationships" with Cohere, Jane Street and Mistral (claim 9319) name logos but give no terms. Several material obligations are left anonymous. DDTL 3.0 finances "a customer contract" (claim 4595), DDTL 4.0 serves a "leading AI enterprise" (claim 11224), and DDTL 5.0 backs "two large, non-investment grade customers" (claim 9119). All of these are facilities that single out specific counterparties without naming them. Concentration has shifted but not spread out. In May 2025, OpenAI's $11.2 billion accounted for about 43% of the $25.9 billion backlog (claims 4969, 4970). By late 2026, Meta's roughly $35 billion in cumulative commitments sits alongside OpenAI, and Microsoft and OpenAI are still described as "among the biggest" (claim 648). The arithmetic reconciles only in part. RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), and backlog reached $99.4 billion by Q1 2026 (claim 1302). Backlog, though, includes amounts that depend on delivery conditions (claim 4970), so the two series cannot be compared directly. Meta's $21 billion, signed March 31, explains roughly half of the Q1 gap, and the remainder is not attributed to any named deal. Backlog grew only $4.8 billion in Q2, to $104.2 billion, even though Jane Street's $6 billion and the Anthropic deal were announced that quarter. That points to heavy burn-down from recognized revenue, which cannot be verified here. The later "$21 billion" Meta references (claims 8668, 10729, 10694) appear to restate the March order and should not be counted again. The ">$25 billion" signed early in Q3 (claims 1303, 3011) has no identified counterparties.

CoreWeave's record splits into three tiers of evidentiary weight. The strongest are order forms executed under existing master services agreements and disclosed in filings: NVIDIA (4352), OpenAI (4296), and Meta's $14.2 billion commitment through 2031 (4282) and its March 31, 2026 order form (2123). These are signed, named, and in Meta's case priced. A second tier is named and signed but carries no disclosed value, such as the multi-year Anthropic agreement (9318, 10731) and Hudson River Trading's August deal (10732). Jane Street's roughly $6 billion commitment (10304, 10730) sits between those tiers, because it is quantified in the company's own release but has no matching filing on this record. The third tier is relationship language: the IBM Granite 'partnership' (4971), the NVIDIA 'collaboration framework' (3892), CrowdStrike (11768), the Genesis Mission (11778), Exemplar Cloud status (10332, 9324), and the rotating 'partner of choice' rosters (11767, 9320). None of these establishes contracted revenue. Several counterparties are unidentified altogether. The customer contract that funded DDTL 3.0 is unnamed (4595), as is the 'leading AI enterprise' behind DDTL 4.0 (11224). DDTL 5.0 serves 'two large, non-investment grade customers' (9119), a credit disclosure that deserves more attention than it gets.

Concentration has shifted from OpenAI to a two-anchor structure. At March 2025, OpenAI's $11.2 billion (4969) was about 43% of $25.9 billion in backlog (4970). Meta's commitments now total roughly $35.2 billion (4282, 9317). The August 2026 report describing Meta as having 'added $21 billion' (8668) most plausibly restates the March deal rather than a third tranche. By August, Microsoft and OpenAI were still cited among the largest customers (648).

The arithmetic reconciles only loosely. RPO rose $45.6 billion during 2025 (3555), but Meta's $14.2 billion is the only 2025 addition quantified on the record, since the NVIDIA and OpenAI order-form values are truncated here. The step from $60.7 billion RPO (3571) to $99.4 billion backlog at March 2026 (1302) is not a like-for-like comparison, because backlog also counts amounts beyond RPO (4970). Meta and Jane Street alone account for about $27 billion of the quarter's gain. Backlog then grew only $4.8 billion in Q2 (1302). The more than $25 billion signed early in Q3 (1303, 3011) has not been attributed to named counterparties.

CoreWeave's record separates cleanly into executed order forms and softer relationship language. The hardest evidence comes from order forms under existing Master Services Agreements: OpenAI in May 2025, adding $11.2 billion of backlog (claim 4969), and again in September 2025 (4296); NVIDIA's capacity order in September 2025 (4352); Meta's initial commitment of up to $14.2 billion through 2031 (4282); and Meta's roughly $21 billion March 31, 2026 order form (2123, 10340, 9317). Jane Street's approximately $6 billion commitment (10304) and the Anthropic multi-year agreement described as signed (9318) are also contractual, although no value is disclosed for Anthropic. The Hudson River Trading agreement (10732) rests on secondary reporting only. Several items read as contracts but carry weaker language: the IBM 'partnership' (4971), the NVIDIA 'collaboration framework' (3892), the CrowdStrike 'global partnership' (11768, 10330), Genesis Mission participation (11778), the Exemplar Cloud designations (10332, 9324) and the 'partner of choice' rosters (11767, 9320). None of these discloses a committed dollar figure. Meta's option to expand through 2032 (4283) was an option, not a commitment, until the March 2026 order form. Unidentified counterparties are a recurring feature. They include the customer behind DDTL 3.0 (4595), the 'leading AI enterprise' behind DDTL 4.0 (11224), the 'two large, non-investment grade customers' financed by DDTL 5.0 (9119), and the more than $25 billion of new commitments signed after the second quarter (1303, 3011), none of which is attributed. Concentration has shifted without clearly declining. In early 2025, OpenAI's $11.2 billion equaled about 43% of the $25.9 billion backlog (4970). By 2026, disclosed Meta commitments total about $35 billion, while Microsoft and OpenAI are still described as among the largest customers (648). The arithmetic reconciles only in part. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), an increase of about $45.6 billion. Only about $25.4 billion of that increase is dollar-identified, from OpenAI's May deal and Meta's initial commitment, because the values of the September OpenAI and NVIDIA order forms are truncated in the record. Backlog rose just $4.8 billion to $104.2 billion in the second quarter of 2026 (1302), even though the Jane Street and Anthropic signings fell in that quarter. That implies substantial run-off, or a value for Anthropic smaller than headlines suggest. The same $21 billion Meta deal appears in at least five claims (10729, 10694, 10340, 9317, 8668), and some claims record it as '$21,000', presumably in millions. Readers should count it once.

CoreWeave's record separates cleanly into a small set of executed, dollar-denominated order forms and a much larger layer of partnership and logo language. The signed contracts are OpenAI ($11.2 billion added to backlog, claim 4969, plus a September 2025 order form under the existing MSA whose value the quote omits, claim 4296), Meta (up to approximately $14.2 billion through 2031, claim 4282, and a roughly $21 billion order form executed March 31, 2026, claim 2123), NVIDIA as a capacity buyer (claim 4352), and Jane Street at about $6 billion (claim 10304). Anthropic is described as a signed multi-year agreement with no value (claim 9318), and Hudson River Trading appears only through secondary reporting of an August agreement (claim 10732). The Meta option to expand through 2032 (claim 4283) is an option, not a commitment. IBM (claim 4971), the NVIDIA collaboration framework (claim 3892), CrowdStrike (claims 11768, 10330), the Genesis Mission (claim 11778) and Exemplar Cloud status (claims 10332, 9324) carry no disclosed economics. The 'partner of choice' rosters (claims 11767, 9320) and the Cohere and Mistral expansions (claim 9319) are logos, not quantified contracts. Several material counterparties remain unnamed: the customer behind DDTL 3.0 (claim 4595), the 'leading AI enterprise' behind DDTL 4.0 (claim 11224), and the two large non-investment-grade customers behind DDTL 5.0 (claim 9119). Microsoft is cited as among the biggest customers (claim 648) but never appears as a disclosed contract in this record. Concentration has shifted rather than eased. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog in May 2025 (claims 4969, 4970). Meta's two order forms now total roughly $35 billion, about a third of the $104.2 billion Q2 2026 backlog (claim 1302). The newer trading-firm and lab names add breadth but little dollar weight. On arithmetic, RPO rose from $15.1 billion to $60.7 billion during 2025 (claims 3572, 3555), consistent with the Meta and OpenAI order forms. Backlog, however, is a broader measure than RPO (claim 4970), so the jump to $99.4 billion by Q1 2026 is not directly comparable. Q2 backlog grew only $4.8 billion despite Jane Street's $6 billion April commitment, which implies consumption or timing offsets that are not itemized. The $21 billion Meta deal is restated across at least six claims (10729, 10694, 10340, 2123, 9317, 8668), and it should be counted once. The $25 billion in post-quarter commitments (claims 1303, 3011) has no named counterparties.

CoreWeave's record separates cleanly into executed order forms and softer partnership language, and the executed layer is narrow. The binding agreements are order forms under existing Master Services Agreements with NVIDIA (claim 4352), OpenAI (claim 4296) and Meta (claims 4282, 2123). Meta's commitments are the only ones with disclosed dollar values in a filing: up to about $14.2 billion through 2031 (claim 4282) and a further roughly $21 billion through December 2032 signed March 31, 2026 (claims 2123, 9317). The Anthropic agreement is described as signed but has no value attached (claim 9318). Jane Street's roughly $6 billion comes from company and press statements rather than an order-form filing (claims 10304, 10730). Hudson River Trading's multi-year deal is sourced only to press (claim 10732). By contrast, IBM (claim 4971), the NVIDIA collaboration framework (claim 3892), CrowdStrike (claims 11768, 10330), Leidos (claim 8669), the Genesis Mission (claim 11778) and the Exemplar Cloud designations (claims 10332, 9324) are collaborations or recognitions with no disclosed economics. The logo lists in claims 11767 and 9320 name counterparties but not terms.

Several material counterparties go unnamed. These include the customer behind DDTL 3.0 (claim 4595), the "leading AI enterprise" financed by DDTL 4.0 (claim 11224), the "two large, non-investment grade customers" behind DDTL 5.0 (claim 9119) and the colocation landlord for Indonesia (claim 8760). The two non-investment-grade customers are the most diligence-relevant disclosure, because they tie debt directly to credit-weaker offtakers. Microsoft appears as a top customer only through press characterization (claim 648).

Concentration has shifted rather than diluted. OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog in March 2025 (claims 4969, 4970). Meta's disclosed commitments now total about $35 billion, roughly a third of the $104.2 billion Q2 2026 backlog (claim 1302), and OpenAI still sits alongside it.

The arithmetic only partly reconciles. RPO rose $45.6 billion during 2025 (claims 3571, 3572), yet the only quantified named deal after Q1 is Meta's $14.2 billion. The OpenAI and NVIDIA order-form values are not in the record. The step from $60.7 billion RPO to $99.4 billion backlog spans different metrics (claims 3555, 1302), so the roughly $17.7 billion not explained by Meta's $21 billion should be treated as unattributed rather than missing. Q2 backlog grew only $4.8 billion despite Jane Street's $6 billion and Anthropic, which implies meaningful burn-down. The "more than $25 billion" of early-Q3 commitments (claim 1303) is not yet tied to named counterparties. Claim 10729 is dated before the event it describes, and the record is truncated at claim 8668.

CoreWeave's record separates into a small set of contractually documented commitments and a much larger layer of partnership language. The signed agreements are the order forms filed under master services agreements. These cover OpenAI (4296, with an earlier $11.2 billion backlog addition in 4969), NVIDIA (4352), and Meta, which signed an initial commitment of up to $14.2 billion (4282) and a further roughly $21 billion order form on March 31, 2026 (2123, 10340, 9317). Jane Street's approximately $6 billion commitment (10304) and the Anthropic agreement described as signed (9318) also belong in this tier, although Anthropic carries no disclosed value. Hudson River Trading's multi-year agreement (10732) is reported only secondhand and without terms. Everything else should be read as non-binding or undimensioned. That includes the IBM partnership (4971), the NVIDIA collaboration framework (3892), CrowdStrike's global partnership (11768, 10330), Leidos's collaboration (8669), Genesis Mission participation (11778, 10331), Exemplar Cloud designations (10332, 9324), 'expanded relationships' with Cohere and Mistral (9319), and the 'partner of choice' logo lists (11767, 9320).

The unnamed counterparties matter more than the named ones. DDTL 3.0 financed an unidentified customer contract (4595). DDTL 4.0 served an unnamed 'leading AI enterprise' (11224). DDTL 5.0 is dedicated to two 'large, non-investment grade customers' (9119), a description that excludes Meta. The more than $25 billion of new commitments signed early in the third quarter (1303) is also unattributed, as is the Indonesia demand (8762). The debt-financed contracts, which are the most capital-intensive, are the least identified.

Concentration has shifted rather than diluted. In May 2025, OpenAI's $11.2 billion represented roughly 43% of $25.9 billion in backlog (4969, 4970), and Microsoft and OpenAI remained described as the largest customers in August 2026 (648). Meta's two disclosed commitments now total about $35 billion, so exposure has spread across two or three hyperscale-scale buyers rather than a broad base.

The arithmetic only partly reconciles. RPO rose from $15.1 billion to $60.7 billion during 2025 (3572, 3555), and the weighted duration was about five years (3584). Backlog, a broader measure that includes amounts beyond RPO (4970), reached $99.4 billion at the end of the first quarter of 2026. Meta's $21 billion explains roughly half of the implied increase, and the remainder is unattributed. Second-quarter backlog grew only $4.8 billion to $104.2 billion (1302), despite Jane Street's $6 billion and the Anthropic signing. That is consistent only if revenue burn offset these additions or Anthropic's value was modest, and neither can be verified from these disclosures. Claim 3011 carries no supporting text and should not be relied upon.

CoreWeave's record separates into three tiers. The signed tier consists of order forms executed under existing Master Services Agreements with named counterparties: NVIDIA (claim 4352), OpenAI (4296, following the $11.2 billion strategic deal in 4969) and Meta (4282, then the March 31, 2026 order form in 2123, which 10340 and 9317 value at about $21 billion). Jane Street's roughly $6 billion commitment (10304) and the Anthropic and Hudson River Trading multi-year agreements (9318, 10732) are described as signed, but no contract value is disclosed for Anthropic or Hudson River Trading. The softer tier is a partnership and framework layer that carries no stated economics. It includes the IBM partnership (4971), the NVIDIA collaboration framework (3892), CrowdStrike (11768), Leidos (8669), the Genesis Mission (11778) and the Exemplar Cloud designations (10332, 9324). The "partner of choice" logo lists (11767, 9320) and the "expanded relationships" with Cohere and Mistral (9319) also belong here, and none of them should be read as quantified commitments.

The unnamed tier matters most for diligence because it is where the financing sits. DDTL 3.0 funds "a customer contract" (4595). DDTL 4.0 serves a "leading AI enterprise" (11224). DDTL 5.0 is dedicated to "two large, non-investment grade customers" (9119). The more than $25 billion in commitments signed early in the third quarter is also unattributed (1303). Microsoft's status as a top customer comes only from press characterization (648), not from a company filing.

Concentration has broadened without dispersing. In May 2025, OpenAI's $11.2 billion was about 43% of the $25.9 billion backlog (4969, 4970). By early 2026, Meta's disclosed commitments totalled roughly $35 billion if the $21 billion is additive, as "new commitment" in 9317 suggests. That figure alone exceeds half of the $60.7 billion in RPO at year-end 2025 (3571). The debt-financed unnamed contracts, two of them explicitly below investment grade, add credit concentration on top of the customer concentration.

The arithmetic reconciles only partly. Backlog and RPO are different measures, because backlog adds amounts that are conditional on delivery (4970). That makes the move from $60.7 billion of RPO to $99.4 billion of backlog at the end of the first quarter (1302) partly definitional. Meta's $21 billion explains only about half of that increase. Second-quarter backlog rose just $4.8 billion to $104.2 billion (1302), even though Jane Street, Anthropic and the Cohere and Mistral expansions were announced in the same window. That implies substantial recognition burn-off, small undisclosed contract values, or both. With a weighted-average duration of about five years (3573), the record does not let the named deals be summed to the headline backlog.

CoreWeave's record contains a small group of agreements backed by filings and a much larger set of partnership and collaboration language. The documented agreements are order forms under existing master services agreements. These cover OpenAI (claims 4969, 4296), NVIDIA's capacity purchase (4352), and Meta: an initial commitment of up to $14.2 billion (4282) and a new order form dated March 31, 2026 (2123), described as a $21 billion commitment (9317, 10340). Jane Street's roughly $6 billion commitment (10304) and the Anthropic agreement, which the company calls signed (9318), also carry counterparty names, but no Anthropic dollar value is disclosed. The IBM partnership (4971), the NVIDIA 'collaboration framework' (3892), CrowdStrike (11768), Leidos (8669), and Genesis Mission participation (11778) should be read as framework or ecosystem language, not quantified contracts. The 'partner of choice' rosters (11767, 9320) name logos without terms. The Hudson River Trading agreement (10732) and Microsoft's status as a top customer (648) come from secondary reporting, not company disclosure.

Several material counterparties go unnamed. The DDTL 3.0 financing supports an unidentified 'customer contract' (4595). DDTL 4.0 serves a 'leading AI enterprise' (11224). DDTL 5.0 funds contracts with 'two large, non-investment grade customers' (9119). None of the more than $25 billion in commitments signed early in the third quarter (1303) is attributed to a named counterparty.

Concentration has changed shape but has not clearly declined. In May 2025, OpenAI's $11.2 billion accounted for about 43% of a $25.9 billion backlog (4969, 4970). Meta has since become at least as large, at $14.2 billion plus $21 billion, if the March order is additive, which the 'expansion' wording in 10729 leaves unclear. OpenAI and Meta together probably still dominate. A trading-firm group (Jane Street, HRT) adds some diversification at the margin.

The arithmetic only partly reconciles. RPO grew from $15.1 billion to $60.7 billion during 2025 (3572, 3571), which the Meta and OpenAI order forms can plausibly explain. Adding Meta's $21 billion and Jane Street's $6 billion to year-end RPO gives about $87.7 billion, against the $99.4 billion backlog three months later (1302). That leaves roughly $12 billion unassigned. Part of the gap is definitional, because backlog includes amounts beyond RPO that are subject to delivery conditions (4970). Q2 backlog rose a net $4.8 billion to $104.2 billion (1302) with no quantified named deal in the quarter, so readers should treat backlog and RPO as separate measures and rely on filed order forms for counterparty attribution.

CoreWeave's record contains a core of signed contracts, surrounded by a wider layer of partnership and logo language that should not be read as committed revenue. The signed contracts are order forms executed under existing Master Services Agreements. They cover OpenAI (claims 4969, 4296), Meta at $14.2 billion (4282) and again at roughly $21 billion on March 31, 2026 (2123, 10340, 9317), and NVIDIA (4352). Jane Street's approximately $6 billion commitment (10304) and the Anthropic multi-year agreement (9318) are also described as signed. Note that the Meta $21 billion figure appears in five claims (10729, 10694, 10340, 9317, 2123); these describe one contract, not five. The softer tier is different in kind. It includes the IBM Granite partnership (4971), the NVIDIA collaboration framework (3892), the CrowdStrike 'global partnership' (11768, 10330), the Leidos collaboration (8669), Genesis Mission participation (11778) and the Exemplar Cloud designation (10332, 9324). None of these carries a dollar figure or contract term. The 'partner of choice' logo lists (11767, 9320) are customer relationships of unknown size. Hudson River Trading's agreement (10732) comes only from secondary reporting and has no value attached.

Several material obligations are left unidentified. The DDTL 3.0 financing supports 'a customer contract' (4595), and DDTL 4.0 serves an unnamed 'leading AI enterprise' (11224). DDTL 5.0 funds contracts with 'two large, non-investment grade customers' (5.0, 9119). The more than $25 billion in commitments signed early in Q3 2026 (1303) are not attributed to any counterparty. Debt-financed buildouts tied to anonymous and explicitly sub-investment-grade counterparties are the main diligence gap.

Concentration has broadened, but it remains heavy. In March 2025, OpenAI's $11.2 billion equalled about 43% of the $25.9 billion backlog (4969, 4970). By August 2026, Microsoft and OpenAI were still described as the largest customers (648). Meta's disclosed commitments now total about $35.2 billion, roughly a third of the $104.2 billion backlog (1302), before accounting for any burn-down.

The arithmetic only partly reconciles. RPO rose $45.6 billion during 2025, to $60.7 billion (3555, 3572). Meta's $14.2 billion explains about 31% of that increase, and the September OpenAI and NVIDIA order forms have no disclosed values. There is also a definitional problem. 'Revenue backlog' includes amounts beyond RPO that are subject to delivery conditions (4970). The $99.4 billion and $104.2 billion backlog figures (1302) are therefore not directly comparable to the $60.7 billion RPO. Q2 backlog grew only $4.8 billion even though Jane Street's $6 billion landed that quarter, which is consistent with revenue burn-off but has not been confirmed. Claim 3011 carries no text and cannot be assessed.

CoreWeave's customer narrative reveals significant distinction between signed purchase commitments and promotional partnership language. Named, executed agreements include Meta's $21 billion cloud computing contract through December 2032 (claims 10340, 2123); Jane Street's $6 billion commitment (claim 10304); and order forms with OpenAI and NVIDIA under existing master service agreements dated 2023 (claims 4296, 4352). Anthropic and Hudson River Trading are referenced as having signed multi-year agreements, though amounts remain undisclosed (claims 9318, 10732). By contrast, NVIDIA is positioned through a "collaboration framework" rather than a purchase commitment (claim 3892), and IBM's partnership focuses on Granite model delivery without quantified supply obligations (claim 4971). CrowdStrike and Leidos are described as strategic partnerships or collaborations (claims 11768, 8669) rather than customer contracts. A lengthy roster of companies—Cognition, Midjourney, Runway, Cohere, Mistral, World Labs—appear in marketing language as "partners of choice" or referenced customers without disclosed contract values (claims 11767, 9320), functioning more as brand association than revenue transparency.

Counterparty concentration has become heavily Meta-dependent. Meta accounts for approximately $35 billion of clearly identified committed contracts as of March 2026, representing roughly 35–60% of disclosed backlog depending on methodology. Jane Street adds $6 billion. The remaining $60–70 billion in reported remaining performance obligations ($60.7 billion as of December 31, 2025, growing to $104.2 billion in Q2 2026) lacks granular attribution to named counterparties, creating opacity around secondary customer contributions (claims 3555, 1302).

Revenue arithmetic raises questions. CoreWeave reported $25.9 billion revenue backlog as of March 31, 2025, which reportedly included $11.2 billion from an OpenAI "strategic deal" announced May 2025 (claims 4970, 4969). By December 2025, RPO had tripled to $60.7 billion—an increase of approximately $35 billion in nine months (claims 3555, 3572). Named large deals (Meta, Jane Street) total roughly $27 billion, leaving $8–10 billion attributable to smaller or previously unannounced customers. The further jump to $104.2 billion backlog in Q2 2026, with $25 billion in commitments cited for the current quarter alone (claim 1303), suggests material undisclosed counterparties or expansion of previously announced relationships. The weighted-average five-year contract duration underscores long-tail revenue assumptions (claims 3584, 3573). Without itemization of the $35+ billion in non-Meta, non-Jane Street backlog, concentration risk and cash-generation sustainability remain difficult to validate independently.

CoreWeave distinguishes between formalized customer contracts and looser partnership arrangements, though the disclosure is inconsistent and raises questions about deal quality and concentration. The company's named, signed-customer base comprises primarily hyperscalers and financial-trading firms with documented order forms or multi-year commitments: Meta (claims 4282, 10340), OpenAI (claims 4296, 4969), NVIDIA (claim 4352), Jane Street (claim 10304), Anthropic (claim 9318), and Hudson River Trading (claim 10732). These six entities are the only counterparties disclosed with specific dollar values or formalized "order form" language under existing master services agreements. By contrast, early partnerships such as IBM (claim 4971) and CrowdStrike (claims 11768, 10330) are described as "partnership" or "collaboration framework" without revenue metrics, suggesting either non-binding frameworks or immaterial commitments relative to the headline figures.

The customer list published in February 2026 (claim 11767) includes Cognition, Cursor, Mercado Libre, Midjourney, and Runway alongside the named commitments, yet these firms appear nowhere else in earnings disclosure or in subsequent announcements with quantified commitments. This suggests either undisclosed order forms or marketing-driven listing of smaller customers without material financial impact. Concentration risk is acute: Meta alone represents approximately $35.2 billion of disclosed commitments (claim 10340), while OpenAI's named $11.2 billion backlog (claim 4969) and Jane Street's $6 billion (claim 10304) comprise the next material tranches. The remaining "strategic partnerships" and named customer relationships lack transparency regarding revenue contribution.

Revenue reconciliation shows gaps. The company reported $25.9 billion in total revenue backlog as of March 31, 2025, with $14.7 billion in remaining performance obligations (claim 4970). By December 31, 2025, RPO had expanded to $60.7 billion (claim 3555), a $46 billion increase within nine months. Announced deals in that period—Meta's $14.2 billion (claim 4282) and the OpenAI order form (claim 4296)—account for only partial closure of that gap, suggesting undisclosed commitments. By August 2026, backlog reached $104.2 billion (claim 1302), with an additional $25 billion in new commitments in early Q3 (claim 1303) remaining unaccounted for in customer press releases. This opacity obscures whether CoreWeave's growth derives from formalized, long-duration contracts or from less certain, smaller arrangements that benefit headline backlog figures.

CoreWeave's customer roster reveals a sharp distinction between named signed contracts and announced partnerships that remain vague on commercial substance. Of the company's $104.2 billion revenue backlog as of Q2 2026 (claim 1302), the largest commitments rest on explicit order forms: Meta's stacked agreements totaling approximately $35.2 billion through December 2032 (claims 4282, 10340), formalized via signed order form in March 2026 (claim 2123); OpenAI's $11.2 billion strategic deal (claim 4969), later documented via order form in September 2025 under an MSA dated April 2023 (claim 4296); NVIDIA's September 2025 order form (claim 4352) under the same April 2023 MSA; Jane Street's $6 billion commitment finalized in April 2026 (claim 10304); Anthropic's multi-year signed agreement (claim 9318); and Hudson River Trading's multi-year deal announced August 2026 (claim 10732). These six counterparties account for at least $67.2 billion of disclosed backlog, representing extreme concentration.

Announced partnerships employ softer language obscuring commercial certainty. IBM's "partnership to deliver compute capacity" for Granite models (claim 4971) lacks signed agreement language or quantified commitment. NVIDIA's January 2026 "collaboration framework" (claim 3892) remains distinct from the transactional order form executed months earlier. CrowdStrike's "major global partnership" (claim 10330) mentions collaboration without binding commitment; Leidos's August 2026 "collaboration" (claim 8669) similarly describes intent rather than signed order.

Concentration risk has intensified markedly. At March 31, 2025, total backlog stood at $25.9 billion (claim 4970). By December 2025, RPO had grown to $60.7 billion (claim 3555)—a 2.3x expansion—dominated by OpenAI's strategic deal and Meta's initial $14.2 billion entry. Meta's March 2026 expansion by $21 billion (claims 10694, 10729) transformed it into the dominant counterparty, now representing roughly one-third of total backlog. The company lists generic "major customer wins" (claim 10338) including Cognition, CrowdStrike, Cursor, Mercado Libre, Midjourney, and Runway without quantifying these relationships, leaving unclear whether they are pilots or material commitments.

Revenue arithmetic provides incomplete reconciliation. The $104.2 billion Q2 2026 backlog (claim 368) appears consistent with prior periods given Meta's $35.2 billion, OpenAI's $11.2 billion, and Jane Street's $6 billion. However, $25 billion in additional commitments reportedly signed through August 2026 (claims 1303, 3011) remain unattributed to named counterparties, suggesting either unannounced customers or undisclosed framework additions to existing contracts. This opacity complicates assessment of whether growth is driven by depth with existing customers or genuine customer diversification.

CoreWeave's announced customer base distinguishes clearly between signed order forms under existing master service agreements and partnership or framework arrangements. The company has named six primary customers with explicit signed commitments: Meta (initially $14.2 billion through December 2031, expanded to $21 billion through December 2032 in March 2026), Jane Street (approximately $6 billion), OpenAI (order form dated September 23, 2025), NVIDIA (new order form as of September 14, 2025), Anthropic (multi-year agreement signed by May 2026), and Hudson River Trading (multi-year agreement announced August 2026). Claims 4282, 10304, 4296, 4352, 9318, and 10732 respectively document these as signed contractual commitments rather than framework discussions. By contrast, the IBM partnership (claim 4971) remains limited to a collaboration framework for delivering Granite model compute capacity, and the NVIDIA relationship evolved from a signed order form into a broader "collaboration framework" language (claim 3892, January 2026), suggesting a shift from discrete customer transactional arrangements toward strategic partnering. The company also cites unidentified counterparties as "leading AI enterprise" (claim 11224) and "two large, non-investment grade customers" (claim 9119), each requiring dedicated infrastructure financing, without naming them publicly.

Counterparty concentration has narrowed materially. CoreWeave's revenue backlog composition shifted from OpenAI representing approximately 43 percent of the $25.9 billion total in May 2025 (claim 4969) to Meta anchoring the disclosed base at $21 billion by Q2 2026 (claim 10340), with Jane Street contributing $6 billion (claim 10304). The named customer cohort of six signed contracts represents a meaningful portion of the $104.2 billion Q2 2026 RPO figure (claim 1302), but unnamed large customers and the approximately $25 billion in commitments signed through August 2026 (claim 3011) suggest concentration is distributing wider. The weighted-average contract duration of five years across the committed base (claim 3556) supports durability claims, though the revenue arithmetic reveals incomplete transparency: the company discloses approximately $27 billion in named large-contract commitments (Meta $21 billion plus Jane Street $6 billion) but does not quantify OpenAI or Anthropic's contract values in later disclosures, leaving material revenue backlog components unattributed to named counterparties.

CoreWeave's disclosed customer commitments exhibit significant asymmetry between named, signed agreements and promotional language around partnerships, alongside material revenue backlog growth that exceeds the sum of identified customer commitments. The company has clearly documented signed customer orders from Meta (claims 4282, 10340, 2123: $14.2 billion through December 2031, expanded with a $21 billion order through 2032 announced March 2026, and an additional $21 billion commitment by August 2026), Jane Street (claims 10304, 10695: $6 billion multi-year), and signed order forms from OpenAI (claim 4296, September 2025) and NVIDIA (claim 4352, September 2025) under existing master service agreements. Anthropic (claims 10731, 9318) and Hudson River Trading (claim 10732) announced multi-year agreements without disclosed dollar amounts.

However, significant customer relationships are either unnamed or described through partnership rather than customer-contract language. IBM's "partnership" (claim 4971) for Granite compute lacks commitment specification. NVIDIA's engagement appears split: formal order forms exist (claim 4352), yet a later "collaboration framework" (claim 3892) suggests expanded but less rigidly committed work. CrowdStrike (claim 10330) is anchored as a "major global partnership," not a named customer commitment. Leidos (claim 8669) appears through a "collaboration" rather than contracted capacity. Unnamed cohorts persist: a "material customer contract" (claim 4595) remains unidentified; claim 9119 references "two large, non-investment grade customers" without naming.

Counterparty concentration has become extreme. Meta alone represents roughly $56 billion of disclosed commitments across three tranches (2025 commitment of $14.2 billion, March 2026 expansion of $21 billion, August 2026 addition of $21 billion). OpenAI's $11.2 billion strategic deal (claim 4969, May 2025) and subsequent order form (claim 4296) anchor two of CoreWeave's largest relationships. The balance of disclosed commitments—Jane Street's $6 billion, Anthropic and Hudson River Trading's unquantified multi-years—suggests a customer base concentrated among a handful of hyperscalers and large AI labs.

Revenue backlog progression reveals a gap. Total backlog was $25.9 billion in May 2025 (claim 4970), then grew to $60.7 billion by year-end 2025 (claim 3555), and reached $104.2 billion by Q2 2026 (claim 1302). The disclosed named commitments (Meta $56 billion, OpenAI ~$11 billion, Jane Street $6 billion, others unquantified) approach but do not obviously reconcile with the $104 billion backlog, particularly when accounting for the five-year weighted-average contract duration (claims 3584, 3561). The apparent gap suggests either undisclosed large customers, preliminary commitments later formalized, or the inclusion of option and capacity-expansion provisions as contracted backlog rather than mere optionality.

CoreWeave's disclosed customer base exhibits a stark contrast between named, contractually committed capacity and vaguely identified partnerships. Of the major announcements, Meta ($35.2 billion across claims 4282 and 10340), OpenAI (strategic deal language in claim 4969 plus September 2025 order form in claim 4296), Jane Street ($6 billion in claim 10304), and Anthropic (multi-year signed agreement in claim 9318) represent explicitly named, signed customer relationships. By contrast, several announcements employ softer language: IBM is described as a partnership (claim 4971) rather than a signed customer contract, and NVIDIA's relationship is framed as a collaboration framework (claim 3892) despite underlying MSA activity (claims 4352, 3892). A material gap exists between CoreWeave's claimed market presence and identified customers. Claim 8926 asserts the company serves most of the world's leading AI labs and hyperscalers, yet claims 11224 and 9119 reference "leading AI enterprise" and "two large, non-investment grade customers" without identification. Claim 648 mentions Microsoft and OpenAI as "among the biggest" customers, yet Microsoft has never been formally announced as a customer by CoreWeave. Claim 8338 lists named enterprise wins (Cognition, CrowdStrike, Cursor, Mercado Libre, Midjourney, Runway) but provides no quantified commitments. Counterparty concentration has evolved dramatically. In May 2025, OpenAI dominated announced commitments at $11.2 billion (claim 4969). By March-April 2026, Meta emerged as the single largest customer following its $21 billion expansion (claim 10340), representing approximately 20 percent of disclosed backlog. Jane Street ($6 billion) and Hudson River Trading (claim 10732, multi-year agreement) occupy secondary positions. Revenue backlog arithmetic reveals substantial unnamed customer value. The disclosed backlog grew from $25.9 billion in Q1 2025 (claim 4970) to $60.7 billion by year-end 2025 (claim 3555), then to $104.2 billion by Q2 2026 (claim 1302), with $25 billion in new commitments signed through August 2026 (claim 1303). The named deals—Meta ($35.2 billion), OpenAI ($11.2 billion), Jane Street ($6 billion), plus Anthropic and Hudson River Trading (unquantified)—account for approximately $52-55 billion of the $104+ billion backlog. This suggests unnamed customers represent roughly 50 percent of disclosed commitments. Financing disclosures highlight this opacity: claims 9119 and 11224 reference major customer contracts used to justify capital raises without customer identification, indicating material revenue concentration among parties CoreWeave does not publicly name.

CoreWeave's announced customer commitments exhibit a stark bifurcation between named, signed deals and vaguely attributed revenue. The company has disclosed three explicitly binding agreements with order forms under master services agreements: Meta ($14.2 billion through December 2031, expanded by an additional $21 billion in March 2026 per claims 4282 and 2123), OpenAI (new order form signed September 2025 per claim 4296, following the strategic deal that added $11.2 billion in backlog per claim 4969), and NVIDIA (new order form September 2025, claim 4352). Jane Street's $6 billion multi-year commitment (claims 10304, 10730) and Hudson River Trading's unnamed multi-year agreement (claim 10732) appear substantive, while Anthropic's multi-year agreement (claims 10731, 9318) carries no disclosed amount. However, a second cohort lacks contractual specificity: CrowdStrike, Cognition, Cursor, Mercado Libre, Midjourney, and Runway are listed as "major customer wins" (claim 10338) without any financial disclosure, while Adaption Labs, Advaita Bio, Perplexity, and World Labs are labeled "partner of choice" (claim 9320) using language that avoids customer-contract semantics. Additionally, NVIDIA's relationship is framed as a "collaboration framework" (claim 3892) rather than a services contract, and Leidos is positioned as a "collaboration" (claim 8669) for sovereign AI services without disclosed commitments. Concentration is extreme: Meta alone represents approximately $35 billion of the June 2026 backlog, or roughly one-third of the $104.2 billion total (claim 1302), with OpenAI constituting another material but unquantified share. The remaining backlog includes unnamed customers—specifically "two large, non-investment grade customers" (claim 9119) and an unnamed "leading AI enterprise" (claim 11224) receiving dedicated financing facilities. The revenue arithmetic strains credibility: CoreWeave reported $15.1 billion in RPO at December 2024 (claim 3572), grew to $60.7 billion by December 2025 (claim 3555), then to $104.2 billion by Q2 2026 (claim 1302), with an additional $25 billion in unintegrated new commitments signed through August (claims 1303, 3011). The $25.9 billion backlog disclosed in May 2025 (claim 4970) already embedded the $14.7 billion RPO, suggesting the Meta and OpenAI figures dominate historical disclosure. Unless the 2026 $25 billion in new commitments or portions of the existing $104.2 billion backlog derive from named customers, CoreWeave's customer base remains materially opaque despite headline announcements, with revenue concentration and counterparty credit quality poorly disclosed relative to the magnitude of commitments.

CoreWeave's disclosed customer base reveals significant concentration around named counterparties and substantial opacity regarding committed-contract identities. As of August 2026, the company reported $104.2 billion in contracted revenue backlog, yet publicly named customer agreements account for approximately $41.2 billion: Meta Platforms at $35.2 billion (combining the original $14.2 billion commitment through December 2031 with a March 2026 expansion worth $21 billion), Jane Street at $6 billion, and three others—Anthropic, OpenAI, and Hudson River Trading—with signed order forms but undisclosed dollar amounts. Each of these represents executed customer order forms under existing master service agreements (claims 4282, 10304, 9317, 8570, 10730, 9318, 4296, 10732), distinguishing them from partnership or collaboration frameworks.

Framework and partnership announcements occupy a separate category. IBM, NVIDIA, CrowdStrike, and Leidos are described as "partnerships" or "collaborations" (claims 4971, 3892, 10330, 8669) rather than signed customer contracts with disclosed revenue impact. The company lists enterprises including Cognition, Cursor, Mercado Libre, Midjourney, and Runway as wins (claim 10338), yet provides neither binding contract language nor contract values for these names.

The backlog arithmetic raises questions about counterparty opacity. Named and quantified customer contracts total roughly $41.2 billion, leaving approximately $63 billion—or 60 percent—of the reported $104.2 billion backlog unexplained by disclosed names. The company disclosed that its May 2026 debt facility supports infrastructure "dedicated to customer contracts with two large, non-investment grade customers" (claim 9119), neither of which has been publicly identified. This concentration pattern—where the three largest named customers (Meta, Jane Street, and one unnamed pair via financing disclosures) account for a material share of backlog—suggests either undisclosed concentration or a long tail of smaller commitments that the company declines to itemize. The evolution from $15.1 billion RPO at December 31, 2024 to $104.2 billion by Q2 2026 occurred largely through named-customer expansion, yet the rate of backlog growth outpaces disclosed contract announcements, implying material commitments remain unnamed in public filings.

CoreWeave's customer disclosure shows a material distinction between executed contracts and partnership announcements, with significant revenue backlog concentration among named parties that remains partially unattributed. The company has confirmed binding agreements with OpenAI, Meta, Jane Street, Anthropic, and Hudson River Trading through references to executed order forms under existing master service agreements (claims 4296, 2123, 10304, 9318, 10732). Meta represents the largest named commitment at approximately $35.2 billion when combining its initial $14.2 billion commitment (claim 4282) and subsequent $21 billion expansion (claim 10729), while OpenAI contributed $11.2 billion to the backlog (claim 4969) and Jane Street committed $6 billion (claim 10304). By contrast, partnerships with IBM, NVIDIA, and CrowdStrike employ framework or collaboration language without signed customer orders or quantified revenue (claims 4971, 3892, 10330). Cursor, Cognition, Midjourney, Runway, and Mercado Libre are listed as "major customer wins" (claim 10338) without supporting documentation of binding agreements or contract values, while Perplexity and World Labs appear only as aspirational "partners of choice" (claim 9320).

Revenue backlog reconciliation reveals a significant gap. CoreWeave disclosed $104.2 billion in total backlog as of August 2026, with $25 billion in new commitments signed through early Q3 (claims 1302, 1303). Named customers account for approximately $52.4 billion: Meta $35.2 billion, OpenAI $11.2 billion, Jane Street $6 billion, plus unspecified amounts for Anthropic and Hudson River Trading. This leaves roughly $52 billion unattributed to identifiable counterparties. While Microsoft and OpenAI are noted as "among the biggest" customers (claim 648), Microsoft has no disclosed signed order form, suggesting either undisclosed bilateral arrangements or reliance on generic customer categories. The company's own reporting emphasizes "serving most of the world's leading AI labs, hyperscalers and AI enterprises" (claim 8926), language that masks individual customer concentration and makes the backlog largely opaque to investor scrutiny. Counterparty concentration has intensified around Tier-1 AI platform companies—Meta, OpenAI, Anthropic—while claimed enterprise breadth lacks verifiable signatures.

CoreWeave has selectively named customers while using partnership language for strategically important relationships. Signed customer agreements with disclosed amounts include OpenAI ($11.2 billion backlog announced May 2025, order form executed September 2025 per claim 4296), Meta ($14.2 billion initial commitment September 2025 per claim 4282, expanded by $21 billion in March 2026 per claim 10340), and Jane Street ($6 billion commitment announced April 2026, claim 10304). These three counterparties account for $52.4 billion of explicitly sized commitments. Anthropic and Hudson River Trading have announced signed multi-year agreements (claims 9318 and 10732), but no dollar figures are disclosed for either.

Elsewhere, CoreWeave uses partnership or framework terminology that does not clearly indicate binding purchase agreements. IBM is cited as having an announced partnership for Granite capacity (claim 4971). NVIDIA is described as having a collaboration framework (claim 3892) as of March 2026, distinct from an earlier order form for capacity (claim 4352). CrowdStrike is characterized as a major global partnership (claim 10330), Leidos as a collaboration (claim 8669), and Cohere and Mistral as expanded relationships (claim 9319) without dates or sizing. The May 2026 reference to CoreWeave as partner of choice for Adaption Labs, Advaita Bio, Perplexity, and World Labs (claim 9320) uses aspirational language with no signature or commitment language attached.

Concentration has evolved in ways the backlog figures do not fully clarify. OpenAI represented $11.2 billion of $25.9 billion total backlog in May 2025, a 43 percent share. The Q2 2026 backlog of $104.2 billion (claim 1302) is not reconciled with named customer commitments: OpenAI, Meta, and Jane Street together total $52.4 billion, leaving $51.8 billion unexplained before early Q3 additions of $25 billion-plus (claim 8573). Microsoft is cited as among the biggest customers (claim 648) without corresponding announcement. CoreWeave's own disclosure identifies two unnamed non-investment grade customers whose infrastructure needs justified facility-level debt financing (claim 9119), indicating material concentration among a core group of undisclosed AI labs.

CoreWeave's customer portfolio divides between explicitly named signed orders under existing Master Service Agreements and partnership frameworks lacking quantified commitments or clear binding force. Named customers with disclosed dollar commitments cluster narrowly. Meta appears across three separate announcements: an initial $14.2 billion commitment through December 2031 (claim 4282, September 2025), an expanded agreement for $21 billion through December 2032 (claim 10340, April 2026), and a further $21 billion commitment added by August 2026 (claim 8668). The temporal sequence—initial deal in September 2025, expansion announced in April 2026, additional commitment noted in August 2026—suggests either additive capacity expansion or material repricing and restatement of terms, a distinction with concentration implications. Jane Street secured $6 billion in commitment (claim 10304, April 2026), Anthropic a multi-year agreement of unspecified magnitude (claim 9318, May 2026), OpenAI a September 2025 order form (claim 4296) building on an earlier $11.2 billion backlog disclosure (claim 4969, May 2025), and NVIDIA both an order form (claim 4352, September 2025) and a "collaboration framework" (claim 3892, March 2026). The framework language is subordinate to signed orders, yet CoreWeave elevates it in partnership messaging.

Microsoft recurs as one of the "biggest customers" (claims 648, 8656) without disclosed contract quantification or signature date, creating asymmetry versus named competitors. CoreWeave further references "two large, non-investment grade customers" in infrastructure financing (claim 9119) entirely without identification, and lists Cursor, Cognition, and Midjourney as "major customer wins" (claim 10338) absent any contract detail. MasterClass selected CoreWeave "to evaluate" AI teaching agents (claim 8617), evaluation language distinct from contracted commitment.

Revenue backlog reconciliation reveals underlying tension. Remaining Performance Obligations jumped fourfold from $15.1 billion (December 2024) to $60.7 billion (December 2025, claim 3571), inadequately explained by disclosed named deals. CoreWeave cites a "contracted revenue backlog" of $104 billion (claim 8538) while separately excluding $25 billion in early Q3 commitments (claim 8573), implying non-binding LOI or framework language pads the backlog number. A single Meta commitment at $21 billion represents roughly twenty percent of the $104 billion total, concentrating portfolio risk in one counterparty across potentially ambiguous deal terms.

CoreWeave's customer portfolio reveals a sharp demarcation between legally binding arrangements with named counterparties and aspirational partnership language. Signed, executable agreements are documented through order forms under existing master services agreements: Meta's two commitments totaling $35.2 billion (claim 4282 for $14.2 billion in September 2025, claim 9317 for $21 billion signed March 2026), OpenAI's order form of September 2025 (claim 4296) layered atop a strategic deal adding $11.2 billion (claim 4969 from May 2025), NVIDIA's order form (claim 4352, September 2025), and Anthropic's multi-year agreement (claim 9318). By contrast, IBM is described only as a partnership to "deliver compute capacity" (claim 4971) without contractual specificity. The May 2026 announcement naming Adaption Labs, Advaita Bio, Hudson River Trading, Perplexity, and World Labs as partners of choice (claim 9320) discloses no commercial terms or binding commitments. NVIDIA's January 2026 "collaboration framework" (claim 3892) sits between partnership and order, and MasterClass was merely selected "to evaluate" CoreWeave's platform (claim 8617), positioning this as a pilot rather than a committed customer contract.

Counterparty concentration has intensified dramatically. Meta and OpenAI dominate the revenue backlog: Meta's $35.2 billion across two executed orders and OpenAI's cumulative backlog constitute the core of CoreWeave's disclosed $104.2 billion Q2 2026 backlog (claim 8520). The financing of infrastructure for two "large, non-investment grade customers" through the DDTL 5.0 facility (claim 9119) indicates credit risk concentration and a tier below hyperscaler counterparties.

Revenue backlog arithmetic shows internal consistency but extreme concentration. The backlog expanded from $15.1 billion (December 2024, claim 3572) to $60.7 billion (December 2025, claim 3571) to $104.2 billion (Q2 2026, claim 8520), with additional $25 billion committed in early Q3 2026 (claim 8573). This four-quarter arc derives almost entirely from Meta's dual tranches and OpenAI's cumulative arrangement, alongside the two undisclosed non-investment-grade contracts financed through DDTL 5.0. The company's revenue base exhibits minimal diversification below these principal counterparties despite rhetoric positioning smaller firms as strategic partners.

CoreWeave's disclosed customer portfolio presents a two-tier structure: named, contracted counterparties with disclosed revenue commitments, and framework partnerships or unnamed customers whose contribution to the $104.2 billion backlog remains opaque. The named customer set includes OpenAI, Meta, Microsoft, NVIDIA, Anthropic, and IBM, each with signed order forms or multi-year agreements. Meta's commitment evolved from $14.2 billion through December 2031 (claim 4282) to an additional $21 billion signed in March 2026 (claim 9317, 8570), representing the most material single customer escalation. OpenAI's May 2025 "strategic deal" disclosed $11.2 billion in revenue backlog (claim 4969) against a total backlog of $25.9 billion (claim 4970), indicating 43 percent concentration in that single counterparty at that time.

By contrast, several partnerships employ framework or pilot language that obscures commitment clarity. NVIDIA's "collaboration framework" announced January 2026 (claim 3892) lacks a disclosed revenue figure, as does IBM's "announced partnership" for Granite compute (claim 4971). Weights & Biases is described as an "expanded capability" rather than a customer commitment (claim 9328). Cohere, Jane Street, and Mistral are characterized as "expanded relationships" (claim 9319) without new revenue quantification. Most problematically, CoreWeave references "leading AI pioneers and enterprises" as "partner of choice" (claim 9320)—naming only Adaption Labs, Advaita Bio, Hudson River Trading, Perplexity, and World Labs without attributing revenue—and two "large, non-investment-grade customers" driving the DDTL 5.0 facility financing (claim 9119) remain entirely unnamed.

The revenue backlog arithmetic strains credibility. From May 2025 to June 2026, disclosed backlog grew 302 percent, from $25.9 billion to $104.2 billion (claims 4970, 8282), with an additional $25 billion in new Q3 commitments (claim 8573). This $129.2 billion projected total does not reconcile against publicly disclosed individual customer values: OpenAI's $11.2 billion plus Meta's $21 billion accounts for only 24 percent of the June backlog. Unnamed or framework-partner additions must represent the preponderance of growth, yet the company simultaneously emphasizes it "serves most of the world's leading AI labs" (claim 8926) without naming them. This opacity suggests either material customers lack public disclosure consent, or backlog definitions or calculation methodology have shifted. The weighted-average five-year contract duration across the backlog (claims 3561, 3584) provides contract stability but masks concentration risk if two or three named counterparties represent disproportionate value.

CoreWeave's announced customer base separates into three clarity tiers. Signed customer agreements with binding order forms are the strongest: OpenAI and Meta have executed new order forms dated September 2025 and March 2026 respectively (claims 4296, 2123); NVIDIA signed an order form in September 2025 (claim 4352); Anthropic received a multi-year agreement signature in May 2026 (claim 9318); Hudson River Trading, Leidos, and MasterClass were announced with definitive language by August 2026 (claims 1234, 8669, 8617). IBM, by contrast, received only a "partnership" announcement for Granite compute delivery (claim 4971), which lacks signed contract clarity. Cohere, Jane Street, and Mistral are described as "expanded relationships with existing enterprise and AI native customers" without specifying new contract closure (claim 9319)—this language suggests continuation rather than newly signed deals. MasterClass's announcement uses "selected CoreWeave Cloud to evaluate and improve" language (claims 8617, 7782), which reads evaluative or pilot-stage rather than a full production commitment. NVIDIA's May 2026 collaboration is explicitly termed a "collaboration framework" (claim 3892), distinct from a customer order form. Critically, infrastructure financing for DDTL 3.0 and DDTL 5.0 facilities references unnamed "large, non-investment grade customers" (claims 4595, 9119), obscuring material revenue concentrations from public view. Counterparty concentration has remained elevated and sticky. OpenAI's $11.2 billion strategic deal backlog was announced in May 2025 (claim 4969) within a $25.9 billion total backlog (claim 4970)—representing 43 percent concentration in one customer. Meta's September 2025 commitment of $14.2 billion (claim 4282) was followed by a March 2026 order for an additional $21 billion (claims 9317, 8570), creating cumulative Meta visibility of approximately $35.2 billion. OpenAI plus Meta together represent roughly $46.2 billion of the June 2026 revenue backlog of $104.2 billion (claim 1302)—approximately 44 percent concentration in two customers, consistent with the May 2025 ratio. The $25 billion in new commitments signed in early Q3 2026 (claim 1303) remains unattributed to specific named customers, limiting transparency into whether concentration has improved. Applied Digital's Polaris Forge 1 lease represents $11 billion of prospective revenue (claim 371), but disclosure of which customer drives this facility remains absent. The revenue arithmetic reveals material backlog growth from $25.9 billion (May 2025) to $60.7 billion (December 2025) to $104.2 billion (June 2026)—consistent with the announced deals, though the absence of named counterparties for unnamed infrastructure contracts and the opacity around the $25 billion early Q3 commitments prevent full audit trail validation.

CoreWeave's customer announcements reflect a portfolio heavily concentrated among a small set of named hyperscalers, with significant revenue backlog allocated to unnamed counterparties. Signed customer agreements—those disclosing binding order forms under existing MSAs—are confirmed for OpenAI (claim 4296, Sept 23, 2025), Meta (claims 4282, 2123, Feb–Mar 2026), NVIDIA (claim 4352, Sept 14, 2025), and Anthropic (claim 9318, May 2026). Each agreement specifies execution mechanics (order form, MSA date, or contract language) and, where disclosed, term length or commitment ceiling. By contrast, partnerships with IBM (claim 4971, announced May 2025) and NVIDIA's January 2026 collaboration framework (claim 3892) employ looser language—'announced partnership' and 'collaboration framework'—without order form or binding commitment disclosures. MasterClass (claims 8617, 7782, Aug 2026) is framed as an evaluation and pilot, not a committed purchase. Several named counterparties—Cohere, Jane Street, Mistral, Hudson River Trading, Perplexity—appear in claim 9319 and 9320 as 'expanded relationships' or 'partners of choice' without contract specificity or dollar quantification.

Counterparty concentration has widened dramatically. As of December 31, 2025, RPO stood at $60.7 billion (claim 3555); by June 30, 2026, revenue backlog reached $104.2 billion (claim 1302). Meta alone accounts for $35.2 billion ($14.2 billion signed Sept 29, 2025, plus $21 billion signed March 2026, claims 4282, 9317). OpenAI's backlog contribution is pegged at $11.2 billion from a strategic deal announced May 2025 (claim 4969), though September 2025 and subsequent order forms lack disclosed amounts. NVIDIA's capacity purchase (Sept 2025) and Leidos collaboration (Aug 2026) remain unquantified. This leaves $46.8 billion of the $104.2 billion backlog unattributed to named customers—roughly 45 percent. Claim 9119 references 'two large, non-investment-grade customers' for infrastructure deployment without identifying them by name, further obscuring concentration risk.

Revenue arithmetic does not reconcile disclosed backlog to named deals. Identified commitments (Meta $35.2 billion, OpenAI $11.2 billion) sum to $46.4 billion, leaving a gap of $57.8 billion as of mid-2026. Anthropic's multi-year agreement is confirmed as 'signed' (claim 9318) but carries no dollar amount. MasterClass and other named partners carry no explicit revenue commitments. CoreWeave discloses $25 billion in new commitments secured in early Q3 2026 (claims 1303, 8573), bringing the backlog to approximately $129 billion by inference, yet the customer composition of these incremental commitments is entirely unnamed (claims 8267, 8261). The weighted-average contract duration of five years (claim 3556) and the shift from $15.1 billion RPO (Dec 2024) to $60.7 billion (Dec 2025) suggest material customer wins, but the absence of granular disclosure on unnamed counterparties limits visibility into concentration and revenue stability.

CoreWeave's partnership announcements reveal a material distinction between signed customer contracts and softer framework or partnership language. Named customers with executed order forms or binding agreements include OpenAI (claim 4296), Meta (claims 2123, 8668), NVIDIA (claim 4352), Anthropic (claim 9318), Hudson River Trading (claim 7598), and MasterClass (claim 8617). By contrast, IBM is described only as a partner to deliver capacity for Granite models without contractual detail (claim 4971), while NVIDIA's collaboration framework language (claim 3892) appears exploratory rather than revenue-bearing. The announcement on May 6, 2026, listing partnerships with Adaption Labs, Advaita Bio, Perplexity, and World Labs (claim 9320) contains no indication of signed commitments or quantified spend.

Counterparty concentration has shifted markedly. Through mid-2025, OpenAI's $11.2 billion strategic deal (claim 4969) dominated disclosed backlog. Meta's emergence as a customer with $14.2 billion committed through December 2031 (claim 4282) and subsequent expansion to $21 billion (claim 8668) by March 2026 now positions Meta as the largest named customer, representing approximately 20 percent of the June 2026 backlog of $104.2 billion (claim 8520). Yet this concentration may mask deeper risk: the company discloses that DDTL 5.0 facility proceeds support infrastructure for two large, non-investment-grade customers (claim 9119) whose identities remain undisclosed, suggesting either commercial sensitivity around credit quality or significant customer hiding.

The revenue arithmetic presents a reconciliation gap. Identifiable named customers—OpenAI, Meta ($35.2 billion combined), Anthropic, Hudson River Trading, MasterClass, and Applied Digital's Polaris Forge commitment ($11 billion, claim 371)—account for roughly $57 billion of the $104.2 billion backlog disclosed in June 2026. The remaining $47 billion remains attributed to unnamed counterparties. While the company notes $25 billion in new Q3 commitments (claim 1303), the identity of these customers is not disclosed. This gap between named and unnamed customer backlog, coupled with specific mention of two undisclosed non-investment-grade customers requiring dedicated infrastructure financing, suggests either significant opaqueness around customer identity or concentration among customers CoreWeave prefers not to name in public filings.

CoreWeave's customer announcements conflate signed agreements with framework partnerships. Named counterparties with binding order forms include Meta (signed September 2025 and March 2026 for $14.2 billion and $21 billion respectively, totaling $35.2 billion through 2031; claims 4282, 9317), OpenAI (order form finalized September 2025 with amount unspecified, though May 2025 language citing a "strategic deal" added $11.2 billion to revenue backlog; claims 4296, 4969), NVIDIA (order form executed September 2025, amount undisclosed; claim 4352), and Anthropic (multi-year agreement announced May 2026 with commitment amount withheld; claim 9318). These carry binding contractual language via master service agreements and order forms.

Partnerships without disclosed signed contracts are labeled "framework," "collaboration," or simply "partnership." IBM's involvement is described solely through "partnership" terminology (claim 4971). NVIDIA's January 2026 announcement positions their relationship as a "collaboration framework" (claim 3892), distinct from the September order form, indicating supplier and strategic partner roles. Microsoft appears repeatedly as "among the biggest" customers (claims 648, 8656) yet CoreWeave discloses zero order form amounts, commitment dates, or purchase obligations—a striking omission for a supposed anchor customer. Leidos is announced as a "collaboration" for sovereign AI services (claim 8669). Five companies named as "partners of choice"—Adaption Labs, Advaita Bio, Hudson River Trading, Perplexity, World Labs—carry no disclosed commitment terms (claim 9320). MasterClass "selected CoreWeave Cloud to evaluate...AI teaching agents" (claim 8617), indicating pilot status, not a signed customer contract. Most opaque: two "large, non-investment grade customers" drive the DDTL 5.0 facility's capital deployment (claim 9119) without identification, precluding concentration analysis.

Customer concentration has tightened dramatically toward Meta and OpenAI. Meta's $35.2 billion in quantified commitments dwarfs peers. OpenAI's true commitment remains ambiguous—the May 2025 "strategic deal" cites $11.2 billion, yet the September order form specifies no amount. Customer backlog arithmetic broadly reconciles: RPO grew from $14.7 billion within March 2025's $25.9 billion revenue backlog (claim 4970) to $60.7 billion by December 2025 (claim 3555), then $104.2 billion in Q2 2026 (claim 8546). Meta's disclosed commitments ($35.2 billion) plus OpenAI ($11.2 billion+) account for roughly $46 billion of the $89.5 billion backlog growth from March 2025 to Q2 2026. The remainder—including $25+ billion of new commitments added in early Q3 2026, NVIDIA's and Anthropic's undisclosed amounts, and the two unnamed DDTL 5.0 customers—remain opaque. This opacity obscures whether CoreWeave's top-three customers represent 50%, 70%, or 90% of backlog, a material concentration metric for assessing customer diversification and revenue sustainability.

CoreWeave's disclosed customer commitments reveal concentration among three named hyperscalers—Meta, OpenAI, and Anthropic—while substantial unnamed parties account for the majority of announced backlog. Meta is the only customer with fully disclosed, staged commitments: $14.2 billion through December 2031 (signed order form, September 2025, claim 4282), an additional order form (April 2026, claim 2123), and a $21 billion incremental commitment (August 2026, claims 8668, 8570), totaling $35.2 billion identified. OpenAI secured an $11.2 billion strategic deal allocation (May 2025, claim 4969), formalized by signed order form (September 2025, claim 4296), but no incremental commitment amount is disclosed. Anthropic signed a multi-year compute agreement for Claude development (August 2026, claim 8571) without disclosed magnitude.

Secondary-tier named customers include Hudson River Trading (multibillion-dollar multiyear agreement, August 2026, exact value withheld, claim 7609), Jane Street ($6 billion commitment, August 2026, claim 7611), and IBM (partnership for Granite capacity, May 2025, no revenue specified, claim 4971). MasterClass is described as "selected" for evaluation, indicating pilot rather than binding commitment (claim 8617). Leidos was framed as a "collaboration framework" for sovereign defense AI services (August 2026, claim 8669), not a traditional customer agreement. Microsoft appears only as "among the biggest" customers (August 2026, claim 648) with no dated order form.

The backlog arithmetic presents a significant reconciliation gap. As of June 30, 2026, CoreWeave disclosed $104.2 billion contracted backlog plus $25+ billion in early Q3 commitments (claims 1302, 8282). Named commitments—Meta $35.2 billion, Jane Street $6 billion, plus unquantified allocations for OpenAI, Anthropic, and Hudson River Trading—account for less than half the disclosed backlog. A May 2026 financing disclosure references "two large, non-investment-grade customers" supporting infrastructure (claim 9119), but these counterparties remain publicly unnamed. Backlog expanded from $15.1 billion RPO (December 2024, claim 3572) to $60.7 billion RPO (December 2025, claims 3555, 3560), then to $104.2 billion (June 2026), yet the announced customer roster does not explain this trajectory.

NVIDIA's relationship requires clarification: a September 2025 order form (claim 4352) may indicate CoreWeave hardware purchases, while a January 2026 "collaboration framework" (claim 3892) is positioned as partnership infrastructure, not direct revenue commitment.

CoreWeave's disclosed customer portfolio exhibits acute concentration risk masked by backlog growth rhetoric. Between May 2025 and June 2026, named customers account for approximately 40–50% of disclosed commitments, with Meta and OpenAI dominating. Meta's position crystallized through two signed order forms: $14.2 billion committed through December 2031 (claim 4282, Sept 29, 2025) and an additional $21 billion signed in March 2026 (claim 8570), aggregating roughly $35.2 billion of the $104.2 billion backlog reported at Q2 2026. OpenAI secured $11.2 billion in revenue backlog by May 2025 (claim 4969) and signed a new order form in September 2025 (claim 4296), though incremental commitment size remains undisclosed. Hudson River Trading announced a multi-year deal characterized as multi-billion-dollar (claim 1234, 1058) but without specific revenue quantification.

The remaining disclosed backlog—roughly 50–60%—belongs to entities identified only through vague language. Anthropic committed through a "multi-year agreement" (claim 8571) with no dollar figure. NVIDIA, initially appearing as a customer, operated under an existing Master Services Agreement (claim 4352) with a new order form signed September 2025, yet NVIDIA remains CoreWeave's primary GPU supplier and is better characterized as a partner than customer. IBM and Leidos received announcement treatment as "partnerships" and "collaborations" (claims 4971, 8669) without binding order form language. Two large, non-investment-grade customers (claim 9119) driving facility financing remain entirely unnamed. Microsoft, mentioned as "among the biggest" (claim 648), has no disclosed signed agreement in the record.

Revenue arithmetic reveals aggressive backlog expansion. Remaining performance obligations grew from $15.1 billion at December 31, 2024, to $60.7 billion at December 31, 2025—a 300% increase—then reached $104.2 billion by June 30, 2026 (claims 3555, 1302). CoreWeave cited $25 billion in net new commitments secured in early Q3 2026 (claims 8283, 8573) but did not publicly identify the counterparty. This unattributed quarter represents 24% of Q2 backlog. The weighted-average contract duration of approximately five years (claim 3584) suggests backlog concentration creates execution risk and customer substitution risk, particularly if Meta or OpenAI materially reduce utilization or renegotiate.

CoreWeave's customer disclosures distinguish sharply between binding signed agreements and strategic framework language. Named customers with signed order forms under existing master service agreements include Meta (claims 4282, 8570: $35.2 billion committed through 2031 plus expansion optionality), OpenAI (claims 4296, 4969: $11.2 billion explicitly disclosed for strategic deal, with additional September 2025 order form of unspecified magnitude), NVIDIA (claim 4352: order form under 2023 MSA, amount undisclosed), Hudson River Trading (claims 1234, 958: multi-year deal, exact volume unstated but described as multi-billion), and Anthropic (claim 8571: multi-year agreement, value not disclosed). Applied Digital represents approximately $11 billion in contracted lease revenue (claim 371). By contrast, partnerships lack binding language: IBM's Granite arrangement (claim 4971) is merely an announcement to deliver capacity; NVIDIA's January 2026 collaboration framework (claim 3892) explicitly uses framework terminology rather than order-form language; Leidos' August 2026 collaboration (claim 8669) is similarly styled; and MasterClass (claim 8617) appears to be an evaluation or pilot, described as selection to "evaluate and improve" teaching agents. Two large customers financing the DDTL 3.0 facility (claim 4595) and two others in DDTL 5.0 (claim 9119) remain completely unnamed, with only credit ratings disclosed.

Concentration has intensified. Meta alone represents 34% of disclosed backlog commitments ($35.2 billion of $104.2 billion as of June 2026, claims 1302, 8282). OpenAI, Applied Digital, and Hudson River Trading collectively account for an estimated additional $32+ billion in named capacity, implying roughly 64% of the $104.2 billion backlog (claim 1302) derives from four counterparties. The backlog exploded from $15.1 billion at year-end 2024 to $60.7 billion nine months later (claim 3555) and $104.2 billion by Q2 2026 (claim 1302)—a 588% increase in eighteen months. The company notes weighted-average contract duration of five years (claims 3584, 3556), suggesting backlog scales with acquisition velocity rather than price inflation.

Revenue arithmetic presents transparency gaps. Disclosed customer commitments—Meta ($35.2B), OpenAI ($11.2B), Applied Digital (~$11B), and Hudson River Trading (unspecified but multi-billion)—account for perhaps $57.4 billion of the $104.2 billion backlog, leaving $46.8 billion (45%) unexplained. Unnamed customers in infrastructure financings, Microsoft's mention (claim 648) without quantification, and several other partnerships lack specificity. The distinction between RPO (remaining performance obligations: $60.7 billion, claim 3555) and total revenue backlog ($104.2 billion) itself remains undefined in the claims provided, limiting reconciliation of accounting versus marketing narratives.

CoreWeave's disclosed customer base exhibits a stark divide between named, binding agreements and partnership rhetoric lacking contractual clarity. The company has secured actual signed order forms from OpenAI (September 2025), Meta (December 2023 MSA with order forms in April 2026), NVIDIA (April 2023 MSA with September 2025 order form), and newer multi-year agreements with Anthropic and Hudson River Trading, all of which carry explicit contract execution language and defined payment terms. Conversely, the IBM partnership announced May 2025 remains described only as a 'strategic deal' contributing $11.2 billion to backlog without a disclosed order form or MSA date—leaving the enforceability and timing of IBM's $11.2 billion commitment materially ambiguous relative to the Meta and OpenAI arrangements. The NVIDIA relationship evolved from existing MSA to a January 2026 'collaboration framework,' language that typically signals non-binding intent rather than revenue-bearing obligation. Supporting details like MasterClass and Leidos are characterized as 'evaluations' and 'collaborations' respectively, suggesting pilot rather than committed-revenue status.

Counterparty concentration has intensified dramatically. Meta and OpenAI dominate: Meta committed $14.2 billion through December 2031 and added $21 billion in March 2026, alone representing a substantial portion of the $104.2 billion Q2 2026 backlog. OpenAI's backlog contribution remains undisclosed post-September 2025 order form. Critically, CoreWeave references two unnamed 'large, non-investment grade customers' (May 2026 debt facility disclosure) and generic language like 'serving most of the world's leading AI labs' without identifying material counterparties, obscuring true customer concentration risk.

Revenue arithmetic raises material reconciliation questions. The disclosed named contracts—Meta $35.2 billion cumulative, IBM $11.2 billion, Applied Digital Polaris Forge $11 billion—total approximately $57.4 billion, yet RPO reached $60.7 billion by December 2025 and $104.2 billion by June 2026. The gap between identified counterparties and stated backlog has widened, not narrowed. August 2026 announcements of '$25 billion in new commitments' lack counterparty identification entirely. This structure suggests either massive undisclosed customer concentration or a reliance on shorter-duration commitments and estimated future demand rather than binding orders, materially complicating assessment of backlog quality and revenue visibility.

CoreWeave's disclosure record distinguishes sharply between binding orders and framework language, though material opacity persists around customer identity and concentration. Signed customer agreements include: Meta ($14.2 billion through December 2031 under an order form dated September 2025, claim 4282, plus a $21 billion commitment via order form on March 31, 2026, claim 2123; these appear to be separate tranches totaling ~$35 billion); OpenAI (order form executed September 23, 2025, claim 4296, amount undisclosed); Jane Street ($6 billion capacity commitment, claim 7611); Hudson River Trading (multi-year infrastructure agreement, claim 1234); Anthropic (multi-year agreement for Claude deployment, claim 8571); Backblaze (multi-year data storage agreement, claim 7645); and NVIDIA (new order form under existing MSA, claim 4352, amount undisclosed). By contrast, IBM's May 2025 announcement (claim 4971) uses vague "partnership" language without contract specifics. NVIDIA's January 2026 announcement (claim 3892) describes a "collaboration framework" rather than a customer contract. MasterClass's August 2026 selection (claim 8617) reads as evaluation pilot language ("evaluate and improve"), not signed capacity commitment. Leidos (claim 8669) is characterized as "collaboration" for defense services without binding contract disclosure.

Counterparty concentration has intensified toward hyperscalers. Early 2025 rhetoric emphasized OpenAI and IBM; by August 2026, Meta dominates disclosed commitments at ~$35 billion, with OpenAI and Microsoft identified as "among the biggest" (claim 648) but without quantified commitments. The six-month backlog expansion from $60.7 billion RPO (December 31, 2025, claim 3555) to $104.2 billion (June 30, 2026, claim 1302) mirrors $25+ billion in unattributed new Q3 commitments (claim 8573), suggesting significant customer wins remain undisclosed by name.

Revenue arithmetic exhibits material gaps. Meta's contracts account for roughly one-third of disclosed June 2026 backlog; Jane Street adds $6 billion. Remaining ~$63 billion includes OpenAI, Microsoft, NVIDIA, Anthropic, and others—but specific customer contract values are absent. The claim that $104.2 billion excludes $25+ billion in early-Q3 commitments (claim 8283) implies total signed backlog near $129 billion by mid-2026, yet customer-level breakdowns necessary to verify concentration risk remain undisclosed. No reconciliation addresses whether Microsoft (listed as customer, claim 8656) has signed binding agreements with disclosed values, or whether Microsoft's position derives from pilot arrangements or framework discussions.

CoreWeave's disclosure of customer agreements distinguishes clearly between binding order forms and collaborative frameworks, though material gaps in counterparty identification and deal attribution limit revenue reconciliation.

Signed customer contracts with explicitly named counterparties include OpenAI (claim 4296, Sept 2025), which executed an order form under an existing master services agreement; Meta Platforms (claims 4282 and 2123, Sept 2025 and March 2026), which committed $14.2 billion through December 2031 and subsequently added $21 billion in March 2026 (claim 8570); NVIDIA (claim 4352, Sept 2025), under its existing MSA dated April 2023; and Hudson River Trading (claim 7598, Aug 2026), described as a multi-year, multi-billion-dollar agreement without specific dollar disclosure. Jane Street is named with a $6 billion commitment (claim 7611, Aug 2026), and Anthropic with a multi-year agreement (claim 8571, Aug 2026). These represent binding revenue backlog with enforceable legal instruments.

Framework arrangements and evaluations lack revenue weight. IBM's partnership (claim 4971, May 2025) uses explicitly non-transactional language. NVIDIA's collaboration framework (claim 3892, Jan 2026) is similarly promotional. MasterClass (claim 8617, Aug 2026) was selected to evaluate and improve AI teaching agents—a platform pilot, not a committed customer contract.

Counterparty concentration is dominated by Meta, which represents $35.2 billion across two tranches, or roughly 34 percent of the $104.2 billion revenue backlog as of June 30, 2026 (claim 8538). Microsoft and OpenAI are identified as among the largest customers (claim 648, Aug 2026), but recent deal values are undisclosed; the May 2025 $11.2 billion OpenAI contribution (claim 4969) predates the September order form and lacks update.

Revenue attribution reveals substantial gaps. Meta ($35.2B) plus Jane Street ($6B) plus Applied Digital's prospective Polaris Forge lease revenue ($11B, claim 371) total $52.2 billion, leaving approximately $52 billion of the $104.2 billion backlog unattributed to named customers. The company separately disclosed $25 billion in early Q3 commitments (claim 8521) without counterparty identification. Either unnamed customers represent material backlog concentration, or deal structures—including pilot programs or framework arrangements—are embedded within backlog figures.

CoreWeave's disclosed customer base reveals sharp concentration among AI giants alongside material vagueness about counterparty identity. Named, signed agreements cluster around hyperscalers: Meta accounts for approximately $35.2 billion in contracted capacity through 2032, combining the $14.2 billion initial commitment of December 2025 per claim 4282 with the subsequent $21 billion expansion signed in March 2026 per claim 8570. OpenAI represents $11.2 billion in committed revenue backlog (claim 4969). Jane Street, Hudson River Trading, and Applied Digital contribute material but smaller positions—Jane Street $6 billion (claim 7611), Applied Digital $11 billion as leased infrastructure revenue (claim 371)—while Hudson River Trading's multiyear arrangement lacks a specific figure (claim 7609). Anthropic and Backblaze appear signed to multiyear agreements (claims 8571, 7645) without disclosed volumes.

Yet the largest share of CoreWeave's $104.2 billion Q2 2026 backlog (claims 8520, 1302) remains sourced from unnamed counterparties. Leidos is described through "collaboration" framework language rather than binding orders (claim 8669). MasterClass is explicitly characterized as an evaluation pilot, not a production commitment (claim 8617). Microsoft appears only as an unnamed "major customer" referenced alongside OpenAI (claim 648). An undisclosed colocation provider and "existing customer demand" drive Southeast Asia deployment (claims 8760, 8762) without identification.

The revenue arithmetic raises reconciliation questions. Named, publicly committed deals total roughly $63 billion when summed: Meta ($35.2B), OpenAI ($11.2B), Applied Digital ($11B), Jane Street ($6B). Anthropic, Backblaze, Hudson River Trading, and other named signatories lack disclosed contract values. This leaves approximately $41 billion in backlog unexplained by public announcements—either drawn from unnamed customers, framework arrangements not yet formalized into orders, or pending disclosure. The progression from $60.7 billion RPO at year-end 2025 (claim 3555) to $104.2 billion by Q2 2026, plus claimed $25 billion in "new commitments" in early Q3 (claim 8573), suggests aggressive acceleration, though the distinction between signed orders and LOI-stage commitments remains opaque outside the named tier.

CoreWeave has disclosed a substantive customer roster, though the nature and materiality of commitments varies significantly across counterparties. Named customers with explicitly signed order forms include OpenAI (claim 4296, September 2025), Meta (claims 4282, 2123, $14.2 billion initial commitment plus $21 billion expansion in March 2026 per claim 8570), NVIDIA (claim 4352, September 2025), and Hudson River Trading (claims 958, 1234, multi-year agreement announced August 2026). Anthropic (claim 8571) and Jane Street (claim 7611, $6 billion commitment) also appear with signed agreements. By contrast, IBM is described only as a partnership (claim 4971) for delivering capacity, lacking specificity on commitment terms, and NVIDIA's more recent collaboration framework announcement (claim 3892, January 2026) uses structurally softer language than the preceding order-form disclosures.

The roster reveals sharp concentration risk. Meta has evolved from $14.2 billion contractual obligation (claim 4282, September 2025, through December 2031) to $35.2 billion total commitment when combining the March 2026 expansion (claim 8570). OpenAI accounts for $11.2 billion of backlog (claim 4969). These two counterparties represent the material spine of disclosed arrangements. Microsoft and OpenAI are cited as among the biggest customers (claim 648) without quantified commitments; this opacity is material.

Revenue backlog reconciliation reveals a critical disclosure gap. Remaining performance obligations surged from $15.1 billion (December 2024) to $60.7 billion (December 2025, claim 3555), then to $104.2 billion (June 2026, claim 1302). The six-month jump from year-end 2025 to mid-2026 exceeded $43 billion. Named additions in that window include Meta's $21 billion (March 2026, claim 8570) and Jane Street's $6 billion (claim 7611, announced August). These account for approximately $27 billion of the $43 billion growth. The remaining $16-18 billion of growth and the separately disclosed $25 billion in net new commitments signed in early Q3 (claims 1303, 8261) remain unattributed to named counterparties, suggesting material undisclosed customers or framework commitments not yet reflected as binding order forms. Pilot arrangements—notably MasterClass (claim 8617, described as evaluation), which lacks committed spend—sit outside this backlog arithmetic and should not be conflated with revenue-bearing contracts. The IBM partnership similarly lacks committed revenue specification despite its headline prominence.

CoreWeave's customer portfolio distinguishes sharply between binding signed order forms and strategic framework arrangements. Named signed customers with quantified commitments include Meta ($14.2 billion through 2031, claim 4282, plus $21 billion commitment, claim 8570), OpenAI (new order form signed September 2025, claim 4296, building on $11.2 billion strategic deal from May 2025, claim 4969), Anthropic (multi-year agreement, claim 8571), Hudson River Trading (multi-year infrastructure deal, claim 7558), Jane Street ($6 billion commitment, claim 7611), and NVIDIA (new order form September 2025, claim 4352). By contrast, IBM's relationship is labeled partnership framework for Granite compute (claim 4971), NVIDIA's January 2026 arrangement is explicitly a collaboration framework (claim 3892), and Leidos is described as a collaboration for sovereign AI services (claim 8669), none of which disclose customer revenue. MasterClass represents pilot language—selected to evaluate AI teaching agents (claim 8617), not a committed spending obligation.

Counterparty concentration has intensified dramatically. Meta and OpenAI represent the disclosed centerpiece: Meta alone accounts for approximately $35.2 billion of named commitments (combining claims 4282 and 8570), while OpenAI's total remains partially opaque beyond the $11.2 billion strategic deal and subsequent order forms. By June 2026, Microsoft and OpenAI together were described as among the biggest customers (claim 648), suggesting a two-to-three customer dependency. A single unnamed material customer drove DDTL 3.0 facility financing (claim 4595), implying additional concentration risk not publicly identified.

Revenue arithmetic broadly reconciles but with opacity. March 2025 backlog stood at $25.9 billion including $14.7 billion RPO (claim 4970). By December 2025, RPO had expanded to $60.7 billion (claim 3555), implying $46 billion in net new commitments. Named signed deals—Meta's $14.2 billion plus option, OpenAI's ongoing orders, Anthropic, Hudson River ($6+ billion implied), Jane Street ($6 billion)—account for approximately $45+ billion, matching the gap. By June 2026, total backlog reached $104.2 billion (claim 1302), with an additional $25 billion signed in early Q3 (claim 1303). The arithmetic holds consistently across five-year weighted-average contract duration (claim 3556), though the concentration of two or three hyperscalers and the shadowed customer suggests revenue quality risk concentrated in AI infrastructure spending cycles.

CoreWeave's customer portfolio has shifted materially from heavily concentrated to more diversified, though quantified exposure to specific counterparties remains opaque in critical cases. Through mid-2025, the backlog narrative hinged on OpenAI's $11.2 billion strategic commitment (claim 4969), which represented nearly half of the stated $25.9 billion revenue backlog as of March 31, 2025 (claim 4970). By September 2025, Meta entered as a named customer with a signed order form committing $14.2 billion through December 2031 (claim 4282), supplemented by an expansion option through 2032 (claim 4283). This Meta commitment increased further to $21 billion through a March 2026 order form (claim 8570), suggesting either renewal or expansion rather than replacement. NVIDIA appears in binding terms via order forms dated September 14, 2025 (claim 4352), though the January 2026 announcement shifted to looser "collaboration framework" language (claim 3892), signaling a downgrade from execution certainty. OpenAI similarly announced a new order form on September 23, 2025 (claim 4296) under an existing Master Services Agreement, confirming continued relationship but leaving incremental value undisclosed.

Recent announcements exhibit concerning vagueness. Anthropic and Hudson River Trading are described via multi-year agreements (claims 8571, 7598), yet Hudson River Trading's commitment remains labeled only as "multi-billion-dollar" without specification (claim 7609). Jane Street's $6 billion commitment is named (claim 7611), but MasterClass's involvement is framed as evaluation—"selected CoreWeave Cloud to evaluate and improve AI teaching agents" (claim 8617)—implying pilot rather than committed capacity. IBM's arrangement remains at "partnership" level without binding language (claim 4971). The backlog arithmetic tracks approximately: base $25.9 billion growing to $60.7 billion RPO by December 2025 (claim 3555), then $104.2 billion by June 2026 (claim 8282), plus $25 billion claimed in early Q3 2026 (claim 8573). Visible additions (Meta's combined $35.2 billion, Jane Street's $6 billion, OpenAI's base $11.2 billion) approximate $52 billion, leaving ~$26 billion attributable to Hudson River Trading, Anthropic, Backblaze, and other undisclosed parties. The widening gap between backlog expansion and disclosed customer names—particularly the unquantified $25 billion Q3 contribution—raises questions about whether the remaining commitments represent binding orders or looser framework commitments.

CoreWeave's publicly identified customers represent only a portion of its reported revenue backlog, highlighting material concentration risk in unnamed counterparties. Named signed customers with specific dollar commitments include OpenAI ($11.2 billion backlog as of May 2025, with a September 2025 order form under its existing Master Services Agreement, claim 4296), Meta ($21 billion commitment through December 2032 announced March 2026, claim 8570, building on an earlier $14.2 billion through-2031 agreement from September 2025, claim 4282), Jane Street ($6 billion, claim 7611), and Hudson River Trading (multibillion-dollar multiyear agreement, claims 7604/7609). Anthropic is disclosed with a multiyear agreement for Claude model compute (claims 8571/7630) and Backblaze with a multiyear storage partnership (claim 7645), though their dollar commitments remain undisclosed. IBM is named for a Granite compute partnership (claim 4971); NVIDIA appears in both an order form (September 2025, claim 4352) and a collaboration framework (January 2026, claim 3892, non-binding language). The explicitly named commitments—approximately $38 billion to OpenAI, Meta, and Jane Street combined, plus undisclosed sums from Hudson River Trading and Anthropic—fall substantially short of CoreWeave's stated $104.2 billion revenue backlog as of June 30, 2026 (claim 8520), leaving $50+ billion in unannounced customer concentration. RPO escalated from $15.1 billion (December 2024, claim 3572) to $60.7 billion (December 2025, claim 3571) to $104.2 billion (June 2026, claim 8520), with an incremental $25 billion in commitments secured in early Q3 2026 (claim 8573). Microsoft is mentioned as among the biggest customers (claim 648) without quantification. The weighted-average five-year contract duration (claim 3561) provides revenue visibility, yet the opacity regarding unnamed customer concentration—specifically whether any undisclosed counterparty exceeds Meta's disclosed $21 billion—creates material risk in assessing revenue durability and customer relationship resilience.

CoreWeave's named-customer base is exceptionally transparent and verifiable through SEC filings and press releases, distinguishing most deals from vague "strategic partnership" nomenclature. Confirmed signed agreements include OpenAI (claim 4969: $11.2 billion backlog as of May 2025; claim 4296: September 2025 order form under existing MSA), Meta (claim 4282: $14.2 billion commitment through December 2031, later revised to $21 billion through 2032 per claim 7628), Jane Street (claim 7611: $6 billion capacity spend), Hudson River Trading (claims 7604–7611: multiyear agreement explicitly "multibillion-dollar" per claim 7609, with early Vera Rubin chip access), Anthropic (claim 7630: multiyear compute deal), and Backblaze (claim 7645: multiyear data storage agreement). NVIDIA's engagement is more ambiguous, characterized as a "collaboration framework" (claim 3892) rather than a binding customer contract, and IBM's Granite partnership (claim 4971) carries partnership semantics rather than purchase commitments.

Counterparty concentration has compressed vertically. Meta and OpenAI dominate at an estimated $35+ billion combined against a total backlog of $104.2 billion as of Q2 2026 (claim 8520). The quantitative trading bloc—Jane Street, Hudson River Trading, IMC, and Flow Traders (claim 1792)—forms a secondary cluster. This oligopolistic structure diverges sharply from traditional infrastructure providers.

Revenue arithmetic reveals material gaps. CoreWeave's backlog expanded from $25.9 billion (March 2025, claim 4970) to $60.7 billion RPO (December 2025, claim 3555) to $104.2 billion (June 2026, claim 8520)—a fourfold increase in 15 months. The intermediate jump to $60.7 billion suggests a major contract booking not fully articulated in concurrent announcements, likely reflecting amendments to Meta's or OpenAI's commitments. The $25 billion in "new commitments signed early Q3" (claim 1303) remain unnamed. Hudson River Trading's value is explicitly withheld as "multibillion-dollar" (claim 7609) without magnitude, preventing independent reconciliation. The weighted-average five-year contract term (claim 3561) implies backlog includes distant-year obligations, not near-term revenue certainty. Cumulatively, named customer commitments (Meta $21B + OpenAI $11.2B+ + Jane Street $6B + Anthropic unspecified + Hudson River unspecified) account for a material but indeterminate fraction of the $104.2 billion total, leaving substantial backlog unattributed to publicly disclosed counterparties.

CoreWeave has disclosed a material shift in customer concentration and commitment structure between mid-2025 and mid-2026. Through September 2025, the company's revenue narrative centered on two strategic relationships: OpenAI, contributing $11.2 billion to backlog (claim 4969), and a material customer contract financing the DDTL 3.0 facility (claim 4595). By March 2026, RPO had surged to $60.7 billion from $15.1 billion year-over-year (claim 3555), yet the company disclosed no single named agreement accounting for this 304% increase. The verbatim record suggests that expansion derived from existing MSAs: new order forms with OpenAI (September 2025, claim 4296), NVIDIA (September 2025, claim 4352), and Meta (September 2025, claim 4282 for $14.2 billion through 2031, expanded March 2026, claim 2123).

By August 2026, the concentration risk had visibly diversified. Meta emerged as a named $21 billion commitment through 2032 (claim 7628), Hudson River Trading signed a disclosed multibillion-dollar multiyear agreement (claim 7609), and Jane Street committed $6 billion to data center capacity (claim 7611). Microsoft remained identified only as "among the biggest" customers without deal specificity (claim 648). IBM's May 2025 "partnership" language (claim 4971) lacked contractual terms and revenue quantification; it remained a framework announcement rather than a binding capacity agreement.

Revenue arithmetic presents a reconciliation gap. Q2 2026 backlog reached $104.2 billion (claim 1302), with $25 billion in net new commitments added in early Q3 (claim 1303). Named deals—Meta $21 billion, Hudson River Trading multibillion (unquantified), Jane Street $6 billion, Anthropic multiyear (claim 7630, unquantified), and earlier OpenAI $11.2 billion—do not sum to disclosed totals, indicating either substantial undisclosed counterparties or higher aggregation from existing customers. The weighted-average contract duration of five years (claim 3584) suggests long-term lock-in; however, the gap between publicly named commitments and reported backlog magnitude implies conservative disclosure around existing customer expansions, obscuring true concentration metrics.

CoreWeave's customer disclosure exhibits a critical distinction between named signed agreements and strategic framework language that obscures the actual customer base. The company has publicly named five substantial customers with explicit order forms or financial commitments: OpenAI (claim 4296, September 2025), Meta (claims 4282, 2123, 7628—$14.2B through 2031, expanding to $21B through 2032), Hudson River Trading (claim 7598, multi-billion dollar agreement, August 2026), Jane Street (claim 7611, $6B), and NVIDIA (claim 4352, order form). By contrast, IBM appears only as a "partnership" to deliver compute capacity for Granite models (claim 4971), representing a supplier relationship rather than a customer commitment. Microsoft and OpenAI are mentioned as "among the biggest" clients (claim 648) and "major clients" (claim 618), yet CoreWeave has disclosed specific deal amounts only for OpenAI and not for Microsoft, suggesting material asymmetry in revenue concentration disclosure.

Counterparty concentration has intensified dramatically. Meta alone represents at least $35.2 billion in announced commitments across three separate order forms (September 2025, March 2026, August 2026), with an expansion option through 2032. When combined with the multi-billion-dollar Hudson River Trading deal and $6 billion from Jane Street, these three customers account for over $41 billion in disclosed commitments. Against a total revenue backlog of $104.2 billion as of June 2026 (claim 1302), this means roughly forty percent of backlog derives from three named customers, with OpenAI and Microsoft comprising an undisclosed but likely material portion of remaining obligations.

The revenue arithmetic reveals material gaps. The reported progression from $25.9 billion backlog (March 2025, claims 4970) to $60.7 billion remaining performance obligations (December 2025, claim 3555) to $104.2 billion (June 2026, claim 1302) represents a four-fold increase over fifteen months. Yet the disclosed deals—Meta $35.2B, Jane Street $6B, and Hudson River Trading's unnamed amount—do not reconcile with an $104 billion backlog without substantial undisclosed customer commitments. The August 2026 announcement of "$25 billion in net new customer commitments in early Q3" (claim 8283) further suggests material customers remain either unnamed or characterized only through partnership language, creating opacity around whether CoreWeave's growth reflects diversified customer acquisition or concentration risk masked by framework agreements with strategic suppliers.

CoreWeave's customer announcements reveal a sharp evolution from strategic framework language to binding contractual agreements. Through mid-2025, the company emphasized "strategic deals" with OpenAI and IBM—language that often masks LOI or MOU structures. OpenAI was cited (claim 4969) as a "$11.2 billion revenue backlog" partnership in May 2025, yet the actual binding order form wasn't announced until September 23, 2025 (claim 4296), suggesting a lag between commercial interest and executable contracts. IBM's Granite partnership (claim 4971) similarly employed partnership terminology without quantified delivery commitments or order form disclosure. NVIDIA's initial announcement used "collaboration framework" language (claim 3892), distinguishing it from the later signed order form (claim 4352).

Starting in September 2025, CoreWeave shifted to naming signed order forms explicitly. Meta's September 2025 commitment—$14.2 billion through December 2031 (claim 4282)—and subsequent March 2026 order form (claim 2123) introduced dated, binding language, followed by a higher $21 billion agreement announced August 2026 (claim 7628). The August 2026 announcements of Hudson River Trading's "multi-year, multi-billion-dollar agreement" (claim 7598), Anthropic's "multi-year deal" (claim 7630), Jane Street's "$6 billion" commitment (claim 7611), and Backblaze's "multi-year foundational data storage agreement" (claim 7645) all specified counterparty identity and term. MasterClass (claim 7782) lacks this precision, referring only to platform selection—closer to pilot than committed purchase.

Counterparty concentration remains heavily weighted: Meta and OpenAI dominate the backlog, with Microsoft described generally as "among the biggest" (claim 648) and 1241 without specific contract disclosure. Diversification to Hudson River Trading, Anthropic, Jane Street, and IBM suggests evolution toward a broader customer base, though three to four hyperscalers likely represent the majority of revenue. Revenue arithmetic shows consistency: $14.7 billion RPO in May 2025 (claim 4970) grew to $60.7 billion by December 2025 (claim 3555), then $104.2 billion by Q2 2026 (claim 1302), with an additional $25 billion in new commitments through early Q3 2026 (claim 1303). These figures align with disclosed agreements, though reliance on hyperscaler concentration introduces material customer concentration risk despite backlog scale.

CoreWeave has successfully named and executed binding contracts with several major counterparties, though concentration risk is acute and revenue reconciliation raises questions. The company has disclosed signed order forms under existing Master Service Agreements with OpenAI (claim 4296, September 2025), NVIDIA (claim 4352, September 2025), and Meta (claim 2123, March 2026), substantiated by quantified commitments: Meta $14.2 billion through December 2031 (claim 4282, September 2025) and a subsequent $21 billion commitment through 2032 (claim 7628, August 2026). Additional named customers with binding agreements include Hudson River Trading (multi-billion-dollar multiyear deal, claim 7598, August 2026), Anthropic (multi-year arrangement, claim 7630, August 2026), Jane Street ($6 billion capacity commitment, claim 7611, August 2026), and Backblaze ($335 million five-year storage agreement, claim 7448, August 2026). IBM and NVIDIA, however, appear in framework language—IBM as an announced "partnership" for Granite compute capacity (claim 4971) and NVIDIA transitioning from order forms to a "collaboration framework" to advance AI adoption globally (claim 3892, January 2026)—suggesting these may be softer commitments than true signed contracts.

The counterparty concentration has shifted dramatically toward hyperscalers and AI leaders. Meta and OpenAI together account for at least $50.4 billion of named commitments, while Microsoft appears only as a referenced "major client" (claims 648, 632, 618, 1100) without any disclosed signed deal, creating an asymmetry in transparency. This naming gap raises questions about whether Microsoft's contribution to backlog is material but undisclosed, or whether the characterization overstates actual commitments.

Revenue reconciliation presents arithmetic tension. CoreWeave disclosed $60.7 billion in remaining performance obligations as of December 31, 2025 (claim 3555), yet Meta alone committed $35.2 billion across two separately announced deals (September 2025 and August 2026). When adding OpenAI's $11.2 billion (claim 4969, May 2025), Jane Street's $6 billion, Hudson River Trading's unquantified but explicitly "multibillion-dollar" commitment, and Applied Digital's $11 billion prospective lease revenue (claim 371, August 2026), the disclosed named commitments exceed the reported RPO figure. This gap suggests either that some announcements represent framework options rather than fully committed RPO, that deals announced as separate tranches overlap the same underlying MSA, or that some "commitments" remain subject to conditions (like delivery and availability requirements mentioned in claim 4970). The August 2026 announcement of $25 billion in additional commitments secured in early Q3 (claims 1303, 3011, 368)—not yet reflected in the $104.2 billion Q2 backlog (claim 624)—further indicates that disclosure lags execution, complicating verification of actual binding exposure.

CoreWeave has disclosed a materially differentiated set of arrangements across its customer roster, mixing binding customer contracts with framework partnerships that require parsing. The majority of named, major customers—Meta, OpenAI, Hudson River Trading, Jane Street, Anthropic, Backblaze, and Microsoft—are confirmed through executed order forms or definitive statements of commitment. Meta's arrangements are particularly clear: a $14.2 billion contract through December 2031 (claim 4282, Sept 2025) followed by an expanded $21 billion agreement through 2032 (claim 7628, Aug 2026), both signed order forms. OpenAI similarly executed an order form in September 2025 (claim 4296) building on an existing master services agreement from April 2023. Hudson River Trading, Jane Street, and Anthropic all feature explicit "multi-year" or multi-billion-dollar commitment language consistent with signed deals (claims 7605, 7611, 7630).

IBM and NVIDIA, by contrast, are positioned as framework partnerships rather than disclosed customer revenue contracts. IBM's Granite compute capacity arrangement (claim 4971) reads as a strategic partnership announcement. NVIDIA's relationship has explicitly evolved into a "collaboration framework" (claim 3892, Jan 2026), suggesting alignment but not a binding customer contract with specified revenue.

Counterparty concentration has shifted dramatically. In May 2025, OpenAI represented 43% of disclosed revenue backlog at $11.2 billion of $25.9 billion total (claim 4969). By August 2026, OpenAI remains material but diversification is evident: the customer roster disclosed includes Meta, Microsoft, Jane Street, Hudson River Trading, Anthropic, Perplexity AI, Google, IBM, and others (claim 1241). Meta has emerged as or remains the largest single commitment.

Revenue arithmetic raises reconciliation questions. OpenAI's contract value is reported both as $22.4 billion as of September 2024 (claim 1242) and $11.2 billion in May 2025 (claim 4969)—suggesting either partial recognition, an interim reporting window, or expansion between those statements. Meta's expansion from $14.2B to $21B year-over-year confirms material growth in a core account. Total RPO quadrupled from $15.1 billion (end of 2024) to $60.7 billion (end of 2025, claim 3571), and revenue backlog reached $104.2 billion by Q2 2026 (claim 624), with an additional $25 billion in commitments signed in early Q3 (claim 1303). This growth is consistent with disclosed deal volume, though OpenAI contract value requires clearer reconciliation.

CoreWeave's disclosed customer announcements conflate binding customer contracts with softer partnership language, creating opacity around revenue concentration and contractual substance. Named customers with explicit order-form or master services agreement references include Meta (claims 4282, 2123), OpenAI (claim 4296), Hudson River Trading (claims 7602–7609), Anthropic (claim 7630), Backblaze (claim 7448), and Jane Street (claim 7611). By contrast, IBM's May 2025 announcement (claim 4971) uses "partnership" language without order-form documentation, and NVIDIA's January 2026 "collaboration framework" (claim 3892) explicitly eschews binding commercial terms.

Counterparty concentration has narrowed around two majors: Meta and OpenAI. Meta's disclosed commitment almost doubled from $14.2 billion through 2031 (claim 4282, September 2025) to $21 billion through 2032 (claim 7628, August 2026)—a 48-percent increase warranting explanation as either new scope or contract repricing, yet neither rationale is disclosed. OpenAI appears twice in the record: first as a "strategic deal" contributing $11.2 billion to revenue backlog in May 2025 (claim 4969), then as a signed order-form counterparty in September 2025 (claim 4296) with no updated dollar amount attached to the executed agreement. Microsoft's status as a "major client" (claims 618, 632, 648) lacks any signed agreement detail or financial disclosure.

The revenue backlog trajectory does not reconcile cleanly with disclosed signed deals. Remaining performance obligations grew from $15.1 billion (December 2024, claim 3572) to $60.7 billion (December 2025, claim 3571) to $104.2 billion (June 2026, claim 624), alongside unspecified $25 billion in early Q3 2026 commitments (claim 1303). Named-customer contracts with specific values—Meta ($21 billion), Jane Street ($6 billion), Backblaze ($335 million)—account for roughly $27 billion of the $104 billion backlog. Hudson River Trading is described as multibillion-dollar but lacks a disclosed amount, and OpenAI's executed contract carries no stated value. This leaves substantial unattributed backlog, suggesting either undisclosed customers or that earlier "strategic partnerships" have been reclassified into financial backlog without explicit customer identification.

CoreWeave's customer disclosure pattern reveals a sharp distinction between contractually signed agreements and strategic naming without valuation. Meta emerges as the longest-standing counterparty, with an explicit order form under an existing Master Services Agreement dated December 2023 (claim 2123), renewed on March 31, 2026. OpenAI represents the second material commitment: CoreWeave disclosed contracts worth $22.4 billion as of September 2025 (claim 1242), though this figure carries a historical date qualifier and no subsequent refresh. Hudson River Trading was introduced as a newly signed customer in late August 2026, described as a "multiyear, multibillion-dollar agreement" (claim 1809) with explicit chip-access provisions, but without quantified revenue attribution.

The remaining large-name roster—Anthropic, Google, IBM, Perplexity AI—appears in claim 1241 as members of a "diverse and expanding set of customers" without attached contract language or financial terms. Jane Street, IMC, and Flow Traders are noted as customers who "have already adopted" CoreWeave's infrastructure (claim 1792), equally bereft of specifics. This pattern—named without signed-agreement evidence—inflates the appearance of customer diversification while obscuring concentration. Rescale shifted from strategic narrative to explicit partnership in August 2025 (claims 1804–1806), with integration commitments but no revenue guarantee quantified. Applied Digital's 400 MW lease commitment totals $11 billion in "prospective lease revenue" (claim 371), language that defers actual cash collection, and Backblaze's five-year, $335 million contract (claim 1243) is the clearest single-named deal outside the three largest.

CoreWeave's $104.2 billion Q2 revenue backlog (claim 1810) stands as the critical reconciliation problem. Named customers with disclosed valuations—OpenAI ($22.4 billion), Applied Digital ($11 billion), Backblaze ($335 million)—sum to roughly $33.7 billion, leaving $70 billion unexplained by public counterparty identification. The backlog's composition reveals profound customer concentration among a narrow set (Meta, OpenAI, Microsoft) whose exact allocations within the $104.2 billion are undisclosed. The $25 billion in net new commitments claimed for early Q3 (claim 1811) adds opacity; no named customers are attributed to this acceleration.

CoreWeave's headline capacity has grown on two tracks, and the gap between them is widening. Active power rose from roughly 420 MW in May 2025 (4972, 29785) to 470 MW (25577), then 590 MW (20199), then 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). That follows the 70 MW and 360 MW year-end figures for 2023 and 2024 (3551, 3552). Contracted power grew faster, from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW (26367). The unenergized backlog went from about 1.2 GW to more than 3 GW. Contracting is outrunning energization, and the long-range targets keep rising. The 5+ GW NVIDIA-linked goal for 2030 (31932, 3893, 27948, 9329) was joined in May by an 8 GW active-power target (9322) with no bridging schedule. Reaching it would mean adding about 7 GW in under five years, against roughly 1 GW delivered to date.

The largest single capacity claim did not hold. The Core Scientific deal promised about 1.3 GW of owned power (28335), with closing guided to Q4 2025 (28334). It was terminated (19825, 4166) and reverted to a partnership (20313). After that, the company itself disclosed an OpenAI installation delay and Core Scientific weather and design setbacks (19828, 19829). The 1.3 GW never formally left the narrative. Contracted totals kept climbing as if nothing had been lost.

Independent corroboration is thin. Apart from one financing date (4593), every capacity and timeline claim on record is self-stated. The record contains no permits, interconnection-queue positions or site reporting for Kenilworth (25583), Lancaster (20201), Indonesia (26366) or the other announced campuses. None of these projects carries a capacity or cost figure at the project level, so the company-wide totals cannot be traced to specific sites.

Over the next 90 days, Kenilworth is the clearest test. Its initial phase is promised for 2026 (25584), the deadline falls within this window, and no update has followed. The Q3 report should show whether active power stayed on its roughly 150 to 250 MW per quarter pace. It should also show whether contracted power rises again without matching energization, and whether the Vera Rubin deployments (17621, 10341) reach named sites. The final record (27051) arrived truncated and is not assessed here.

CoreWeave's active footprint has grown on a steep and mostly consistent curve: 70 MW at end-2023 (3551), 360 MW at end-2024 (3552), roughly 420 MW in spring 2025 (4972, 29785), 470 MW by August (25577), 590 MW in November (20199), 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). Contracted power has climbed faster still, from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). The gap between contracted and energized capacity has therefore widened rather than closed, to roughly four to one. Long-range ambition has also escalated. A 5 GW-by-2030 framing tied to NVIDIA (3893, 27948) was joined in May 2026 by an 8 GW active-power target for the same year (9322). Reaching it implies sustained additions near 1.75 GW a year, well above anything delivered so far.

On dates, the largest revision was structural. The Core Scientific deal promised about 1.3 GW of owned gross power (28335) and a Q4 2025 close (28334), but it was terminated in October 2025 (4166, 19825) and recast as a continuing partnership (20313). Contracted-capacity figures kept rising afterward with little public reconciliation of what the lost owned megawatts meant. Weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828) were disclosed soon after management had characterized performance problems as isolated (19827). That sequence warrants scepticism toward later reassurances. Kenilworth's initial phase remains pinned to 2026 (25584), with about ten weeks left. Indonesia's 360 MW is dated to 2028 (26332).

Independent corroboration in this record is thin. Every named campus apart from Polaris Forge 1 sits at announced status with no capacity or cost fields populated. No permits, interconnection-queue positions or site reporting are attached for Kenilworth's 250 MW (25583) or Lancaster's 300 MW (20201). The supplied record also ends partway through the Indonesia entries. The announced pace rests almost entirely on company disclosures.

Over the next 90 days, watch the Q3 2026 report for active power meaningfully above the 1 GW mark and any change to the 4.2 GW contracted figure. Watch for confirmation or quiet deferral of Kenilworth energization before year-end. PJM queue or local permitting activity for Lancaster and Kenilworth would be the first external evidence. Look also for any reconciliation of the 8 GW target with current build rates.

CoreWeave's capacity story has two curves, and the gap between them is the central fact. Active power has risen steadily: roughly 420 MW in early 2025 (4972, 29785), 470 MW by August (25577), 590 MW in November (20199), 850 MW at year-end (11769, 3569), and past 1 GW by May 2026 (9321). That is a real ramp. The 10-K history of 70 MW in 2023 and 360 MW in 2024 (3551, 3552) shows annual additions growing from about 290 MW to about 490 MW. Contracted power has grown faster, from 1.6 GW (4973) through 2.2, 2.9, 3.1 and 3.5 GW (25578, 20200, 3554, 9323) to 4.2 GW (26367). Contracted capacity has stayed at roughly four times active capacity throughout. So the backlog is growing in step with delivery rather than being worked down. The 2030 framing has also grown, from 5 GW of NVIDIA-linked AI factories (3893, 27948) to an 8 GW active target (9322). The company has not reconciled the two figures. Reaching 8 GW from about 1 GW requires roughly 1.5 GW of energization a year, about three times the 2025 pace.

On dates, the largest revision was structural. The Core Scientific deal carried 1.3 GW of owned power and a Q4 2025 close (28335, 28334). It was terminated in October (4166), and the relationship continued only as a partnership (20313). Weeks later the company disclosed weather and design delays at Core Scientific sites and a delayed OpenAI installation (19829, 19828), while describing the problems as isolated (19827). Most project-level dates in the record are loose or missing. Kenilworth's initial phase is pegged to 2026 (25584), with no later reaffirmation on file. Lancaster has a 300 MW figure but no timeline (20201). Indonesia is the only project with a capacity and year attached: 360 MW online in 2028 (26331, 26332).

Independent corroboration is thin. Nearly every facility claim is company-stated. The only verified timeline item, the DDTL 3.0 termination date (4593), concerns financing, not construction. The record contains no permits, PJM interconnection filings for New Jersey or Pennsylvania, or site reporting that confirms the announced pace.

Over the next 90 days, the Q3 report is the key test. Watch whether active power has moved meaningfully beyond the 1 GW mark, whether Kenilworth's 2026 phase is reaffirmed or quietly moved, and whether the 8 GW target is restated alongside a contracted figure that has now passed 4.2 GW. Watch also for any update on delivery at the Core Scientific-hosted sites.

CoreWeave's contracted power has grown at almost every update: 1.6 GW in May 2025 (4973), 2.2 GW in August (25578), 2.9 GW in November (20200), 3.1 GW at year-end (11770, 3554), 3.5 GW in May 2026 (9323), and 4.2 GW by August 2026 (26367). Active capacity has also risen. The company reported 360 MW at end-2024 (3552), about 420 MW in spring 2025 (4972, 29785), 850 MW at end-2025 (3569), and more than 1 GW by May 2026 (9321). The gap between the two has not narrowed. Contracted power has stayed at roughly four times energized power, so the backlog is growing faster than the company can deliver it. The 2030 targets have also grown, from 5 GW of NVIDIA-linked AI factories (3893, 27948, 9329) to an 8 GW active-power target (9322). Getting from about 1 GW to 8 GW in four and a half years means adding roughly 1.5 GW a year. That is about three times the roughly 490 MW added in calendar 2025.

The one large step change in the record reversed. The Core Scientific merger would have brought about 1.3 GW of owned gross power (28335) and more than 1 GW of expansion potential (28336), with closing expected in Q4 2025 (28334). It was terminated on 2025-10-30 (4166, 19825). Later filings describe the relationship as a continuing partnership (20313) rather than a lost owned-capacity path, which leaves the capacity loss unaddressed. Within weeks, the record also shows weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828), which the company described as isolated (19827). Explicit site-level dates are rare. Kenilworth's 250 MW campus (25583) carries a 2026 initial-phase commitment (25584) with no later update on record. Indonesia's 360 MW is pegged to 2028 (26331, 26332). Lancaster's 300 MW (20201) has no date at all. Most listed campuses show no capacity or cost figures.

This record contains no independent corroboration: no permits, interconnection-queue positions, or site reporting. Every capacity figure is company-stated. Without outside evidence, the announced pace cannot be checked against physical progress. The claim list supplied here also ends partway through the August 2026 entries, so later revisions may be missing.

Over the next 90 days, watch three things. First, whether the Q3 2026 report shows active power growing meaningfully beyond 1 GW or whether contracted power alone keeps climbing. Second, whether Kenilworth's initial phase is energized by year-end or quietly moved. Third, whether utility filings in New Jersey, Pennsylvania, and New York begin to show the load these campuses would require.

CoreWeave's capacity story has two curves that are pulling apart. Active power has grown steadily, from roughly 420 MW in early 2025 (claims 4972, 29785) to about 470 MW (25577), then 590 MW (20199), 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). Contracted power has grown faster, from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323), and 4.2 GW by August 2026 (26367). The gap between what is energized and what is contracted has widened from about 1.2 GW to more than 3 GW. Ambition has also stepped up. The NVIDIA-linked goal of 5+ GW by 2030 (3893, 27948, 9329) was joined in May 2026 by an 8 GW active-power target for the same year (9322). That would require roughly an eightfold increase in energized capacity in under five years.

The one large shrinkage was Core Scientific. Its roughly 1.3 GW of owned power (28335) was presented as arriving by Q4 2025 (28334), but the merger was terminated (4166, 19825) and the relationship reverted to a partnership (20313). Later disclosures of weather and design delays at Core Scientific sites (19829) and a slipped OpenAI installation (19828), described by the company as isolated (19827), suggest some of that capacity is late even as leased supply. Neither delay came with a revised delivery date in the record, which is effectively a revision without acknowledgement.

Independent corroboration is thin. Nothing in this record ties Kenilworth (25583), Lancaster (20201), Indonesia (26331, 26332) or the other announced campuses to permits, interconnection queue positions, or site reporting. Most projects carry no capacity or cost figures at all. The reported active megawatts are company-stated, though their consistency across quarterly and annual filings (3551, 3552, 3553) lends them some weight. The forward pipeline is largely unverified. Note that the claim record supplied here is truncated after claim 26331, so later disclosures are not reflected.

Over the next 90 days, the Q3 2026 report should show whether active power has kept pace with the 4.2 GW contracted figure or whether the gap keeps widening. Kenilworth's initial phase is promised within 2026 (25584), so year-end is a hard test of whether any announced greenfield campus delivers on schedule. Readers should also look for any restated timing on the Core Scientific sites. Finally, any utility, zoning, or queue filings for Lancaster or Kenilworth would be the first external evidence behind the pipeline.

CoreWeave's capacity story has grown on two tracks at very different speeds. Contracted power rose from 1.6 GW (claim 4973) through 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323) to 4.2 GW (26367), roughly 2.6x in fifteen months. Active power also grew quickly, from about 420 MW (4972, 26764) to 850 MW at year-end 2025 (3553) and past 1 GW by May 2026 (9321). But the gap between contracted and energized capacity widened from about 1.2 GW to more than 3 GW. The 2030 ambition has escalated in the same way. The 5 GW NVIDIA-linked goal (3893, 27948, 9329) sat alongside an 8 GW active-power target within four months (9322). Hitting 8 GW implies a build rate several times what the company delivered in 2025.

On dates, the clearest reversal is Core Scientific. The deal was presented as adding roughly 1.3 GW of owned power plus 1 GW of expansion headroom (28335, 28336) with a Q4 2025 close (28334). It was terminated in October (19825, 4166), and only a continuing partnership was offered in its place (20313). Weeks later, the company acknowledged an OpenAI installation delay and weather and design delays at Core Scientific sites (19828, 19829), having earlier described data center underperformance as isolated (19827). The near-term site promise still in play is Kenilworth's initial phase in 2026 (25584), part of a campus of up to 250 MW (25583). The record shows no update on it since August 2025, with about ten weeks left to deliver.

Independent corroboration is thin. Apart from a financing date (4593), every capacity and timeline claim here is company-stated. None of the named projects, including Lancaster's 300 MW (20201), Kenilworth, Indonesia's 360 MW for 2028 (26331, 26332) and Polaris Forge 1, has a disclosed capacity, cost, permit trail or interconnection queue reference in the record. The reported active-power gains are consistent with real deployment. The contracted backlog, however, cannot yet be tied to specific energized sites.

Over the next 90 days, watch the Q3 2026 report for whether active power has moved meaningfully beyond the 1 GW mark and whether contracted power keeps rising faster than it. Watch for any acknowledgement, or silence, on Kenilworth's 2026 phase. Watch PJM queue and local permitting records for Kenilworth and Lancaster. And watch whether Vera Rubin deployments (17621, 10341) and the Meta commitment (17654) are attributed to named facilities rather than to aggregate totals.

CoreWeave's capacity story has two curves. Contracted power has climbed steadily: 1.6 GW in May 2025 (claim 4973), 2.2 GW by August (25578), 2.9 GW in November (20200), 3.1 GW at year-end (11770, 3554) and 3.5 GW by May 2026 (9323). Active power has also risen fast, from roughly 420 MW (4972, 29785) to 470 MW (25577), 590 MW (20199), 850 MW at December 31, 2025 (11769, 3569) and past 1 GW (9321). The 70 MW to 360 MW to 850 MW year-end series (3551, 3552, 3553) shows real energization, not just signed paper. The ceiling has moved the most. It went from a 5 GW NVIDIA-linked 2030 target (3893, 27948, 9329) to an 8 GW active-power goal for the same year (9322). That target sits about eight times above current active capacity, and the gap between contracted and active capacity is still roughly 2.5 GW. The one material shrinkage was never fully acknowledged as such. The Core Scientific deal would have added about 1.3 GW of owned power and 1 GW-plus of expansion (28335, 28336), with a Q4 2025 close (28334). It was terminated in October 2025 (19825, 4166) and reframed as a continuing partnership (20313). That capacity quietly left the narrative rather than being restated. Date discipline is mixed. Management called delivery problems isolated (19827), yet within weeks it disclosed weather and design delays at Core Scientific sites (19829) and an OpenAI installation slip (19828). Neither came with a revised in-service date on the record. Project-level evidence is thin. Every site except Polaris Forge 1 remains at announced status with no capacity or cost on its project record. That includes Kenilworth at 250 MW (25583), Lancaster at 300 MW (20201) and Indonesia at 360 MW by 2028 (27051, 27052). The record supplied here contains no permits, interconnection queue filings or site reporting, so the pace is supported only by company statements. The claim list is also truncated after the Indonesia entries. Over the next 90 days, the clearest test is Kenilworth, whose initial phase is promised within 2026 (25584). If it slips or arrives undescribed, that would echo the pattern seen in 2025. The next quarterly print should show whether active power is keeping pace with the 3.5 GW contracted book, and whether Vera Rubin bring-up (17621, 10341) turns into named-site capacity. Utility and PJM queue records for Lancaster and Kenilworth should be checked as the first independent corroboration.

CoreWeave's capacity story has expanded steadily, with no announced figure ever walked back. Contracted power went from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323) in about a year. Active power has also compounded. The company's own year-end series runs from 70 MW (3551) to 360 MW (3552) to 850 MW (3553), and it reported passing 1 GW by May 2026 (9321). The long-range ambition grew too. The 5+ GW of AI factories with NVIDIA by 2030 (3893, 27948, 9329) sat alongside a newer 8 GW active-power target for the same date (9322). Reaching 8 GW would take roughly 1.7 GW of net additions a year. The best year on record added about half a gigawatt.

The Core Scientific episode is the clearest case of a promised path that disappeared. The deal was framed as delivering about 1.3 GW of owned gross power plus 1 GW or more of expansion potential (28335, 28336), with a Q4 2025 close (28334). It was terminated in October 2025 (19825, 4166) and recast as a continuing partnership (20313). Contracted capacity kept rising afterward, which suggests the gap was filled through leases rather than ownership. The filings do not reconcile this explicitly.

Timeline candor has been partial. The company acknowledged weather and design delays at Core Scientific sites (19829) and an OpenAI installation slip (19828), but described the problems as isolated (19827). Most site-level claims carry no dated commitment. Kenilworth's 250 MW campus with initial-phase delivery in 2026 (25583, 25584) is the main near-term dated promise. Lancaster's 300 MW (20201) and Indonesia's 360 MW across three facilities by 2028 (27051, 27052) have longer horizons. Every named project except Polaris Forge 1 remains at announced status, with no capacity or cost fields populated.

The record contains no independent corroboration: no permits, no interconnection queue positions, no on-the-ground reporting. The announced pace therefore rests entirely on company disclosures. The claim list provided here also cuts off mid-entry after claim 26331, so later items may not be reflected in this read.

Over the next 90 days, the Q3 2026 results should show whether active power keeps growing at more than 150 MW per quarter and whether contracted capacity moves past 3.5 GW. Kenilworth's 2026 initial-phase deadline expires within the window, so energization, or a quiet re-dating, will be telling. Utility or PJM filings tied to Kenilworth and Lancaster would be the first external test of the stated schedule.

CoreWeave's headline capacity numbers have only moved upward, and the more useful question is how the gap between contracted and active power has behaved. Contracted power rose from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323) in about a year. Active power went from roughly 420 MW (4972, 29785) to 590 MW (20199), 850 MW at year-end 2025 (3569), and past 1 GW by May 2026 (9321). The year-end series of 70, 360 and 850 MW (3551, 3552, 3553) shows real energization, not just announcements. The ratio of contracted to active power has held near 3.5x, but the absolute undelivered backlog has roughly doubled, to about 2.5 GW. The 2030 ambition has also grown, from 5+ GW of NVIDIA-aligned AI factories (3893, 27948, 9329) to an 8 GW active-power target (9322). The filings do not explain how the two figures reconcile.

The Core Scientific episode is the clearest case of announced capacity disappearing. The 1.3 GW of owned power (28335) and the 1 GW+ expansion option (28336) depended on a deal expected to close in Q4 2025 (28334). That deal was terminated in October (19825, 4166), and the record contains only a statement that the partnership continues (20313). Contracted capacity kept rising afterward, which suggests the leasing pipeline absorbed the loss. The record does not show a capacity figure being restated downward. The December disclosures of weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828) came after earlier language describing performance problems as isolated (19827). That sequence weakens the credibility of that framing.

Very few dated, project-level commitments exist, so few have had a chance to slip. The independent record is thin. Every named project except Polaris Forge 1 sits at announced status with no capacity or cost fields. No permits, interconnection queue positions, or site reporting in this record corroborate Kenilworth (25583), Lancaster (20201), or the 360 MW Indonesia plan targeted for 2028 (27051, 27052). The supplied record also ends partway through the August 2026 entries, so later claims may be missing.

In the next 90 days, the Kenilworth initial phase promised for 2026 (25584) becomes testable. The next quarterly update should show whether active power is growing fast enough to keep the 8 GW target plausible. Watch also for PJM queue or county permitting filings for Lancaster and Kenilworth, and for evidence that the Vera Rubin deployments (10341, 17621) are landing at specific named sites rather than unnamed ones.

CoreWeave's reported active power has grown steadily and roughly as described. The company's own filings trace a path from 70 MW at end-2023 (3551) to 360 MW at end-2024 (3552) and 850 MW at end-2025 (3553, 11769), with active power passing 1 GW by May 2026 (9321). Contracted power has grown faster and has never fallen between disclosures: 1.6 GW (4973), 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323), and 4.2 GW by August 2026 (26367). Contracted power has stayed at roughly three and a half to five times active power, so the delivery backlog is now the main thing to track. The 2030 targets raise the bar further. Five gigawatts with NVIDIA (3893, 27948, 9329) and 8 GW active (9322) would require adding about seven gigawatts in under five years. The best year on record added roughly 490 MW.

The largest revision was never framed as one. The Core Scientific deal brought about 1.3 GW of owned power (28335) and more than 1 GW of expansion potential (28336) into the story, along with close dates of October 2025 (4517), Q4 2025 (28334) and April 2026 (4731). The merger was terminated in October 2025 (4166, 19825). Contracted totals kept rising afterward, and the record does not explain how that capacity was reclassified, apart from a continuing partnership (20313). Delays at the operating level are documented: an OpenAI installation delay (19828) and weather and design delays at Core Scientific sites (19829). Management had characterized the data center performance issues as isolated (19827) about a month before those delays were disclosed.

The record contains no independent corroboration, such as permits, interconnection queue positions or site reporting, for any named project. Project-level capacity and cost fields are blank. Kenilworth (25583), Lancaster (20201) and Indonesia (27051) rest only on company statements. The announced pace therefore cannot yet be checked against outside evidence.

Over the next 90 days, the clearest test is Kenilworth's initial phase, which was promised for 2026 (25584). The record shows no progress update with about eleven weeks remaining. The Q3 2026 results should show whether active power is keeping pace with the 4.2 GW contracted figure. Readers should also watch for any restatement of the 8 GW 2030 target, any status change at Polaris Forge 1, and any utility or county filings for Lancaster or Kenilworth that would give the first outside check on timing.

CoreWeave's capacity story has two curves, and the gap between them is the central fact. Active power has grown steadily: roughly 420 MW in spring 2025 (4972, 29785), 470 MW by August (25577), 590 MW in November (20199), 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). That is real, compounding delivery from a 70 MW base at end-2023 (3551). Contracted power, however, has grown much faster: 1.6 GW (4973), then 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323) and 4.2 GW (26367). Contracted capacity has added about 2.6 GW over the period while active capacity added about 600 MW, so the backlog of power not yet energized keeps widening. The 2030 framing has also moved. The original ambition was 5 GW-plus of AI factories with NVIDIA (3893, 27948, 9329). A separate 8 GW active-power target followed (9322), and the record does not explain how the two figures relate.

The one capacity line that shrank was inorganic. The Core Scientific deal promised about 1.3 GW of owned gross power plus 1 GW-plus of expansion potential (28335, 28336). Its close date was stated variously as October 2025, Q4 2025 and April 2026 (4517, 28334, 4731). The deal was then terminated (19825, 4166), leaving only a continuing partnership (20313). Within weeks, disclosures cited weather and design delays at Core Scientific sites and a delayed OpenAI installation (19829, 19828), and management characterized the issues as isolated (19827). Neither delay has a later claim on record confirming it was resolved.

Independent corroboration is thin. Every facility claim here is company-stated, and the project table carries no capacity or cost figures. No permits, interconnection-queue filings or site reporting are on record for Kenilworth (25583), Lancaster (20201) or Indonesia (26331, 26332). Polaris Forge 1 is the only project marked partially operational. The pace of active-power growth is credible because it shows up in audited year-end figures (3552, 3553), but the conversion of named campuses into energized megawatts is not yet visible at the project level.

Over the next 90 days, the key test is Kenilworth's promised 2026 initial-phase delivery (25584). With about ten weeks remaining, the absence of any follow-up claim is notable. The Q3 2026 report should show whether active power has moved meaningfully beyond 1 GW and whether the active-to-contracted ratio is stabilizing. It is also worth watching for any acknowledgement that the late-2025 delays affected 2026 timelines, and for any interconnection or permitting filings that would anchor Lancaster or Kenilworth outside company disclosures.

CoreWeave's headline capacity has grown steadily. Active power rose from roughly 420 MW (4972, 29785) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (11769, 3569) and past 1 GW by May 2026 (9321). That is real, compounding delivery. Contracted power has grown faster, from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323) and 4.2 GW (26367). The backlog of undelivered power is therefore widening in absolute terms even as energization speeds up. The long-range target has also moved. The 5 GW of AI factories with NVIDIA by 2030 (3893, 27948, 9329) was joined in May by an 8 GW active-power target for the same year (9322). The company did not explain how the two relate.

The largest discontinuity is Core Scientific. The roughly 1.3 GW of owned power (28335), plus 1 GW of expansion potential (28336), was presented as arriving through an acquisition expected to close in Q4 2025 (28334). The deal was terminated in October (19825, 4166). Within weeks, management described data center performance as isolated (19827). The December filings then cited an OpenAI installation delay (19828) and weather and design delays at Core Scientific sites (19829). Those are slips, but they were disclosed rather than quietly revised. Since then the contracted-power figures have recovered past the level the merger would have implied, apparently through leasing instead.

Project-level evidence is thin. Kenilworth (25583, 25584), Lancaster (20201), Indonesia (26332) and the other named sites all carry status=announced, with no capacity or cost fields populated. Only Polaris Forge 1 is marked partially operational. The record contains no permits, interconnection queue positions or site reporting that would independently confirm the announced pace at any named campus. The aggregate MW disclosures come from company filings, not third parties.

The next 90 days should test the near-dated commitments. Kenilworth's initial phase is due in 2026 (25584), so year-end will show whether it energizes or the date moves without comment. Q3 results should show whether active power has kept its roughly 150-250 MW per quarter cadence and whether contracted power has moved beyond 4.2 GW. Watch also whether the Vera Rubin multi-site deployment (10341, 17621) appears as energized capacity. For Indonesia's 360 MW for 2028 (26332, 27051), look for land or grid filings. The source record was truncated at claim 27051, so later claims may not be reflected here.

CoreWeave's capacity story has only gone up, and it has gone up faster in contracts than in delivered power. Contracted power rose from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW by May 2026 (9323). Active power moved from roughly 420 MW (4972, 26764) through 470 MW (25577) and 590 MW (20199) to 850 MW at year-end 2025 (3553) and past 1 GW (9321). Active power has stayed near 26 to 29 percent of contracted power. That ratio is steady, but in absolute terms more than 2.5 GW of contracted power is still not energized. The 2030 targets also changed. The NVIDIA-linked figure was 5+ GW of AI factories (3893, 27948, 9329). A separate 8 GW active-power target appeared later (9322) with no stated bridge between the two, and it implies roughly seven additional gigawatts in about four years. The largest single capacity item, about 1.3 GW of owned power plus 1 GW of expansion headroom from Core Scientific (28335, 28336), lost its ownership path when the merger was terminated (19825, 4166). The Q4 2025 close guidance (28334) did not hold. The relationship was recast as a continuing partnership (20313), and contracted totals kept rising afterward without any reconciliation of how much of that power is leased rather than owned. On timing, the record contains admitted slippage: an OpenAI installation delay (19828) and weather and design delays at Core Scientific sites (19829). Management framed these as isolated (19827), which is a characterization, not evidence. Site-level dates are sparse. Kenilworth's initial phase is due in 2026 (25584) against a 250 MW envelope (25583). Lancaster's 300 MW (20201) carries no date. Indonesia's 360 MW is set for 2028 (26332, 27051). Independent corroboration is absent from this record. There are no permit filings, interconnection-queue positions, utility agreements or on-the-ground reporting for any named site, and every project carries blank capacity and cost fields. The announced pace rests entirely on company statements. Over the next 90 days, the Q3 2026 print should show whether active power is compounding at the roughly 150 to 250 MW per quarter the 8 GW target requires. Kenilworth either delivers an initial phase by December 31 or the 2026 date is quietly moved. Watch also for any disclosure separating owned from leased capacity, and for New Jersey and Pennsylvania utility or zoning filings that would independently anchor the Kenilworth and Lancaster figures.

CoreWeave's capacity story has two curves, and they are diverging. Active power has climbed steadily from roughly 420 MW in March 2025 (26764, 4972) to about 470 MW (25577), then 590 MW (20199), 850 MW at year-end 2025 (11769, 3553), and past 1 GW by May 2026 (9321). That is a credible ramp, and the year-end figures are reported consistently across filings (3552, 3553, 3569). Contracted power has grown faster: 1.6 GW (4973), 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323), and now 4.2 GW (26367). The unbuilt backlog has roughly tripled, from about 1.2 GW to well over 3 GW. The long-range target has also inflated without much explanation, from 5 GW of AI factories with NVIDIA by 2030 (3893, 27948, 9329) to an 8 GW active-power goal for the same year (9322).

The largest single reversal was the Core Scientific deal. CoreWeave guided to a close in Q4 2025 (28334, 4517) and said the merger would bring about 1.3 GW of owned gross power (28335). The termination (19825, 4166) removed that capacity from the ownership narrative, although the commercial relationship continues (20313). Shortly afterward the company disclosed delays tied to Core Scientific weather and design issues and to OpenAI installation timing (19829, 19828), and characterized them as isolated (19827). That framing is untested. Few project-level dates exist to hold management to. Kenilworth's initial phase is due in 2026 (25584), and the Indonesia capacity is promised for 2028 (27052, 26332).

Independent corroboration is thin. Every facility claim in this record is company-stated. No permit, interconnection-queue, or site-level reporting confirms the 250 MW Kenilworth (25583), 300 MW Lancaster (20201), or 360 MW Indonesia (27051) figures. Of the nine named projects, only Polaris Forge 1 is listed as partially operational, and none carries disclosed capacity or cost.

Over the next 90 days, start with the Q3 2026 print. It should show whether active power grew materially beyond 1 GW and whether contracted capacity keeps outrunning it. Kenilworth's initial phase faces a 2026-12-31 deadline, and any silence or quiet re-dating there would be telling. Two other items are worth tracking: whether Vera Rubin deployments (17621, 10341) appear as energized capacity, and whether the 4.2 GW contracted figure is tied to specific sites with utility interconnection evidence. Watch too for any reconciliation of the 5 GW and 8 GW targets.

CoreWeave's headline capacity has only grown, and it has grown fast. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power tracked upward from roughly 420 MW (4972, 26764) through 590 MW (20199) and 850 MW at year-end 2025 (3553) to more than 1 GW by May 2026 (9321). The record shows no cuts. Active capacity is real and compounding, roughly a twelvefold rise from the 70 MW year-end 2023 baseline (3551). But contracted capacity has run ahead of it. The undelivered backlog widened from about 1.2 GW to more than 3 GW, and the long-range ambition moved from 5+ GW of NVIDIA-linked AI factories by 2030 (3893, 27948) to an 8 GW active-power target (9322) without any explanation of the bridge. One small inconsistency deserves reconciliation: the 360 MW year-end 2024 figure (3552) sits awkwardly against 420 MW reported for March 2025 (26764), which implies about 60 MW was energized in a single quarter.

The largest capacity reversal was the collapse of the Core Scientific deal. It would have added about 1.3 GW of owned gross power (28335), and its slated close (28334, 4517) gave way to termination (4166, 19825). Contracted totals kept rising regardless, so the company never visibly marked down its pipeline. Timeline integrity is weaker than the capacity figures suggest. Data center performance was described as isolated (19827) shortly before weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828) were disclosed. That sequence is a revision in practice, even if it was not presented as one.

The record offers almost no independent corroboration. Every facility claim is company-stated, no project carries a disclosed capacity or cost at the project level, and there are no permit, interconnection-queue, or site-reporting entries for Kenilworth, Lancaster (20201), Plano, or Niagara Falls. Indonesia's 360 MW targeted for 2028 (26331, 26332) is too far out to test.

Over the next 90 days, the critical test is Kenilworth's initial-phase delivery promised within 2026 (25584), against a 250 MW campus plan (25583). The third-quarter report should show whether active power growth keeps pace with the 4.2 GW contracted figure, or whether the backlog widens again. Investors should also watch for utility interconnection filings that would anchor the Lancaster and New Jersey sites in third-party evidence.

CoreWeave's headline capacity has only gone up. Contracted power rose from 1.6 GW (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3582), 3.5 GW (9323) and 4.2 GW (26367) over roughly fifteen months. Active power grew too, from about 420 MW (4972, 26764) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3553) and past 1 GW (9321). That active figure is a real build rate: about 490 MW was added in calendar 2025, compared with 290 MW in 2024 (3552, 3551). But contracted capacity grew faster still. The gap between contracted and energized power widened from about 1.2 GW to more than 3 GW, so most of what the company cites as capacity is backlog, not delivered infrastructure.

The long-dated targets have also grown. The NVIDIA framework promised more than 5 GW of AI factories by 2030 (3893, 27948), and by May 2026 there was a separate 8 GW active-power goal for the same year (9322). Meeting 8 GW means adding roughly 1.75 GW a year, more than three times the best year on record.

On dates, the clearest miss is Core Scientific. A Q4 2025 close (28334) and an October 2025 milestone (4517) gave way to termination (4166, 19825), which the company framed as a continuing partnership (20313). Six weeks later, Core Scientific weather and design delays (19829) and an OpenAI installation delay (19828) surfaced, after an earlier description of data-center problems as isolated (19827). The 1.3 GW owned and 1 GW expansion figures (28335, 28336) were never formally withdrawn. Contracted totals kept rising regardless, which suggests leased Core Scientific capacity is still counted.

Independent corroboration is thin. The record contains no permits, interconnection-queue filings or site reporting for Kenilworth (25583), Lancaster (20201), Indonesia (26331, 26332) or the other announced sites. Every project except a partially operational Polaris Forge 1 carries no recorded capacity or cost. The pace is supported mainly by the company's own disclosures.

Over the next 90 days, three items stand out. The first is whether Q3 reporting shows active power growing meaningfully beyond 1 GW or flattening. The second is whether Kenilworth's initial phase, promised for 2026 (25584), is energized by year-end or quietly moved. The third is whether the Core Scientific sites recover from the delays flagged in December, and whether utility or county filings begin to appear for Lancaster and Kenilworth to back the announced megawatts.

CoreWeave's contracted power figure has risen in nearly every disclosure: 1.6 GW in May 2025 (claim 4973), 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323), and 4.2 GW by August 2026 (26367). Active power has also grown, from roughly 420 MW (4972, 26764) through 590 MW (20199) and 850 MW at year-end 2025 (3553) to above 1 GW (9321). The two lines are diverging, though. The gap between contracted and energized capacity was about 1.2 GW in May 2025 and is now above 3 GW. Contracted capacity has become the headline number while delivery is falling further behind it. The roughly 1.3 GW the company expected to own through Core Scientific (28335) did not materialize when the merger was terminated (4166, 19825). That made the expected close dates in 4517, 28334 and 4731 moot, and the relationship reverted to a partnership (20313).

The long-range target has moved up rather than down. A 5+ GW AI-factory goal with NVIDIA (3893, 27948) was joined in May 2026 by an 8 GW active-power target for 2030 (9322). Hitting 8 GW would require adding about 1.75 GW a year. In 2025 the company added roughly 490 MW (3552, 3553). The record shows no reconciliation of that gap.

On dates, the most concrete project milestone is Kenilworth's initial phase in 2026, against a 250 MW campus (25583, 25584). No follow-up claim reaffirms or revises it, so the deadline is approaching without any public update. The December 2025 disclosures of weather and design delays at Core Scientific sites and a delayed OpenAI installation (19829, 19828) came only weeks after the company characterized data center issues as isolated (19827). That sequence deserves weight when judging current assurances. Lancaster at 300 MW (20201) and Indonesia at 360 MW for 2028 (26331, 26332) have announced capacities but no interim milestones.

The record contains no independent corroboration: no permits, no interconnection-queue entries, and no on-the-ground reporting for any named site. Capacity fields are blank across all nine projects. The announced pace therefore rests entirely on company statements.

The next 90 days have three useful checks. The third-quarter report will show whether active power is still compounding or whether contracted capacity keeps growing faster. Kenilworth must show energized initial capacity by year-end or be visibly rescheduled. Any county permit or utility filings for Kenilworth, Lancaster or Plano would be the first external evidence that the 2030 trajectory is grounded in physical progress.

CoreWeave's headline numbers have moved in only one direction. Active power rose from roughly 420 MW in March 2025 (4972, 26764) to 850 MW at year-end 2025 (11769, 3553) and past 1 GW by May 2026 (9321). Contracted power grew faster, from 1.6 GW (4973) to 3.5 GW (9323). So the backlog of power not yet energized widened from about 1.2 GW to about 2.5 GW, even as deliveries accelerated. The long-range target has also grown. It started as 5+ GW of AI factories with NVIDIA by 2030 (3893, 27948, 9329) and now sits alongside an 8 GW active-power goal (9322). Reaching 8 GW from roughly 1 GW in under five years would take about 1.4 GW a year. The best year on record, 2025, added about 490 MW (3552, 3553).

The date record is mixed. The Core Scientific deal promised 1.3 GW of owned gross power (28335) and a closing in Q4 2025 (28334, 4517). It was terminated on 2025-10-30 (19825, 4166) and recast as a continuing partnership (20313). Contracted power kept rising afterward, which suggests those megawatts moved from owned to leased rather than disappearing. That shift has economic consequences the topline numbers hide. In December 2025 the company disclosed weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828). Management called these isolated (19827), and there has been no later reconciliation of the affected capacity.

The weakest part of the record is independent verification. The file contains no permit filings, interconnection-queue entries, or site reporting for any named project. Six of the nine projects carry no capacity figure at all. Kenilworth's 250 MW (25583), Lancaster's 300 MW (20201) and Indonesia's 360 MW (26331, 27051) rest entirely on company statements. The supplied record also cuts off mid-entry, so any later claims could not be assessed.

Over the next 90 days, the first test is the Q3 2026 report. It should show whether active power grew enough beyond 1 GW to keep pace with the 8 GW path, and whether contracted power grew again or flattened. The second is Kenilworth: its initial phase is promised by 2026-12-31 (25584), the first site-level date due to come due, with no corroborating construction evidence on file. Finally, look for PJM or NYISO queue positions and local permits for Lancaster, Kenilworth and Niagara Falls. Those would be the first outside check on whether the announced pace is real.

CoreWeave's capacity numbers have risen steadily, but contracted power is outrunning energized power. Contracted capacity went from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power rose more slowly but consistently: about 420 MW (4972, 26764), then 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3553), and past 1 GW by May 2026 (9321). The long-range target moved fast. A January 2026 goal of 5+ GW of NVIDIA-linked AI factories by 2030 (3893, 27948) became an 8 GW active-power target four months later (9322). With roughly 1 GW energized and 4.2 GW contracted, that target needs about an eightfold build-out in four years. Today's contracted backlog covers only about half of it.

The largest revision was never presented as one. The roughly 1.3 GW of owned power plus 1 GW of expansion tied to Core Scientific (28335, 28336) depended on a merger expected to close in Q4 2025 (28334). The merger was terminated in October 2025 (19825, 4166). Contracted capacity kept rising afterward, so the lost ownership seems to have been absorbed through the continuing commercial relationship (20313), not written down. CoreWeave did disclose delays. In November 2025 it described data-center performance problems as isolated (19827). Within weeks it cited OpenAI installation delays and Core Scientific weather and design setbacks (19828, 19829). Few site-level dates exist that could slip. Kenilworth's initial phase is due in 2026 (25584). Indonesia's 360 MW is due online in 2028 (27052, 26332). Lancaster's 300 MW (20201) has no stated timeline. Of the nine projects on record, only Polaris Forge 1 shows any operational status, and most list no capacity or cost.

The record contains no independent corroboration: no permits, no interconnection queue positions, no on-the-ground construction reporting. The pace of energization is supported only by the company's own disclosures. Kenilworth and Lancaster are both in PJM territory, where queue filings are public and the absence of matching entries would be informative.

Over the next 90 days, watch whether Kenilworth's 2026 initial phase is delivered, restated or quietly moved. Watch whether the third-quarter update shows active power growing at a pace consistent with 8 GW by 2030, and whether contracted power keeps rising beyond 4.2 GW. Watch for any timeline attached to Lancaster. And watch whether the Vera Rubin rollouts (10341, 17621) turn into named, energized sites rather than portfolio-level statements.

CoreWeave's contracted power has grown in every disclosure. It stood at 1.6 GW in May 2025 (4973), then 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW by May 2026 (9323). Active power followed the same direction, rising from roughly 420 MW (4972, 26764) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3553) and above 1 GW in May 2026 (9321). Active power has stayed at roughly 20 to 30 percent of contracted power throughout. That means the backlog of undeployed capacity has grown in absolute terms, from about 1.2 GW to about 2.5 GW, even as delivery sped up. The long-range target has also moved. The NVIDIA framing of 5+ GW of AI factories by 2030 (3893, 27948, 9329) was joined in May 2026 by an 8 GW active-power target for the same year (9322). We found no explanation in the record for how the two figures relate.

The clearest revision without much acknowledgement involves Core Scientific. About 1.3 GW of owned power through that acquisition (28335) was presented as part of the scale story, with closing expected in Q4 2025 (28334). The deal was terminated in October 2025 (19825, 4166) and recast as a continuing partnership (20313). Within weeks the record showed weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828). Contracted capacity kept rising after the termination, which suggests replacement sourcing, but the record does not show how the lost owned capacity was backfilled.

At the site level, the disclosures are thin. Most named projects, including Niagara Falls, Digital Drive, Sweden, Plano and Taloja, carry no capacity or cost figures. Kenilworth (25583) and Lancaster (20201) have megawatt figures but no follow-up progress claims. Indonesia's 360 MW (26331, 27051) is dated to 2028 (27052, 26332). None of these claims is backed in the record by permits, interconnection queue positions or on-the-ground reporting, so the announced pace rests entirely on company statements. The only verified item is a financing milestone (4593), not a facility milestone. Note that the supplied record ends partway through the claim list, so later entries may change this picture.

Three things matter over the next 90 days. First, Q3 results should show whether active power keeps pace with the 8 GW trajectory, which implies roughly 1.4 GW of additions per year from here. Second, Kenilworth's initial phase is due by year-end 2026 (25584), and this is the first site-level date in the record to fall due; independent permitting or utility evidence should be visible if it is on track. Third, any update on the Core Scientific delays will show whether the hosted capacity CoreWeave relies on is converting to active power.

CoreWeave's headline capacity has grown fast and without reversal. Contracted power rose from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power also climbed, from roughly 420 MW (4972, 26764) to 850 MW at year-end 2025 (3553) and past 1 GW in May 2026 (9321). The scale of ambition grew too: the 2030 goal moved from 5+ GW of NVIDIA-linked AI factories (3893, 27948) to an 8 GW active-power target (9322). The one real contraction was the Core Scientific deal. Its roughly 1.3 GW of owned power (28335) left the plan when the merger was terminated (4166, 19825), yet the contracted totals kept rising. That suggests the contracted figure reflects leases and partner capacity more than owned assets. The gap between contracted and active power has also widened in absolute terms, from about 1.2 GW to more than 3 GW, so delivery is falling further behind signings in megawatt terms even as active capacity grows.

The record of dated promises is thin and mixed. The Core Scientific close slipped from a Q4 2025 target (28334) to termination. In December 2025 the company acknowledged an OpenAI installation delay and weather and design delays at Core Scientific sites (19828, 19829), shortly after describing data center performance problems as isolated (19827). The most important hard date still open is Kenilworth's initial phase in 2026 (25584), against a stated capacity of up to 250 MW (25583). No progress claim has followed it. Lancaster's 300 MW (20201) and the 2028 date for Indonesia (26332, 27052) have no interim milestones on record.

Independent corroboration is essentially absent from this record. Every capacity and timeline claim is company-stated. The only verified item is a financing date (4593). The project ledger lists most sites as announced with no capacity or cost fields, and Polaris Forge 1 is the only one marked partially operational. No permits, interconnection-queue entries or site reporting support the claimed build pace, so the rate of active-power additions is the only usable proxy.

Over the next 90 days, three things matter most. First, whether the Q3 2026 update shows active power rising in step with the 4.2 GW contracted figure. Second, whether Kenilworth energizes any capacity by year-end or its date shifts without comment. Third, whether Lancaster, Taloja, Sweden or Plano gain disclosed megawatts, utility agreements or permits. A filing that quantifies the March 2026 portfolio (156) would also help.

CoreWeave's capacity story has two curves, and the distance between them is the main thing to track. Contracted power has risen steadily, from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3582), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power has also grown quickly: roughly 420 MW in early 2025 (4972, 26764), 590 MW in November (20199), 850 MW at year-end (3553, 11769) and more than 1 GW by May 2026 (9321). In relative terms, energized capacity has held at about a quarter of contracted capacity. In absolute terms the backlog not yet energized has grown from about 1.2 GW to more than 3 GW. The 2030 targets have also moved. The 5+ GW of AI factories with NVIDIA (3893, 27948, 9329) was joined in May 2026 by an 8 GW active-power target (9322), and no bridging explanation is on record. The largest contraction came from Core Scientific. The deal was marketed with about 1.3 GW owned plus 1 GW of expansion potential (28335, 28336), and then terminated (4166, 19825). It was recast as a continuing partnership (20313), with no restated view of how much of that capacity CoreWeave still expects to control.

On timing, the record shows slippage. In November 2025, data-center performance problems were described as isolated (19827). In December came disclosures of weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828). No revised delivery schedule followed either disclosure. Site-level commitments are thin. Kenilworth's initial phase is promised for 2026 (25584), Lancaster has 300 MW with no date (20201), and Indonesia has 360 MW targeted for 2028 (26331, 26332, 27052). Polaris Forge 1 is listed as partially operational without any capacity claim attached.

There is no independent evidence in this record. It contains no permits, no interconnection-queue entries and no on-the-ground reporting, and every capacity figure is self-reported at the company level. The announced pace is therefore neither corroborated nor contradicted. It is simply unverified, and project rows carry no capacity or cost data.

Over the next 90 days, Kenilworth's 2026 initial-phase promise comes due. Any quiet move into 2027 would be the first measurable site-level miss. The Q3 2026 update should show whether active power is keeping up with the 4.2 GW contracted figure or whether the gap widens further. Watch also for whether the 8 GW target is reiterated, reconciled with the NVIDIA figure, or dropped, and for any third-party permitting or utility filings at Kenilworth, Lancaster or the Indonesian sites that would confirm construction is actually under way.

CoreWeave's capacity narrative has grown every quarter. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power has also risen, from roughly 420 MW (4972, 26764) to 590 MW (20199), 850 MW at year-end 2025 (3553), and past 1 GW by May 2026 (9321). Energized capacity has stayed at roughly a fifth to under a third of contracted capacity, so the absolute gap has widened from about 1.2 GW to well over 2.5 GW. The long-range ambition has also grown, from 5+ GW of AI factories with NVIDIA by 2030 (3893, 27948) to an 8 GW active-power target (9322). The company's own year-end figures show additions of about 290 MW in 2024 and 490 MW in 2025 (3552, 3553). Reaching 8 GW would require roughly 1.5 GW a year from here, about three times the best year on record.

On dates, the clearest miss is corporate rather than physical. The Core Scientific merger, guided for completion (4517, 4731), was terminated (4166, 19825) and recast as a continuing partnership (20313). Late-2025 disclosures cite weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828), while the company described its data center performance as isolated (19827). No revised delivery date accompanied those disclosures. Most site-level promises carry no dates at all. Lancaster's 300 MW (20201) and the Kenilworth 250 MW campus (25583) are capacity statements, and only Kenilworth carries a firm initial-phase target of 2026 (25584). Indonesia's 360 MW across three facilities is pinned to 2028 (26331, 27052, 8758).

The record contains no independent corroboration. There are no permits, utility interconnection filings or site-level construction reporting tying any named project to the announced pace. Project-level capacity and cost fields remain blank for every site on file. The aggregate figures are self-reported and largely unverifiable.

Over the next 90 days, the first test is Kenilworth: whether the initial phase energizes by December 31 or the 2026 target is quietly dropped. The second is the Q3 2026 active-power figure, which needs to show additions well above the historical run-rate to keep the 8 GW target credible. The third is whether the next contracted-power number keeps rising while active power lags. Investors should also look for New Jersey or PJM interconnection records and Pennsylvania permitting activity at Lancaster, and for evidence that the Vera Rubin multi-site deployments (10341, 17621) are landing on new sites rather than relabeled existing capacity.

CoreWeave's contracted power has risen at every disclosure: 1.6 GW in May 2025 (4973), 2.2 GW by August (25578), 2.9 GW in November (20200), 3.1 GW at fiscal year-end (11770, 3554), 3.5 GW in May 2026 (9323), and 4.2 GW by August 2026 (26367). Active power has also grown, from roughly 420 MW (4972, 26764) to 850 MW at December 2025 (3569) and above 1 GW by May 2026 (9321). The two lines are diverging, though. The unenergized backlog has grown from about 1.2 GW to more than 3 GW, so new contracts are arriving faster than capacity is being switched on. The 2030 targets also do not agree with each other. The NVIDIA framing says 5+ GW of AI factories (3893, 27948, 9329), while a separate 8 GW active-power target (9322) appeared in May 2026 without a reconciliation. Reaching 8 GW from roughly 1 GW would take about 1.75 GW of new active capacity a year, around three and a half times the roughly 490 MW added in 2025 (3552, 3553).

Date discipline is mixed. The Core Scientific merger had targeted closings of October 2025 (4517) and April 2026 (4731). It was terminated instead (4166, 19825), with a continuing partnership reported (20313). In December 2025 the company acknowledged weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828), and it had already called the underperformance isolated (19827). The Kenilworth initial phase is still pinned to 2026 (25584), with no interim update on record and less than three months left. The DDTL 3.0 commitment termination date of July 31, 2026 (4593) has passed, and the record does not show how it resolved.

Independent corroboration is the weakest part of the record. Every capacity and timeline claim here is company-stated. Most named projects, including Kenilworth (25583), Lancaster (20201) and Indonesia (26331, 27051), carry no capacity or cost entries in the project register and remain at announced status. No permits, interconnection queue positions or site reporting appear that would independently support the pace implied by the 250 MW, 300 MW and 360 MW figures.

Over the next 90 days, the third-quarter report should show whether active power has moved meaningfully past 1 GW and whether the 5 GW and 8 GW targets are reconciled. It should also say whether Kenilworth will deliver in calendar 2026 or quietly move to 2027. Signs that Lancaster and Indonesia are moving from contract to construction would help, especially state permits, utility interconnection filings and named Indonesian sites behind the 2028 date (26332). The status of the DDTL 3.0 commitment should also be disclosed. The supplied record ends partway through early August 2026, so any later disclosures are not reflected here.

CoreWeave's capacity story has two curves, and they are diverging. Active power rose from roughly 420 MW in March 2025 (4972, 26764) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (11769, 3553), and past 1 GW by May 2026 (9321). That is real growth of about 600 MW in fourteen months. Contracted power grew faster: 1.6 GW (4973), 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323), and 4.2 GW by August 2026 (26367). The backlog of contracted but undelivered megawatts has widened in absolute terms with every disclosure. The 2030 ambition also escalated within four months, from more than 5 GW of NVIDIA-aligned AI factories (3893, 27948, 9329) to an 8 GW active-power target (9322). Hitting 8 GW would require adding about 1.5 GW a year, more than double the pace demonstrated so far.

Project-level dates are sparse, which makes slippage hard to pin down. The clearest test is Kenilworth, where the company promised an initial phase of a 250 MW campus in 2026 (25584, 25583). With under three months left in the year, the record contains no update. Lancaster's 300 MW (20201) carries no date at all. Indonesia's 360 MW is pegged to 2028 (26332, 27052). The one acknowledged delay pattern is instructive. In November 2025 data center performance problems were characterized as isolated (19827), but within five weeks the company disclosed an OpenAI installation delay and weather and design delays at Core Scientific sites (19828, 19829). This followed the termination of the Core Scientific merger (19825, 4166), which left those sites as lease relationships (20313) rather than owned capacity. Six named projects remain at announced status with no capacity figure attached.

The record contains no independent corroboration: no permits, no interconnection queue entries, and no on-the-ground reporting for any site. Every figure here is company-stated, so the announced pace is unverified rather than contradicted.

Over the next 90 days, watch four things. First, whether the Q3 2026 disclosure shows active power growth keeping up with the 1.5 GW-per-year run rate that the 8 GW target implies. Second, whether Kenilworth's initial phase is energized by December 31 or quietly re-dated. Third, whether the 4.2 GW contracted figure is broken down to show how much is Indonesia versus domestic sites. Fourth, whether Vera Rubin deployments (10341, 17621) appear at named facilities rather than as generic multi-location claims. Note that the source list ends mid-entry at claim 8760, so later August-onward claims were not assessed.

CoreWeave's capacity story has grown in one direction only, but contracted capacity is growing much faster than delivered capacity. Active power rose from roughly 420 MW in March 2025 (4972, 26764) through 470 MW (25577) and 590 MW (20199) to 850 MW at year-end 2025 (11769, 3569), then passed 1 GW by May 2026 (9321). Contracted power grew faster, from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). The gap between signed and energized capacity has widened from about 1.2 GW to more than 3 GW. The long-range target has also grown. The NVIDIA-linked 5+ GW by 2030 framing (3893, 9329) now sits alongside an 8 GW active-power goal for the same year (9322), and the company has not explained how the two figures relate.

The company's record on dates is mixed and only partly acknowledged. The Core Scientific merger carried target dates of October 2025 and April 2026 (4517, 4731), then was terminated (19825, 4166) and recast as a continuing partnership (20313). In December 2025 the company disclosed weather and design delays at Core Scientific sites and an installation delay on OpenAI infrastructure (19829, 19828), after describing data-center performance problems as isolated the month before (19827). Site-level promises are sparse. Kenilworth's 250 MW campus is the only US project with a dated near-term deliverable, an initial phase in 2026 (25583, 25584). Lancaster's 300 MW has no date (20201). Indonesia's 360 MW is pinned to 2028 (26332, 8758).

The record contains no independent corroboration. None of the nine listed projects has a recorded capacity or cost field. No permits, interconnection-queue filings or on-the-ground reports are attached. Polaris Forge 1 is marked partially operational but has no claims behind it. The pace therefore rests entirely on company disclosures, and those disclosures report portfolio totals rather than site-by-site energization.

Over the next 90 days, the Q3 2026 report should show whether active power keeps pace with the 4.2 GW contracted figure or whether the gap keeps widening. Kenilworth's 2026 deadline expires within this window, so watch whether an energized first phase is reported or the timing is quietly moved to 2027. Utility, zoning or interconnection filings for Kenilworth, Lancaster and the Jakarta sites would be the first outside test of the stated pace. The company should also clarify whether 8 GW active by 2030 supersedes or adds to the 5 GW NVIDIA program.

CoreWeave's contracted power figure has only gone up. It rose from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3582), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power rose with it, from 420 MW (4972, 26764) through 590 MW (20199) and 850 MW at year-end 2025 (3553) to more than 1 GW (9321). The ratio of contracted to active capacity has stayed at about four to one, so energization is keeping pace with signing rather than closing the gap. The long-range target has also grown. A 5+ GW NVIDIA-linked ambition for 2030 (3893) gained an 8 GW active-power target in May 2026 (9322), published next to a restated 5 GW goal (9329) with no explanation of how the two relate. Reaching 8 GW from roughly 1 GW in under five years would mean a much steeper build curve than the 2024-2025 increase of about 490 MW (3552, 3553).

On timing, the clearest break is the Core Scientific merger. Successive filings repeated its terms (4731, 4331, 4294) and it then ended in termination (19825, 4166), downgraded to a continuing partnership (20313). Delays were disclosed gradually. In November 2025 data center underperformance was described as isolated (19827). A month later the company acknowledged weather and design delays at Core Scientific sites (19829) and a slipped OpenAI installation (19828). That sequence suggests early characterizations of execution risk have run optimistic. Most named projects carry no capacity, date or cost on record. These include Niagara Falls, Digital Drive, Sweden, Taloja and Plano. Only Kenilworth (25583, 25584), Lancaster (20201) and Indonesia (27051, 27052) have concrete figures.

No independent evidence is in the record: no permits, interconnection queue positions or site reporting. The announced pace therefore rests entirely on company statements and cannot yet be corroborated. The portfolio snapshot dated March 31, 2026 (156) carries no figure, which limits reconciliation.

The next 90 days hold a hard test. Kenilworth's initial phase is promised within 2026 (25584), so either energization or an acknowledged slip should appear by early January. The Q3 report should show whether active power has moved meaningfully past 1 GW. Watch also for whether contracted capacity keeps climbing while active capacity stalls. Finally, watch for any reconciliation of the 5 GW and 8 GW targets, and for disclosure of how much of the DDTL 3.0 facility was drawn before its July 2026 commitment termination (4593).

CoreWeave's headline numbers have grown steadily and in one direction. Contracted power rose from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323), and most recently 4.2 GW (26367). Active power also rose, from about 420 MW (4972, 26764) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3553), and above 1 GW by May 2026 (9321). The year-end series of 70, 360 and 850 MW (3551 to 3553) shows real energization, roughly doubling each year. The more useful signal is the ratio. Contracted capacity has stayed near four times active capacity throughout, so the company is signing new capacity about as fast as it brings old capacity online. The backlog of capacity still to be delivered is not shrinking. The 2030 targets, 8 GW active (9322) and 5+ GW of NVIDIA AI factories (3893, 9329), would require adding roughly 1.75 GW a year from a base just over 1 GW. That is a step change from the pace shown so far, not a continuation of it.

On timing, the record is mixed. The Core Scientific merger was to close by October 2025 (4517) or April 2026 (4731). It was instead terminated (4166, 19825), and only a commercial relationship remains (20313). In November 2025 the company described data center performance problems as isolated (19827). The following month it disclosed an OpenAI installation delay (19828) and weather and design slippage at Core Scientific sites (19829). That sequence is the clearest instance of guidance being overtaken by events without a prior revision. Most named projects carry no capacity or date at all. Niagara Falls, Digital Drive, Sweden, Taloja and Plano have no quantified claims, and Lancaster's 300 MW (20201) has no schedule.

The record contains no independent corroboration of the announced pace, such as permits, interconnection queue positions or site reporting. Indonesia's 360 MW for 2028 (26332, 27052) is leased from a colocation partner (8760). Delivery there therefore depends on a third party's construction, which CoreWeave does not directly control.

The next 90 days offer a concrete test. Kenilworth's initial phase was promised for 2026 (25584), the deadline falls on December 31, and nothing has been said since August 2025. The Q3 report should show whether active power grew materially beyond 1 GW, and whether contracted capacity rose again without a matching rise in active power. It is also worth checking whether the Vera Rubin deployments (10341, 17621) are counted as new megawatts or as hardware refreshes inside existing halls.

CoreWeave's headline capacity has grown steadily in each reporting cycle. Contracted power rose from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active power has also risen, from roughly 420 MW (4972, 26764) through 590 MW (20199) and 850 MW at year-end 2025 (3569) to more than 1 GW in May 2026 (9321). The ratio between contracted and energized capacity has not narrowed, however. It was about 3.8 to 1 at the start of the record and is about 4 to 1 now, so new contracts are arriving at least as fast as power is being energized. The long-range target has also moved without a clear reconciliation. A 5-plus GW of AI factories with NVIDIA by 2030 (3893, 9329) sits alongside a separate 8 GW active-power target (9322), and the filings do not explain how the two relate. On dates, the record is thin because most facility claims carry no delivery deadline. The ones that slipped did so visibly. The Core Scientific merger, with target closes of October 2025 (4517) and April 2026 (4731), was terminated in October 2025 (4166, 19825). In December 2025 the company disclosed an OpenAI installation delay and weather and design delays at Core Scientific sites (19828, 19829), shortly after calling a data center performance issue isolated (19827). The DDTL 3.0 commitment termination date of July 2026 (4593) has passed with no recorded follow-up. Independent corroboration is essentially absent. Every named project except the third-party Polaris Forge 1 is still listed as announced and has no recorded capacity or cost at the project level. No permits, interconnection queue positions or site reporting support the 250 MW Kenilworth campus (25583), the 300 MW Lancaster facility (20201) or the 360 MW across three Indonesian sites for 2028 (26331, 26332, 8758). The aggregate figures are the company's own statements. Over the next 90 days, the most important test is the Q3 2026 report. It should show whether active power has moved meaningfully above 1 GW and whether contracted capacity keeps climbing faster than delivered capacity. Kenilworth's initial phase is due by year-end 2026 (25584), so evidence of energization, or a quiet revision of that date, should appear before the window closes. PJM queue or utility filings tied to Kenilworth and Lancaster, and named operating partners for the Indonesian sites, would be the first independent signs that the stated pace is real.

CoreWeave's capacity record follows two curves that are pulling apart. Active power has grown steadily, from roughly 420 MW in spring 2025 (4972, 26764) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (11769, 3569) and past 1 GW by May 2026 (9321). Contracted power has grown faster, from 1.6 GW (4973) through 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323) to 4.2 GW by August 2026 (26367). The undelivered backlog has therefore roughly doubled, from about 1.2 GW to well over 3 GW. The long-range target has also moved upward. A 5-plus GW AI-factory goal with NVIDIA by 2030 (3893, 9329) was joined in May 2026 by an 8 GW active-power target for the same year (9322). That second target implies a far steeper build rate than the roughly 600 MW added during 2025.

On dates, the record holds very few dated, project-level commitments. That scarcity is itself a finding: none of the nine listed projects carries a capacity or cost figure at the project level. The one near-term promise is the Kenilworth initial phase in 2026 (25584), against a 250 MW campus (25583). It is still marked announced, and the record shows no reaffirmation since August 2025. Delivery friction surfaced in late 2025. Management described a performance issue as isolated (19827), and a month later weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828) were on record. That sequence deserves scrutiny, given that the Core Scientific merger collapsed (19825, 4166) while the partnership continued (20313). Lancaster's 300 MW (20201) has no stated timeline. Indonesia's 360 MW is consistently pegged to 2028 (26331, 26332, 155).

The record contains no independent evidence, such as permits, interconnection queue positions or site reporting, for any project, so the announced pace cannot be corroborated here. The input was also truncated after claim 155, so later claims may be missing.

Over the next 90 days, watch four things. First, the Q3 2026 report, to see whether active power growth keeps pace with the 4.2 GW contracted figure. Second, any Kenilworth energization or quiet re-dating before year-end. Third, the status of DDTL 3.0 after its July 2026 commitment termination date (4593). Fourth, county or utility filings for Lancaster and Kenilworth that would show physical progress.

CoreWeave's active footprint has grown steadily and its contracted backlog has grown faster. Active power went from 70 MW at end-2023 (3551) to 360 MW at end-2024 (3552), then 420 MW (4972), 470 MW (25577), 590 MW (20199) and 850 MW at end-2025 (3553). The company reported passing 1 GW in May 2026 (9321). Contracted power climbed from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (3554) and 3.5 GW (9323). The gap between contracted and energized capacity has therefore widened from roughly 1.2 GW to about 2.5 GW, so backlog is outrunning delivery. The 2030 ambition has also shifted. Earlier filings framed it as 5+ GW of AI factories with NVIDIA (3893, 17595). The May 2026 disclosure restates that figure (9329) but adds an 8 GW active-power target in the same release (9322), and the record does not explain how the two numbers relate. Hitting 8 GW would mean adding roughly 1.75 GW a year, about three times the 2025 build rate. On dates, the clearest miss is the Core Scientific merger. It was expected to close by October 2025 (4517), was flagged as delayed, and was terminated on 2025-10-30 (4166, 19825), with the companies saying the commercial relationship continues (20313). Six weeks later management disclosed weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828), after describing data center performance as isolated (19827). Site-level commitments are thin. Kenilworth has an initial-phase delivery promised in 2026 (25584) at up to 250 MW (25583), and Lancaster has 300 MW (20201) with no date. Seven other announced projects, including Niagara Falls, Polaris Forge 1, Plano and Taloja, carry no capacity or cost in the record. The record contains no independent evidence, such as permits, PJM or ERCOT interconnection queue entries, or on-site reporting, so the announced pace rests entirely on company statements. The input also ends mid-way through an August 2026 contracted-capacity claim (26367), so the latest backlog figure could not be assessed. Over the next 90 days, watch three things. First, whether third-quarter results show active power moving well beyond 1 GW or stalling. Second, whether Kenilworth's first phase is energized before its 2026-12-31 deadline or quietly pushed out. Third, whether management reconciles the 5 GW and 8 GW framings. Interconnection filings for the New Jersey and Pennsylvania sites would be the first outside check on whether the contracted megawatts are physically on their way.

CoreWeave's capacity record follows two curves that rise at different speeds. Active power has climbed steadily, from 420 MW in May 2025 (claim 4972) through 470 MW (25577) and 590 MW (20199) to 850 MW at year-end 2025 (3569), and past 1 GW by May 2026 (9321). Contracted power has grown faster: 1.6 GW (4973), then 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). The gap between energized and contracted capacity has widened from roughly 1.2 GW to more than 3 GW. The 2030 ambition has also grown. In January it was framed as 5+ GW of AI factories with NVIDIA (17595). By May it had become an 8 GW active-power target (9322), with the 5 GW NVIDIA figure retained as a subset (9329). Raising a long-range target within four months, while the near-term backlog is still unbuilt, signals ambition more than delivery.

On dates, the record is thin, and that matters. Very few facility claims carry a promised date. Kenilworth's initial phase is due in 2026 (25584), and nothing since August 2025 confirms or revises it. Lancaster's 300 MW (20201) has no stated timeline. Indonesia's 360 MW targets 2028 (26332). Niagara Falls, Digital Drive, Sweden, Plano and Polaris Forge 1 have no capacity or date on record. CoreWeave has acknowledged slippage: weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828). It framed the earlier performance problem as isolated (19827). The Core Scientific merger missed its stated window (4517) and was terminated (19825), which left CoreWeave relying on a partner's delivery rather than its own.

The dataset contains no independent corroboration: no permits, no interconnection queue positions and no site reporting. The announced pace therefore rests entirely on company disclosure. The year-end active figures are self-reported, though they appear consistently across filings (3553, 3569).

Over the next 90 days, the Q3 report should give the first active-power figure since May. Growth well short of the contracted additions would confirm the widening gap. Kenilworth's 2026 deadline expires inside this window, so watch for delivery confirmation or a quiet change of language. Watch also for PJM queue filings or local permits for Kenilworth, Lancaster and Digital Drive. Finally, check whether the 8 GW target is repeated or recast.

CoreWeave's contracted power has grown in every disclosure on record: 1.6 GW (4973), 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3559), 3.5 GW (9323) and 4.2 GW by August 2026 (26367). Active capacity has also risen steadily, from 420 MW (4972) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3569) and more than 1 GW in May 2026 (9321). The two lines are not converging. The gap between contracted and energized power grew from roughly 1.2 GW to about 2.5 GW over that year, and the most recent contracted figure has no matching active number. The 2030 ambition also moved quickly. It stood at 5+ GW (17595) in January and became an 8 GW active-power target (9322) in May, while the NVIDIA-linked 5+ GW framing was restated alongside it (9329). That raised the headline by 60 percent in a little over three months without explaining what changed. Timeline discipline is weaker than the capacity narrative suggests. The firm described data center underperformance as isolated (19827), and five weeks later disclosed Core Scientific weather and design delays (19829) and an OpenAI installation delay (19828). The Core Scientific merger, which was the main route to owned capacity, was terminated (19825, 4166) and replaced by a continuing partnership (20313). The 2026-04-07 completion marker (4731) therefore lapsed without being met. Several tuck-in acquisitions also carry delayed flags against their original dates (3578, 3577, 3589). At the project level the record is thin. Only Kenilworth (25583, 250 MW), Lancaster (20201, 300 MW) and Indonesia (26331, 360 MW) have stated capacity. Niagara Falls, Digital Drive, Plano, Stockholm and Polaris Forge 1 have no capacity or cost on file, and Polaris Forge 1 has no location. This dataset contains no independent corroboration of the announced pace, meaning no permits, interconnection queue entries or site reporting. On this record every megawatt figure is self-reported. Over the next 90 days, the most testable commitment is Kenilworth's initial-phase delivery in 2026 (25584). If the year ends without an energization disclosure, that is a slip. The Q3 report should show an active-power figure that can be checked against the 4.2 GW contracted base. If active power has not moved well beyond 1 GW, the 8 GW 2030 path (9322) becomes difficult to defend. Watch also whether the Core Scientific-hosted sites are counted as recovered or quietly reclassified, and whether the Indonesia 2028 date (26332) gains a named site or utility counterparty.

CoreWeave's headline capacity figures have climbed steadily. Active power went from 420 MW in May 2025 (claim 4972) to roughly 470 MW in August (25577), 590 MW in November (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). Contracted power grew in parallel: 1.6 GW (4973), then 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3559), and 3.5 GW (9323). The long-range ambition has grown faster than either. The January 2026 framing of 5+ GW of AI factories with NVIDIA by 2030 (17595) was joined in May by an 8 GW active-power target for the same year (9322), even as the 5 GW line was repeated (9329). Going from about 1 GW to 8 GW in under five years means adding power more than twice as fast as the 2024-2025 run rate (3568, 3569). None of the 2030 targets has been tied to named sites. On dates, the record is mixed. Facility-level promises are thin. The firmest is the 2026 initial phase at Kenilworth (25584), part of a campus of up to 250 MW (25583), and it has no public update since August 2025. Lancaster's 300 MW (20201) carries no delivery date at all. The clearest slippage came through the Core Scientific relationship. The merger was flagged delayed (4731, 4517) and then terminated (20312, 4166). Weeks after the company called data-center performance issues isolated (19827), it acknowledged weather and design delays at Core Scientific sites and a slower OpenAI installation (19828, 19829). The DDTL 3.0 commitment termination date of July 2026 (4593) is also marked delayed, with no restated date on record. Independent corroboration is essentially absent. The record contains no permits, no interconnection-queue entries, and no site reporting. Every project row lists capacity and cost as blank, and the March 2026 portfolio disclosure (156) carries no figure. The announced pace therefore rests on company statements alone. Over the next 90 days, the Q3 2026 report should show whether active power has kept rising from the 1 GW mark and whether contracted power has moved beyond 3.5 GW. Kenilworth needs to show energized capacity before year-end, or the 2026 promise will slip. PJM queue and local permit filings for Lancaster and Kenilworth would be the first outside test of either campus. For Indonesia, watch for a named site or power partner to back the 360 MW, 2028 commitment (26331, 26332).

CoreWeave's contracted power has grown in every disclosure on record. It went from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3559), 3.5 GW (9323) and most recently 4.2 GW (26367). Active power has also risen steadily: 420 MW (4972), 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). The problem is the gap between the two. Contracted capacity has stayed roughly four times active capacity, so the backlog of undelivered power keeps getting larger even as deliveries speed up. The 2030 target has also moved. The NVIDIA-linked goal of 5+ GW (3893, 17595, 9329) was joined in May 2026 by an 8 GW active-power target (9322), and the company has not explained how the two figures relate. Reaching 8 GW means adding about 7 GW in four and a half years. The company built roughly 1 GW cumulatively through mid-2026, so this is a pace it has not yet shown.

The timeline record is mixed. The Core Scientific merger closing date slipped (4517, 4731) before the deal was terminated outright (4166, 20312), with a looser partnership kept in its place (20313). In December 2025 the company acknowledged installation delays tied to OpenAI capacity and weather and design problems at Core Scientific sites (19828, 19829), which it had earlier described as isolated (19827). The delayed flags on the small software acquisitions (3577 through 3589) look like a filing-date artifact rather than real slippage, and should not be read as a delivery signal.

At the project level, most sites on record have no stated capacity or date. Only Kenilworth (250 MW, initial phase in 2026; 25583, 25584), Lancaster (300 MW, undated; 20201) and Indonesia (360 MW, online in 2028; 26331, 26332) carry specifics. The dataset includes no independent corroboration: no permits, no PJM or other interconnection queue positions, and no site reporting. The announced pace therefore rests entirely on company statements. The record we received was cut off at claim 26366, so any Indonesia phasing detail in that claim is not reflected here.

Over the next 90 days, the clearest test is Kenilworth. Its initial phase is due by December 31, and energization or customer-acceptance language in the Q3 results would confirm it. Other things to watch are whether active power rises meaningfully above 1 GW, whether the 4.2 GW contracted figure grows again without a matching rise in active capacity, any public PJM queue or permitting record for Lancaster, and whether management reconciles the 5 GW and 8 GW targets or quietly drops one.

CoreWeave's capacity story is one of steady upward revision at the portfolio level and near-silence at the project level. Contracted power has climbed from 1.6 GW (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3559) and 3.5 GW (9323) in roughly a year. Active power has moved from 420 MW (4972) to just over 1 GW (9321), consistent with the year-end footprint series of 70, 360 and 850 MW (3551, 3552, 3553). Energized capacity has stayed at roughly a quarter to a third of contracted capacity throughout, so the backlog is growing about as fast as delivery. The long-range ambition has also expanded. The 5+ GW of NVIDIA-linked AI factories by 2030 (17595, 9329) now sits alongside a larger 8 GW active-power target for the same year (9322), with no reconciliation between the two figures.

On dates, the record is thin, which is itself the finding. The only dated facility promise is the Kenilworth initial phase in 2026 (25584), attached to a 250 MW ceiling (25583), and it has not been restated since August 2025. Lancaster's 300 MW (20201) carries no date at all. The clearest case of slippage was characterized rather than quantified. Performance issues were first framed as isolated (19827), then attributed to an OpenAI installation delay and to Core Scientific weather and design delays (19828, 19829). That came after the Core Scientific merger itself slid past its promised close (4517, 4731) and was terminated (19825), leaving a partnership in its place (20313). DDTL 3.0's commitment window is also flagged as delayed (4593).

No independent corroboration appears in this record: no permits, no interconnection queue positions, no site reporting. Six of the eight named projects carry no capacity or cost figure, and Polaris Forge 1 has no location. Indonesia's 360 MW (155) rests on leased colocation (8760) with an incomplete timeline. The announced pace is therefore supported only by CoreWeave's own disclosures.

Over the next 90 days, watch the Q3 report. Check whether active power moves meaningfully beyond 1 GW and whether contracted power grows again or plateaus at 3.5 GW. Watch whether Kenilworth's first phase is confirmed, quantified or quietly re-dated before its 2026 year-end deadline. Look for PJM queue or local permitting evidence for Kenilworth and Lancaster. Look for a dated, sized commitment for Indonesia. Watch for any sign that the Vera Rubin deployments (10341, 17621) and DDTL 5.0 funding (9121) are converting into named sites.

CoreWeave's capacity story is one of steady, disclosed growth in contracted power, with active power rising more slowly. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3570) and 3.5 GW by May 2026 (9323). Active power rose from 420 MW (4972) to 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (3569) and past 1 GW (9321). The year-end series of 70, 360 and 850 MW (3567-3569) shows real build-out, but the contracted-to-active ratio has only narrowed from about 3.8x to 3.5x, so the backlog is growing roughly as fast as delivery. The long-range target has also grown. Claims 17595 and 3893 framed 5+ GW of AI factories with NVIDIA by 2030. Claim 9322 now states 8 GW of active power by 2030. Reaching 8 GW from about 1 GW requires roughly 1.5 GW of energization a year, about three times the best year on record (+490 MW in 2025). The filings do not explain how the 5 GW and 8 GW figures relate. On timing, the company has acknowledged slippage. Claims 19828 and 19829 cite OpenAI installation delays and Core Scientific weather and design problems, while 19827 called the shortfall isolated. The Core Scientific merger slid past its 2025-10-01 target (4517) and was terminated (4166), though the hosting relationship continues (20313). Several acquisition claims carry promised-by dates earlier than their own disclosures (3577-3579), which looks like a dating artifact rather than a real slip. Project-level detail is thin. Kenilworth's 250 MW (25583) has a dated first phase in 2026 (25584). Lancaster's 300 MW (20201) has no date, and Indonesia's 360 MW figure (155) rests on leased colocation space (8760). Five of the eight listed projects have neither a capacity figure nor a date. This record contains no permits, interconnection-queue filings or site reporting, so the announced pace is supported only by the company's own disclosures and has not been independently checked. Over the next 90 days, the key test is whether Kenilworth's initial phase is energized by 2026-12-31. The Q3 2026 results should show whether active power keeps growing by at least 150 MW a quarter, which the 8 GW target implies. Watch also for any reconciliation of the 5 GW and 8 GW goals, and for the first utility or interconnection filings tied to Lancaster or Jakarta.

CoreWeave's announced capacity has only grown, and the reported figures line up with one another. Contracted power rose from 1.6 GW in May 2025 (4973) to 2.2 GW (25578), 2.9 GW (20200), 3.1 GW (11770, 3559) and 3.5 GW by May 2026 (9323). Active power followed a slower but steady path: 420 MW (4972), 470 MW (25577), 590 MW (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by the May 2026 report (9321). The year-end series of 70, 360 and 850 MW (3551 through 3553) shows real acceleration. The gap between contracted and energized capacity has still widened, from roughly 1.2 GW to about 2.5 GW. The long-range targets have also moved up without much intermediate scaffolding. The 5+ GW NVIDIA-linked goal (17595, 9329) was joined by an 8 GW active-power target for 2030 (9322). Hitting 8 GW would require adding well over 1.5 GW a year, roughly three times the ~490 MW added in 2025.

On dates, the record holds few dated facility promises, so it offers few chances to test them. The clearest one is Kenilworth's initial phase in 2026 (25584), against a campus of up to 250 MW (25583). Nothing on the record since August 2025 updates that milestone. The company described data center underperformance as isolated in November 2025 (19827). A month later, delays tied to an OpenAI installation (19828) and to weather and design problems at Core Scientific sites (19829) were on the record. The terminated Core Scientific merger (19825, 4166) left CoreWeave dependent on a partner it no longer controls (20313). Lancaster's 300 MW (20201) and Indonesia's 360 MW deployment (155) carry no stated delivery dates. Indonesia will run in leased colocation space (8760), which is faster to deliver but less owned capacity.

The record contains no independent corroboration: no permits, no interconnection queue positions, no site reporting. Every project entry also lacks capacity and cost fields. The pace above therefore rests entirely on company disclosures, and the claim describing the March 2026 portfolio (156) carries no figure.

Over the next 90 days, the third-quarter report is the main test. It should show whether active power keeps climbing by roughly 150 MW or more per quarter, and whether contracted power moves past 3.5 GW. Kenilworth's 2026 deadline expires on 2026-12-31, so energization or a quiet slip will become visible. Further items to watch are any utility or zoning filings for Lancaster, a date for Jakarta, the status of Core Scientific deliveries, and whether the Vera Rubin bring-up (17621, 10341) turns into named sites.

CoreWeave's capacity story has only moved in one direction. Contracted power has stepped up at every disclosure, from 1.6 GW (claim 4973) to 3.5 GW (claim 9323), and active power has risen to just over 1 GW (claim 9321). The pattern matters more than the totals. Active capacity has roughly doubled in a year, but contracted capacity has grown in parallel, so the gap of about 2.5 GW between what is signed and what is energized has not closed. The 2030 framing has also drifted: an NVIDIA-linked goal of 5+ GW of AI factories (claims 3893, 17595, 9329) now sits alongside a separate 8 GW active-power target (claim 9322), and the company has not explained how the two figures relate.

On timing, the record is thin, and that is itself a finding. Kenilworth's initial phase was promised for 2026 (claims 25583, 25584) and has not been restated since August 2025, with under three months of the year left. Lancaster's 300 MW (claim 20201) has no date attached. Five of the eight named projects carry no capacity figure at all. The clearest revision came through partners rather than CoreWeave's own sites. In November 2025 the company characterized data center performance issues as isolated (claim 19827). Weeks later it acknowledged Core Scientific weather and design delays and an OpenAI installation delay (claims 19828, 19829), after the Core Scientific merger had already slipped past its expected close and been terminated (claims 4517, 4166).

This record contains no independent corroboration: no permits, no interconnection queue positions, and no site-level reporting. The announced pace therefore rests entirely on company disclosure. The Indonesia deployment, sized at 360 MW in a truncated claim (155), is described as leased from a colocation provider (claim 8760). That suggests at least some future growth will come from third-party shells, so its credibility depends on the landlord's construction rather than CoreWeave's.

Over the next 90 days, watch the third-quarter results for whether active power has moved meaningfully past 1 GW, and whether contracted capacity rises again without a matching step in energized megawatts. Watch for any Kenilworth energization or a quiet shift of its date into 2027. Look for PJM queue or county filings tied to Lancaster, and for the identity and phasing of the Jakarta colocation partner. Finally, watch whether the Meta commitment (claim 17654) and the Vera Rubin deployments (claims 10341, 17621) get mapped to specific sites.

CoreWeave's capacity story is one of steady upward revision at the portfolio level, paired with thin detail at the project level. Contracted power has grown from 1.6 GW (claim 4973, May 2025) to 2.9 GW (20200), then 3.1 GW (11770, 3559), then 3.5 GW (9323, May 2026). Active power has risen at a similar pace: 420 MW (4972), 590 MW (20199), 850 MW at year-end 2025 (3569), and above 1 GW by May 2026 (9321). That builds on footprints of 70 MW in 2023 (3551) and 360 MW in 2024 (3568). The disclosed active figures have never moved backward, and that matters. Still, the long-dated targets now run well ahead of the contracted base. The 5+ GW of NVIDIA-linked AI factories by 2030 (3893, 17595, 9329) and the 8 GW active target (9322) would require roughly 1.5 GW of energization per year, against 3.5 GW contracted today.

On dates, the record is mixed. The Core Scientific merger missed its stated window (4517), and the company later terminated it (4166, 20312) while keeping the commercial relationship (20313). A month after management called data center performance issues isolated (19827), it acknowledged weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828). Those admissions came without restated delivery dates. The DDTL 3.0 commitment window is also flagged as delayed (4593).

Independent corroboration is the weakest part of the record. None of the eight named projects carries a disclosed cost, and only two carry capacity figures. Lancaster has 300 MW (20201), and Indonesia has 360 MW across three Jakarta sites by end-2028 (8758, 155). Indonesia is explicitly leased from an unnamed colocation provider (8760), and Polaris Forge 1 lacks even a location. Nothing in the file points to permits, PJM or ERCOT interconnection queue positions, or site reporting that would let the announced pace be checked against physical progress. The portfolio-wide snapshot (156) and the final entry (22017) are incomplete as recorded.

Over the next 90 days, watch whether third-quarter disclosures show active capacity moving meaningfully beyond 1 GW, and whether contracted power grows again or plateaus. Also look for permitting or queue evidence at Lancaster, Kenilworth and Plano, the naming of the Indonesia colocation partner, revised delivery dates for the delayed Core Scientific sites, and drawdown activity under the DDTL 4.0 and 5.0 facilities (11225, 9121) as a proxy for build pace.

CoreWeave's headline capacity numbers have only moved upward, and the record supports the direction if not the destination. Contracted power rose from 1.6 GW in May 2025 (claim 4973) to 2.9 GW by November (20200), 3.1 GW in the annual filing (3559, 3570), and 3.5 GW by May 2026 (9323). Active capacity followed a steeper relative curve: 70 MW at end-2023 (3567), 360 MW at end-2024 (3568), 850 MW at end-2025 (3569), and past 1 GW by May 2026 (9321). The long-range framing has also grown. The January 2026 target of 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) was joined in May by an 8 GW active-power target for the same year (9322). That target exceeds the 3.5 GW currently contracted, so it depends on sites that are not yet secured. It also implies roughly 1.4 GW of net additions per year, against about 490 MW added during 2025.

On timing, the clearest slippage is in partner delivery rather than in CoreWeave's own stated dates. The Core Scientific merger, expected to close around October 2025 (4517), was terminated on 2025-10-30 (4166, 20312), with the commercial relationship continuing (20313). Weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828) were disclosed in December, after a November characterization of performance issues as isolated (19827). The OpenPipe, Marimo and Monolith acquisitions are flagged as delayed (3588, 3577, 3589). Those flags look like closing dates reported after the fact rather than missed commitments.

Independent corroboration is thin. Eight projects are on record, all at announced status. Only two carry figures: Lancaster at 300 MW (20201) and three leased Jakarta facilities totaling 360 MW by end-2028 (8758, 8760). None of the provided claims cite permits, utility interconnection queue positions, or site-level reporting, so the announced pace rests on company disclosure alone. The record also truncates at claim 22017, whose content could not be assessed.

Over the next 90 days, the third-quarter report should show whether active capacity has moved meaningfully beyond the 1 GW mark and whether contracted capacity has grown past 3.5 GW. Beyond that report, three things would firm up the project-level picture: the Lancaster facility appearing in PJM queue or local zoning records, the Jakarta colocation partner being named, and evidence that the Vera Rubin deployments (10341, 17621) have reached customer service rather than validation alone.

CoreWeave's capacity story is mostly about contracted power running ahead of what is energized. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (claims 3551, 3552, 3553). An intermediate 590 MW print in November 2025 (20199) fits that path, and the company reported passing 1 GW in May 2026 (9321). Contracted power grew faster: 1.6 GW (4973), then 2.9 GW (20200), 3.1 GW (11770, 3554), and 3.5 GW (9323). That leaves roughly 2.5 GW signed but not live. The long-range target has also moved. 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) was restated in May 2026 next to a new 8 GW active-power goal for the same year (9322, 9329). The filings never say whether 8 GW replaces the 5 GW figure or counts something different. Either way, reaching 8 GW means growing roughly eightfold from today's base in about four years.

Dates have held unevenly, and some changes went unacknowledged. The Core Scientific merger carried close targets of October 2025 and then April 2026 (4517, 4731). It was terminated on 2025-10-30 (19825, 4166) and recast as a continuing partnership (20313). In December 2025 the company disclosed an OpenAI installation delay and Core Scientific weather and design delays (19828, 19829), shortly after calling the problems isolated (19827). The DDTL 3.0 commitment termination date is also flagged as delayed (4593).

Independent corroboration is thin in this record. The only named site with a disclosed size is Lancaster at 300 MW (20201). Niagara Falls, Digital Drive, Kenilworth, Plano, Stockholm and Polaris Forge 1 have no capacity, cost, permit or interconnection-queue entries on file. Indonesia's 360 MW by end-2028 (8758, 155) would be leased from an unnamed colocation provider (8760), so the build risk sits with a third party. The announced pace therefore rests almost entirely on company statements. The most recent claim, 17625, arrived without detail.

Over the next 90 days, the Q3 2026 report is the first test. Watch whether active power keeps adding 150 to 250 MW per quarter and whether contracted capacity moves past 3.5 GW. Check also whether the 5 GW and 8 GW targets are reconciled. Lancaster should show up in PJM queue filings or local land-use approvals if it is real near-term capacity. The Indonesia colocation partner may be named. Vera Rubin deployment claims (10341, 17621) should translate into specific energized sites.

CoreWeave's headline numbers have only moved upward. Contracted power rose from 1.6 GW in May 2025 (claim 4973) to 2.9 GW by November (20200), 3.1 GW in the annual filing (3554, 11770), and 3.5 GW by May 2026 (9323). Active capacity tracked behind it, from 590 MW (20199) to 850 MW at year-end 2025 (3569), then past 1 GW (9321). Year-end footprints of 70, 360 and 850 MW (3567-3569) show real delivery. The gap between contracted and energized capacity is still roughly 2.5 GW, and it has widened rather than narrowed. The long-range target has also grown. A 5+ GW AI-factory goal with NVIDIA (17595, 3893, 9329) now sits beside an 8 GW active-power target for 2030 (9322), and the company has not said whether the two overlap. Reaching 8 GW would require energizing about 1.5 GW a year, roughly three times the 2025 pace. Facility-level dates are sparse, which makes slippage hard to measure. The clearest timeline failure is the Core Scientific merger. It was promised by October 2025 (4517) and later by April 2026 (4731), then terminated on 30 October 2025 (19825, 4166), with the hosting relationship kept in place (20313). In December the company disclosed weather and design delays at Core Scientific sites and a delayed OpenAI installation (19828, 19829), after earlier calling data-center underperformance isolated (19827). The DDTL 3.0 commitment date is also flagged as delayed (4593). Most named projects have no capacity, cost, or date on record. The exceptions are Lancaster at 300 MW with no date (20201) and Indonesia at 360 MW across three leased colocation sites by end-2028 (155, 8758, 8760). The Indonesia figure equals the company's entire 2024 footprint but depends on a third party's construction. The record holds no independent corroboration, such as permits, interconnection-queue positions, or site reporting, for any project, so the pace rests entirely on company statements. Over the next 90 days, the Q3 report should show whether active capacity is climbing fast enough to make 8 GW credible and whether contracted capacity keeps outrunning it. The other things to watch are a dated milestone or utility filing for Lancaster, naming of the Indonesian colocation partner, recovery at the Core Scientific-hosted sites, and how the DDTL 5.0 financing (9121, 8736) is sized against the build plan. The latest capacity claim on record (17625) arrived incomplete and should be checked before it is relied on.

CoreWeave's reported footprint has grown steadily, and the targets have grown faster than delivery. Active power went from 420 MW in May 2025 (claim 4972) to 590 MW in November (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). That is roughly 600 MW of energized capacity added per year. Contracted power rose faster over the same period: 1.6 GW (4973), then 2.9 GW (20200), 3.1 GW (11770, 3570), and 3.5 GW (9323). The gap between contracted and active capacity widened from about 1.2 GW to about 2.5 GW. The long-range target also escalated. In January 2026 the company framed a goal of 5+ GW of AI factories with NVIDIA by 2030 (17595, 3893). By May it was citing 8 GW of active power by 2030 (9322). That figure exceeds even the current contracted base, and at the observed run rate it would take roughly a decade to reach.

The record contains very few project-level dates against which slippage can be measured. Lancaster has a stated 300 MW (20201) but no delivery date. Niagara Falls, Kenilworth, Plano, Polaris Forge 1, Stockholm and Digital Drive carry neither capacity nor schedule on record. The only dated site commitment is Indonesia: 360 MW across three Greater Jakarta facilities by end-2028 (155, 8758), delivered through leased colocation rather than self-build (8760). The clearest timing signal is the Core Scientific sequence. The merger targeted for October 2025 (4517) was terminated (4166, 20312) while the partnership continued (20313). Weeks after a performance issue was characterized as isolated (19827), the company acknowledged weather and design delays at Core Scientific sites and a delayed OpenAI installation (19828, 19829). That pattern suggests downstream schedule risk is disclosed late.

None of the claims on record is corroborated by independent evidence such as permits, interconnection queue positions, utility filings or site reporting. The announced pace therefore rests on company statements alone. The portfolio snapshot (156) and the final entry (18935) carry no usable figures.

Over the next 90 days, the Q3 2026 report is the key checkpoint. Watch whether active capacity is still growing at roughly 150 MW per quarter, whether the 8 GW target is restated or quietly reverted to 5 GW, and whether the Core Scientific-hosted capacity has come online. Outside the company, watch for PJM queue or local permitting activity tied to Lancaster, naming of the Indonesian colocation partner, and evidence that the Vera Rubin deployments (10341, 17621) are energized rather than validated in a lab.

CoreWeave's capacity story has grown at both ends. Contracted power rose from 1.6 GW in May 2025 (claim 4973) to 2.9 GW by November (20200), 3.1 GW in the February and March 2026 disclosures (11770, 3582), and 3.5 GW in May 2026 (9323). Active power moved from 420 MW (4972) to 590 MW (20199), then 850 MW at year-end 2025 (3569, 11769), and past 1 GW by May (9321). The year-end series of 70, 360 and 850 MW (3567, 3568, 3569) shows real compounding. The long-range target has also moved. The January 2026 framing was 5+ GW of AI factories with NVIDIA by 2030 (17595, 3893). By May this sat beside a broader 8 GW active-power target for the same date (9322, 9329). That new figure is more than double current contracted capacity. The company has not explained how much of the gap is expected to come from NVIDIA-linked sites versus new contracts.

Date discipline is mixed. The Core Scientific merger was the clearest planned route to owned capacity. It was marked delayed against its expected close (4517, 4731) and then terminated on 2025-10-30 (19825, 4166), with only a continuing partnership left (20313). The December disclosures of an OpenAI installation delay and of weather and design delays at Core Scientific sites (19828, 19829) followed management's November description of data center underperformance as isolated (19827). That sequence deserves skepticism. The DDTL 3.0 commitment termination date is also flagged delayed (4593). This suggests financing timelines are being revised alongside build schedules.

The record contains no independent corroboration. There are no permits, no interconnection-queue entries, and no site reporting. Most named projects carry no capacity or date on record at all: Niagara Falls, Kenilworth, Plano, Sweden, Digital Drive and Polaris Forge 1. Lancaster's 300 MW (20201) and Indonesia's 360 MW by end-2028 (155, 8758) are the only site-level figures. Indonesia relies on leased colocation space (8760), which shifts execution risk to an unnamed partner. The latest claim in the set (17625) is truncated and could not be assessed.

Over the next 90 days, watch three things. First, whether the third-quarter report restates the 8 GW target and shows active capacity keeping its roughly 150 MW-per-quarter pace. Second, whether PJM queue filings or local permits appear for Lancaster. Third, whether the Indonesia colocation provider is named and given a first energization date.

CoreWeave's capacity story has scaled steadily on paper. Contracted power moved from 1.6 GW in May 2025 (claim 4973) to 2.9 GW that November (20200), 3.1 GW by the March 2026 annual filing (11770, 3582), and 3.5 GW by May 2026 (9323). Active capacity tracked behind it, from 420 MW (4972) to 590 MW (20199) to 850 MW at year-end 2025 (3569), then past 1 GW (9321). The year-end footprint series of 70, 360 and 850 MW (3567, 3568, 3569) shows real energization, not just signed contracts. The long-range targets are another matter. The 8 GW active-power goal for 2030 (9322) sits beside a 5+ GW NVIDIA-linked target (17595, 9329), and both run well ahead of anything contracted today. Reaching the 8 GW goal would require roughly a 2.3x increase over the current contracted base within four years.

On timelines, the clearest failure is the Core Scientific merger. Its October 2025 close target (4517) slipped, and the deal was terminated on 2025-10-30 (19825, 4166), with a continuing partnership put in its place (20313). Weeks later, management acknowledged Core Scientific weather and design delays (19829) and an OpenAI installation delay (19828), having earlier described performance problems as isolated (19827). The delayed flags on the Marimo, OpenPipe and Monolith deals (3577, 3588, 3589) look like closing dates recorded after the fact rather than genuine slippage. The DDTL 3.0 commitment termination date (4593), however, warrants a check on whether it was extended.

Independent corroboration is thin. The record contains no permits, interconnection-queue entries or site reporting for any of the eight named projects. Only Lancaster (20201, 300 MW) and Indonesia (8758, 155, 360 MW by end-2028) carry stated capacities, and Indonesia relies on leased colocation (8760) rather than owned build. Niagara Falls, Kenilworth, Plano, Digital Drive, Sweden and Polaris Forge 1 have no capacity or date claims attached at all. The record as supplied ends mid-entry at claim 17625.

Over the next 90 days, the third-quarter report is the main checkpoint. It should show whether active capacity has moved meaningfully beyond 1 GW and whether contracted power has grown past 3.5 GW. Other things to watch are PJM queue or county filings for Lancaster, the naming of the Indonesian colocation partner, disclosure on the status of DDTL 3.0 commitments, and evidence that Vera Rubin deployments (10341, 17621) are landing in specific sites rather than remaining a multi-location aspiration.

CoreWeave's capacity story has grown at every disclosure point, and the active base has largely kept pace. Contracted power rose from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323) in roughly twelve months. Active power climbed from 590 MW (20199) to 850 MW at year-end 2025 (11769, 3569) and past 1 GW by May 2026 (9321), which extends a footprint series that ran from 70 MW to 360 MW to 850 MW (3567-3569). The long-range ambition has moved faster than either series. The 2030 target began as 5+ GW of AI factories with NVIDIA (17595, 3893), and by May it had become 8 GW of active power (9322), while the 5 GW framing was restated alongside it (9329). Even against today's 3.5 GW contracted figure, roughly 4.5 GW of the 2030 goal has no identified site in this record. Delivering it would require sustaining about 1.75 GW of additions per year, well above the recent quarterly run rate.

The timeline record is weaker. The Core Scientific merger missed its October 2025 target (4517) and was terminated weeks later (4166, 19825), with only a continuing partnership left in place (20313). One month after management called data center performance issues isolated (19827), disclosures cited an OpenAI installation delay (19828) and weather and design setbacks at Core Scientific sites (19829). These were not framed as revisions to earlier guidance. The DDTL 3.0 commitment termination date is also flagged as delayed (4593). The delayed flags on the tuck-in acquisitions (3577, 3588, 3589) look like artifacts of how closing dates were recorded, so they should not be read as operational slippage.

Independent corroboration is thin. Lancaster carries a 300 MW figure (20201) but no date, and six named projects, including Niagara Falls, Plano and Kenilworth, have no capacity, cost or schedule attached. The record contains no permits, interconnection queue positions or site reporting for any of them. Much of the pipeline sits in third-party buildings, including the Indonesian colocation leases behind the 360 MW Jakarta target for 2028 (8760, 8758). As a result, landlord disclosures are the most practical outside check.

The Q3 2026 results are the main event for the next 90 days. Readers should watch whether active power gains keep pace with contracted additions and whether the 8 GW target is reiterated or quietly reframed. Other markers are PJM queue or county permit filings for Lancaster, naming of the Jakarta colocation partner, and any status update on the delayed Core Scientific deliveries. The final claim on record (17625) was truncated in the source data and should be reviewed before publication.

CoreWeave's headline capacity has grown steadily, and the reported active figures largely track what was promised. Active power rose from 70 MW at end-2023 to 360 MW at end-2024 (3551, 3552), then 420 MW in May 2025 (4972), 590 MW in November (20199), 850 MW at year-end (3553, 11769), and past 1 GW by May 2026 (9321). Contracted power grew faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3554) and 3.5 GW (9323). The ambition has scaled more than either. A 5+ GW AI-factory goal with NVIDIA by 2030 (17595, 9329) has been joined by an 8 GW active target for the same date (9322). That target is more than double today's contracted base, and it implies roughly 1.9 GW of activations a year against about 490 MW delivered in calendar 2025.

The schedule record is weaker at the deal and site level than at the aggregate level. The Core Scientific merger, guided to close by October 2025 (4517), slipped and was then terminated (4166, 20312) in favour of a continuing partnership (20313). Weather and design delays at Core Scientific sites and an OpenAI installation delay surfaced weeks later (19829, 19828), after management had characterized data-center performance issues as isolated (19827). Several acquisition closings carry delayed flags (3578, 3587, 3589), though these appear to be timing artifacts rather than capacity misses. The DDTL 3.0 commitment termination date (4593) is also flagged as delayed, and its July 2026 date has now passed.

Independent corroboration is thin. Of eight named projects, only Lancaster (300 MW, 20201) and Indonesia (360 MW across three leased Jakarta-area facilities by 2028, 8758, 155, 8760) carry any capacity figure. Niagara Falls, Digital Drive, Polaris Forge 1, Sweden, Kenilworth and Plano have no capacity, cost or, in one case, location on record. The record contains no permits, interconnection-queue entries or site reporting that would let the announced pace be tested project by project. The March 2026 portfolio disclosure (156) carries no figure, and the record ends with an incomplete entry (17625).

Over the next 90 days, the third-quarter report should show whether active power has moved meaningfully beyond 1 GW and whether contracted power keeps rising toward the 8 GW goal. It is also the point at which a stall in contracted power would become visible. Other things to watch are Lancaster filings with PJM and local permitting bodies, identification of the Indonesian colocation lessor, a status update on the Core Scientific sites, and evidence that the Vera Rubin bring-up (17621, 10341) has become billable capacity supporting the Meta payment schedule (17655).

CoreWeave's capacity story is one of steady upward revision at the portfolio level and thin disclosure at the site level. Active power has grown from 420 MW (claim 4972) to 590 MW (20199), then 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). The historical footprint series of 70, 360 and 850 MW (3567, 3568, 3569) makes clear this is real compounding, not relabeling. Contracted power has climbed faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3559) and 3.5 GW (9323). The gap between contracted and energized capacity is therefore widening in absolute terms, and that gap is where delivery risk sits. The 2030 ambition has also shifted. It began as 5+ GW of AI factories with NVIDIA (3893, 17595), and by May 2026 sat alongside a broader 8 GW active power target (9322) without any explanation of how the two figures relate.

On dates, the record is mixed. The headline capacity figures carry no promised dates, so they cannot formally slip. Execution friction is documented, however: weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (18828 is not on record; see 19828), which management characterized as isolated (19827). The Core Scientific merger, slated to close by October 2025 (4517), was terminated instead (4166, 20312) and replaced by a continuing partnership (20313). That leaves CoreWeave dependent on a third party's build schedule rather than its own. The DDTL 3.0 commitment window is also flagged as delayed (4593).

Independent corroboration is the weakest part of the picture. Every named project, including Lancaster, Kenilworth, Plano, Niagara Falls and Digital Drive, sits at announced status with no capacity or cost attached, apart from Lancaster's 300 MW (20201) and the Indonesia plan for 360 MW by 2028 (8758, 155). This record contains no permits, interconnection queue entries or site reporting, so the announced pace cannot yet be verified from outside the company's own filings. The Indonesia capacity will be leased from a colocation provider (8760), which shifts construction risk to that provider. The record also appears truncated after claim 18935.

Over the next 90 days, watch whether the Q3 2026 report moves active capacity meaningfully beyond 1 GW, or whether only contracted capacity grows. Also watch for PJM queue or local permitting activity at Lancaster, for the Indonesian colocation partner to be named along with site timelines, for any update on Core Scientific delivery delays, and for evidence that the Vera Rubin deployments (10341, 17621) are converting into energized megawatts.

CoreWeave's headline capacity numbers have only moved in one direction, and the targets have grown faster than the delivered base. Active power went from 70 MW at end-2023 (3567) to 360 MW at end-2024 (3568), 590 MW by November 2025 (20199), 850 MW at end-2025 (3569, 11769), and past 1 GW by May 2026 (9321). Contracted capacity rose from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3554), then 3.5 GW (9323). The long-range ambition escalated too. The January 2026 framing of 5+ GW of AI factories with NVIDIA by 2030 (17595, 3893), restated in 9329, sits beside a new 8 GW active-power target by 2030 (9322). Calendar 2025 added roughly 490 MW of active power. Reaching 8 GW by 2030 implies roughly 1.5 to 1.8 GW per year from here. That is a multiple of any annual pace the record has shown so far, and nearly 4.5 GW of it is not yet even contracted.

On dates, the record shows slippage that was reframed rather than plainly acknowledged. The Core Scientific merger carried an expected close of October 2025 (4517), later pushed toward April 2026 (4731), and was terminated in October 2025 (4166, 19825), with a continuing partnership instead (20313). By December, reporting tied OpenAI installation delays (19828) to weather and design problems at Core Scientific sites (19829). Management had earlier described data center performance issues as isolated (19827). The DDTL 3.0 commitment termination date is also flagged as moved (4593).

Independent corroboration is thin. None of the eight named projects carries a disclosed capacity or cost on the project record. The exceptions are Lancaster at 300 MW (20201) and Indonesia at 360 MW across three leased Jakarta-area facilities by 2028 (8758, 155, 8760). Niagara Falls, Digital Drive, Polaris Forge 1, Stockholm, Kenilworth and Plano have no dates, permits, or interconnection-queue evidence on file. The portfolio disclosure as of March 31, 2026 (156) also attaches no figure. The aggregate numbers therefore rest almost entirely on company statements.

Over the next 90 days, the Q3 2026 report should show whether active capacity is compounding toward the 8 GW path or holding near the low single-gigawatt range. Lancaster needs a PJM queue position or local permitting to move from announcement to build. The Indonesia colocation partner should be identified. Watch whether the Vera Rubin bring-up (17621, 10341) turns into named sites, and whether the DDTL 5.0 financing (9121, 8736) is drawn against specific facilities.

CoreWeave's headline capacity figures have climbed steadily across every disclosure. Active power rose from 420 MW (claim 4972) to 590 MW (20199), then to 850 MW at year-end 2025 (11769, 3569), and passed 1 GW by May 2026 (9321). Contracted power grew faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3570), and 3.5 GW (9323). The long-range framing has shifted. The 5+ GW of NVIDIA-aligned AI factories by 2030 (3893, 17595, 9329) now sits alongside an 8 GW active-power target for the same date (9322). That target is more than double the current contracted base. Reaching it from roughly 1 GW in mid-2026 implies adding about 1.6 GW a year, against the roughly 490 MW added during 2025 (3568, 3569). Nothing on record yet bridges that gap.

Date discipline is weaker than the capacity series suggests. The Core Scientific merger was expected to close by October 2025 (4517), slipped, and was terminated on 2025-10-30 (4166, 20312). It was recast as a continuing partnership (20313). Six weeks later the company cited Core Scientific weather and design delays (19829) and an OpenAI installation delay (19828). In November it had characterized data center performance issues as isolated (19827), and that framing did not hold up well against the December disclosures. Several software acquisitions carry flagged expected dates that passed before closing was reported (3577–3579, 3587–3589), though these matter less for physical delivery. The DDTL 3.0 commitment termination date is also flagged as delayed (4593).

Independent corroboration is thin. The record contains no permit filings, interconnection-queue entries, or on-site reporting for any named project. Six of the eight sites, including Niagara Falls, Kenilworth, Plano and Polaris Forge 1, carry no stated capacity or cost. Lancaster's 300 MW (20201) and Indonesia's 360 MW across three leased Greater Jakarta facilities by end-2028 (155, 8758, 8760) are company assertions only. Claim 18935 is truncated in the record and should be re-pulled before publication.

Over the next 90 days, the Q3 2026 results should show whether active power keeps compounding past 1 GW and whether contracted power moves beyond 3.5 GW. Other signals to watch are a PJM queue position or local zoning action for Lancaster, a named colocation partner and permitting for the Jakarta sites, evidence that Vera Rubin deployments (10341, 17621) are energized rather than only validated, and any further delay language tied to Core Scientific-hosted capacity.

CoreWeave's company-wide capacity numbers have grown steadily, while its site-level disclosure remains thin. Active power rose from 420 MW in May 2025 (claim 4972) to 590 MW in November (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). That follows a year-end series of 70 MW in 2023 and 360 MW in 2024 (3567, 3568). Contracted power grew faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3570), and 3.5 GW (9323). The gap between signed and energized capacity is now about 2.5 GW, so contracting is outrunning delivery. The 2030 ambition has also shifted in framing. A target of 5+ GW of AI factories with NVIDIA (3893, 17595, repeated in 9329) now sits beside an 8 GW active-power target (9322), and the record does not reconcile the two. Reaching 8 GW would mean adding roughly 7 GW in about four and a half years, compared with roughly 1 GW built to date.

The dated record is mixed. The Core Scientific merger slipped past its October 2025 target (4517) and was then terminated (20312, 4166), leaving a continuing partnership (20313). In December, the company acknowledged weather and design delays at Core Scientific sites (19829) and a delay in an OpenAI installation (19828), after describing data center performance problems as isolated (19827). The DDTL 3.0 commitment termination date is also flagged as delayed (4593). By contrast, the project-level commitments carry almost no dates. Lancaster is described at 300 MW (20201) with no delivery date. The other U.S. sites, Sweden, and Polaris Forge 1 are listed with no capacity or cost. Indonesia is the one dated buildout: 360 MW across three leased colocation facilities by the end of 2028 (155, 8758, 8760). Claim 156 refers to a portfolio snapshot as of March 31, 2026 without stating a figure, and claim 18935 is truncated in the record.

The record contains no independent corroboration. There are no permits, no interconnection queue entries, and no on-the-ground reporting. The capacity trajectory therefore rests entirely on the company's own statements and should be read that way.

In the next 90 days, watch three things. First, the third-quarter print: whether active power advances meaningfully beyond 1 GW, and whether contracted capacity moves past 3.5 GW or stalls. Second, any update on the Core Scientific-hosted sites and the OpenAI deployment. Third, any outside evidence that ties projects to real progress, such as PJM queue or zoning filings for Lancaster and Kenilworth, Dominion queue activity near Richmond, and a named colocation partner in Jakarta. Draws on DDTL 5.0 (9121) and Vera Rubin deployment disclosures (10341, 17621) would also show whether capital and hardware are flowing at the stated pace.

CoreWeave's capacity story is one of steady upward revision at the aggregate level paired with almost no verifiable detail at the project level. Active power rose from 420 MW (claim 4972, May 2025) to 590 MW (20199), then 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). Contracted power climbed from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3570) and 3.5 GW (9323). The long-range target has grown faster than either. The January framing of 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) was joined in May by an 8 GW active-power goal for the same year (9322, 9329). That goal sits 4.5 GW above the contracted backlog. Reaching it implies roughly 1.75 GW of energization a year. The record shows about 490 MW added in calendar 2025 (3568 to 3569) and about 150 MW in early 2026.

On dates, the clearest miss is the Core Scientific merger. It was guided to close around October 2025 (4517) and was terminated on 2025-10-30 (4166, 20312), with the relationship recast as a continuing partnership (20313). In November, management characterized data-center performance issues as isolated (19827). A month later, weather and design delays at Core Scientific sites (19829) and a slipped OpenAI installation (19828) entered the record, and the earlier framing was never revisited. The delayed flags on the Monolith, OpenPipe and Marimo deals (3587-3589) look like close-date bookkeeping rather than operational slippage.

Independent corroboration is thin. Of eight named projects, only Lancaster (300 MW, 20201) and Indonesia (360 MW across three leased colocation sites by end-2028, 8758, 155, 8760) carry any capacity figure. The record contains no permits, interconnection-queue entries or site reporting for any of them. The aggregate megawatt figures are therefore company-reported only. The Indonesia plan relies on a leased model, which shifts delivery risk to an unnamed colocation provider. The most recent capacity claim (17804, 2026-08-11) arrived incomplete and cannot be assessed.

Over the next 90 days, watch the third-quarter report for whether active capacity has moved meaningfully beyond 1 GW and whether contracted power keeps growing toward the 8 GW goal. Watch for PJM queue or local permitting activity at Lancaster, and for disclosure of the Indonesian colocation partner. Also watch for any update on the Core Scientific sites tied to the OpenAI delay, since those are the only delivery slips the company has acknowledged.

CoreWeave's capacity narrative is one of steady upward restatement on both the active and contracted lines. Active power went from 420 MW in May 2025 (4972) to 590 MW by November (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). That is roughly a 2.4x increase in twelve months, consistent with the retrospective footprint series of 70, 360 and 850 MW (3567-3569). Contracted capacity grew more slowly in percentage terms: 1.6 GW (4973), then 2.9 GW (20200), 3.1 GW (11770, 3570) and 3.5 GW (9323). The long-range framing has shifted. The 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) was restated in May 2026 (9329) alongside a separate 8 GW active-power target for the same year (9322). The record does not explain how the two figures relate, and the larger one arrived without any visible bridge from the 3.5 GW contracted base.

The main schedule failures sit around Core Scientific. The merger's expected close slipped (4517, 4731) before the deal was terminated in October 2025 (20312, 4166), though the hosting relationship continued (20313). Within weeks, the company acknowledged weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828), after earlier describing data center performance problems as isolated (19827). Those admissions are the clearest evidence that headline active-capacity figures depend on third-party delivery the company does not control. The DDTL 3.0 commitment termination date is also flagged as delayed (4593), which suggests financing timelines have stretched along with the buildout.

At the project level, the record is thin. Of eight named sites, only Lancaster (300 MW, 20201) and Indonesia (360 MW across three leased colocation facilities by end-2028, 8758, 155, 8760) carry capacity figures. Niagara Falls, Digital Drive, Polaris Forge 1, Sweden, Kenilworth and Plano remain announced with no capacity, cost or date attached. The record contains no permits, interconnection queue entries or on-the-ground reporting, so the announced pace rests entirely on company statements. The March 2026 portfolio disclosure (156) is unquantified here, and claim 17804 is truncated.

Over the next 90 days, the third-quarter print is the key test. It should show whether active capacity has moved meaningfully beyond 1 GW, whether contracted capacity has grown past 3.5 GW, and whether management reconciles the 5 GW and 8 GW targets. The other items to watch are PJM queue or local permitting activity for Lancaster, identification of the Indonesian colocation partner, any update on the Core Scientific-hosted delays, and draws under DDTL 5.0 (9121) as a proxy for actual construction spend.

CoreWeave's capacity narrative has scaled upward at every reporting point, and the compounding is real at the level of reported active power. Active capacity went from 420 MW (4972) to 590 MW (20199), then 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). The company's own year-end series of 70, 360 and 850 MW (3567, 3568, 3569) implies roughly 490 MW added during 2025. Contracted capacity rose faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3570) and 3.5 GW (9323), so the backlog of contracted but undelivered power has widened rather than narrowed. The long-range target also moved. The January framing of 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) was joined in May by an 8 GW active-power target for the same date (9322), while the 5 GW figure was restated alongside it (9329). These may be different scopes, but the company has not clearly reconciled them. Reaching 8 GW from about 1 GW would require roughly 1.5 GW of energization per year, about three times the 2025 pace.

Date discipline is hard to test because almost none of the named sites carry capacity or delivery dates. Niagara Falls, Digital Drive, Polaris Forge 1, Sweden, Kenilworth and Plano are all listed without either. Lancaster has 300 MW but no date (20201). Indonesia is the only site-level dated commitment: 360 MW across three leased colocation facilities by end-2028 (8758, 155, 8760). Where dates did exist, they slipped. The Core Scientific merger missed its October 2025 target (4517) and was terminated (4166, 20312), with a continuing partnership offered as the replacement (20313). Within weeks, weather and design delays at Core Scientific sites (19829) and an OpenAI installation delay (19828) surfaced, after management had characterized data center underperformance as isolated (19827).

The record contains no independent corroboration: no permits, no interconnection queue positions, no utility filings, no site reporting. Every capacity figure is company-stated, and the March 2026 portfolio disclosure (156) carries no quantified breakdown in this dataset. The final claim on record (17804) is truncated and could not be assessed.

Over the next 90 days, the third-quarter report is the key checkpoint. It should show whether active capacity is growing at a rate consistent with the 8 GW goal, and whether contracted power keeps outrunning delivery. Also watch whether the 5 GW and 8 GW targets are reconciled, whether the Indonesia colocation provider and sites are named, whether Lancaster appears in PJM queue or local permitting records, and whether the Vera Rubin deployments (10341, 17621) are tied to specific energized sites.

CoreWeave's capacity story is one of steady upward revision on both the active and contracted lines, with a widening gap between them. Active power has climbed from 420 MW (claim 4972) to 590 MW (20199), 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). That is real, compounding growth: roughly 490 MW was added in calendar 2025 against 290 MW in 2024 (3552). Contracted power has grown faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3554), and 3.5 GW (9323). The undeployed backlog therefore expanded from about 1.2 GW to roughly 2.5 GW over the year. The 2030 framing has also shifted. The 5+ GW of AI factories tied to NVIDIA (17595, 9329) now sits beside an 8 GW active-power target (9322). Reaching 8 GW implies adding about 1.5 GW a year from here, which is roughly three times the demonstrated 2025 pace.

The date record is weaker than the capacity record. The Core Scientific merger, expected to close by October 2025 (4517), was terminated on 30 October (4166, 19825) and replaced by a continuing partnership (20313). Within weeks, management's description of data center underperformance as isolated (19827) was followed by acknowledged OpenAI installation delays and Core Scientific weather and design setbacks (19828, 19829). The DDTL 3.0 commitment window (4593) is flagged as delayed. None of the later capacity disclosures revisit these slips explicitly.

Independent corroboration is thin. Of the eight projects on record, none carries a filed capacity or cost figure. Lancaster's 300 MW (20201) and Indonesia's 360 MW by 2028 (8758, 155) are the only site-level numbers. Both are company-stated, and Indonesia is explicitly a leased colocation arrangement (8760), which shifts execution risk to an unnamed provider. We have no permits, interconnection queue positions, or on-the-ground reporting linking the 3.5 GW contracted figure to specific energized sites.

Over the next 90 days, the third-quarter disclosure is the key test. Active capacity should land meaningfully above 1.2 GW if the 8 GW path is credible. Any further slip in Core Scientific-hosted deliveries would signal that the late-2025 delays were not isolated. Other markers to watch are utility or PJM filings for Lancaster, naming of the Indonesian colocation partner, and evidence that the Vera Rubin deployments (10341, 17621) are energized rather than merely validated.

CoreWeave's capacity story is one of steady upward restatement, with almost no project-level detail beneath it. Active power rose from 420 MW (claim 4972) to 590 MW (20199), then to 850 MW at year-end 2025 (3569, 11769), and past 1 GW by May 2026 (9321). That pace is consistent with the reported year-end footprint series of 70, 360 and 850 MW (3567, 3568, 3569). Contracted power grew faster, from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3570) and 3.5 GW (9323). The long-range target escalated as well. The NVIDIA-linked goal of 5+ GW by 2030 (17595, 3893) was restated in May 2026 alongside a new 8 GW active-power target (9322, 9329). That target sits well above everything contracted today, so most of the 2030 ambition is not yet backed by signed power.

On timing, the record shows slippage that was conceded only after the fact. In November 2025, data center underperformance was characterized as isolated (19827). A month later, the company attributed delays to weather and design problems at Core Scientific sites (19829) and to slow installation for OpenAI (19828). The Core Scientific merger was flagged as delayed (4731, 4517) before it was terminated (20312, 4166), leaving a continuing partnership (20313) in place of owned capacity. The DDTL 3.0 commitment termination date is also marked delayed (4593).

Independent corroboration is the weakest part of the record. Of eight named projects, only Lancaster (300 MW, 20201) and Indonesia (360 MW by 2028, 8758, 155) carry any stated capacity. None carries a cost, and Polaris Forge 1 has no location attached. Indonesia is to be leased from an unnamed colocation provider (8760). Nothing in the record ties any site to a permit, an interconnection queue position or on-the-ground reporting, so the aggregate megawatt figures cannot yet be reconciled against physical builds. The final claim in the record (17625) arrived truncated and could not be assessed.

Over the next 90 days, the Q3 2026 report is the key test. It will show whether active capacity has moved meaningfully beyond 1 GW and whether contracted power has grown past 3.5 GW or plateaued. Beyond the report, watch for site-level disclosure at Lancaster, Kenilworth, Plano and Niagara Falls, for the naming of the Indonesian colocation partner, and for evidence that Vera Rubin deployments (10341, 17621) are reaching revenue-generating halls rather than validation stages.

CoreWeave's disclosed footprint has grown steadily. It has not shrunk at any point in the record, but the gap between contracted and energized power keeps widening. Active capacity went from 420 MW in May 2025 (claim 4972) to 590 MW in November (claim 20199), 850 MW at year-end 2025 (claims 11769, 3553), and past 1 GW by May 2026 (claim 9321). That is roughly 580 MW added in twelve months. Over the same period, contracted power more than doubled, from 1.6 GW (claim 4973) through 2.9 GW (claim 20200), 3.1 GW (claim 11770), and 3.5 GW (claim 9323) to 4.2 GW in August 2026 (claim 22017). About 3.2 GW is now contracted but not active, compared with roughly 1.2 GW a year earlier. The long-range target has also moved up. The January framing was 5 GW of AI factories with NVIDIA by 2030 (claims 17595, 3893). By May it had become 8 GW active by 2030 (claim 9322), a figure that implies a build rate several times anything demonstrated so far.

Date discipline is mixed. The Core Scientific merger missed its expected close (claim 4517) and was terminated (claim 4166), leaving a continuing partnership (claim 20313). The company described data center performance as isolated in November 2025 (claim 19827). Five weeks later it acknowledged weather and design delays at Core Scientific sites and a slipped OpenAI installation (claims 19829, 19828), which amounts to a revision without much explicit acknowledgement. Several tuck-in acquisitions closed after their initially indicated dates (claims 3587, 3588, 3589), and the DDTL 3.0 commitment window is flagged as delayed (claim 4593).

Independent corroboration is thin. Of eight named projects, only Lancaster (300 MW, claim 20201) and Indonesia (360 MW across three leased colocation sites by 2028, claims 8758, 8760) carry any capacity figure. None of the eight has permit filings, interconnection queue entries, or site imagery on record. The announced pace therefore rests almost entirely on company statements and financing activity (claims 11225, 9121, 8736). The claims record supplied ends mid-entry at claim 17625.

Over the next 90 days, the third-quarter report should show whether active capacity has moved meaningfully beyond 1 GW, and whether the 4.2 GW contracted figure has grown again without a matching step-up in energized power. Lancaster warrants a check for PJM queue positions and local land-use approvals. For Indonesia, the question is whether the colocation lessor is named. For Vera Rubin (claims 10341, 17621), watch for any disclosure of where the multi-location deployment is actually running, and whether the Meta commitment (claims 17654, 17655) is tied to specific sites.

CoreWeave's capacity story is one of steady operational growth running beneath a much faster-growing contracted and aspirational layer. Active power has moved from 70 MW at end-2023 (3551) to 360 MW at end-2024 (3552), 420 MW by May 2025 (4972), 590 MW by November (20199), 850 MW at end-2025 (3553, 11769), and past 1 GW by May 2026 (9321). That is real and roughly linear, adding something like 500 to 650 MW a year. Contracted power has compounded faster: 1.6 GW (4973), 2.9 GW (20200), 3.1 GW (11770, 3554), 3.5 GW (9323), and 4.2 GW by August 2026 (22017). The gap between signed and energized capacity is widening in absolute terms. The 2030 ambition also escalated quietly. It was framed as 5+ GW of AI factories with NVIDIA in January 2026 (17595, 3893), then restated alongside an 8 GW active-power target only three months later (9322, 9329). Reaching 8 GW from roughly 1 GW implies sustaining two to three times the historical annual build rate.

On timing, the record contains few hard facility dates to test. The clearest slippage is adjacent rather than site-specific. The Core Scientific merger missed its expected close (4517) and was terminated (4166, 19825), with the partnership continuing (20313). The company acknowledged weather and design delays at Core Scientific sites and an OpenAI installation delay in December 2025 (19829, 19828), after characterizing performance issues as isolated a month earlier (19827). The Indonesia commitment of 360 MW by end-2028 (8758, 155) rests on leased colocation space (8760), which shifts delivery risk to an unnamed provider.

Independent corroboration is thin. Every claim here is company-stated. All eight listed projects remain at announced status with no recorded capacity or cost, apart from a 300 MW figure for Lancaster (20201). No permits, interconnection-queue positions, or site reporting in this record confirm the pace. The UK commitment of £2.5 billion (20202) is not tied to any named site, and Polaris Forge 1 lacks a location.

Over the next 90 days, the third-quarter report should show whether active power has cleared roughly 1.2 to 1.3 GW, which would keep the historical pace intact. Other signals to watch are whether contracted capacity rises above 4.2 GW, and whether the 8 GW target is repeated or softened. Further checks are PJM queue or county permitting activity for Lancaster, disclosure of the Indonesian colocation partner, and evidence that Vera Rubin deployments (10341, 17621) are landing at named sites rather than in aggregate figures. Note that the supplied record ends mid-entry at claim 17625, so the latest August disclosures may be incomplete.

CoreWeave's headline numbers have only moved up. Active power grew from 70 MW at end-2023 (3551) to 360 MW at end-2024 (3552), 590 MW by November 2025 (20199), 850 MW at end-2025 (3553, 11769), and past 1 GW by May 2026 (9321). Contracted power has grown faster: 1.6 GW (4973), 2.9 GW (20200), 3.1 GW (11770, 3559), 3.5 GW (9323), and most recently 4.2 (17625; the record is truncated, and the unit is presumably GW). So the backlog of contracted but undelivered power has widened in absolute terms, from roughly 1.2 GW to about 3 GW. The long-range target has also grown. The NVIDIA-linked goal of 5+ GW of AI factories by 2030 (3893, 17595, 9329) was joined in May 2026 by an 8 GW active-power target for the same date (9322). The company has not explained how the two figures relate. Reaching 8 GW from roughly 1 GW would take about 1.5 GW of energization per year, three times the roughly 490 MW added in calendar 2025.

The record contains almost no site-level dates to test. The one dated project commitment, 360 MW across three leased Greater Jakarta facilities by end-2028 (155, 8758, 8760), is recent and still unfalsifiable. Lancaster's 300 MW (20201) has no delivery date. Niagara Falls, Richmond, Kenilworth, Plano, Stockholm and Polaris Forge 1 carry no capacity or cost at all. The slippage that has surfaced is attributed to third parties: weather and design delays at Core Scientific sites and a delayed OpenAI installation (19829, 19828), which management called isolated (19827). That assurance came just after the Core Scientific merger collapsed (20312, 4166), leaving CoreWeave dependent on that partner as a landlord rather than an owner (20313). The DDTL 3.0 commitment window also carries a delayed flag (4593).

No independent corroboration is on file. There are no permits, no interconnection-queue positions in PJM, NYISO or ERCOT, and no ground reporting for any named site. The March 2026 portfolio disclosure (156) and the August active-capacity figure (17811) carry no values. The pace is therefore supported only by company statements.

Over the next 90 days, the Q3 2026 report should show whether active capacity kept pace with the jump to 4.2 in contracted power, and whether the 5 GW and 8 GW targets are reconciled. Watch for megawatt figures and energization dates on the US campuses, especially Lancaster, and for queue or utility filings that match them. Also watch for the name of the Jakarta colocation partner and its first phase, and for any updated language on the delayed Core Scientific sites.

CoreWeave's headline capacity has grown quickly, but contracted capacity has grown faster than delivered capacity. Active power went from 420 MW in May 2025 (4972) to 590 MW in November (20199), 850 MW at year-end 2025 (3553, 11769), and past 1 GW by May 2026 (9321). Contracted power rose from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (3554, 11770), 3.5 GW (9323) and 4.2 GW by August 2026 (22017). The absolute gap between contracted and energized capacity has widened from about 1.2 GW to more than 3 GW. The 2030 ambition has also grown. The January framing was 5+ GW of AI factories with NVIDIA (3893, 17595). In May the company added an 8 GW active-power target (9322) alongside it (9329). That doubles the energized base roughly eightfold in under five years.

Dates at the corporate level have not held well. The Core Scientific merger was guided to close by October 2025 (4517), carried forward (4731), and then terminated on 2025-10-30 (4166, 19825, 20312), with only a continuing partnership (20313). Within weeks the company cited weather and design delays at Core Scientific sites (19829) and a delayed OpenAI installation (19828), while describing the problems as isolated (19827). The DDTL 3.0 commitment termination date (4593) is also flagged as delayed. The tuck-in acquisitions carry slipped target dates (3587, 3588, 3589), though they have little bearing on physical delivery.

At the site level, the record is thin. Only Lancaster (300 MW, 20201) and Indonesia (360 MW across three leased colocation facilities by end-2028, 8758, 155, 8760) have stated capacity. Niagara Falls, Digital Drive, Kenilworth, Plano, Stockholm and Polaris Forge 1 have no capacity, cost or dated milestone. The UK commitment (20202) is expressed in pounds, not megawatts. Nothing in this record independently corroborates the announced pace: there are no permits, interconnection queue positions or on-site reporting. Every figure here is company-sourced. The claim list provided was also truncated after 17625.

In the next 90 days, watch the Q3 results for the active-capacity print against the 1 GW baseline, and for whether contracted power holds at 4.2 GW or is restated. Look for PJM queue or local permitting activity at Lancaster, and for named colocation partners and energization dates in Jakarta. Check whether the Vera Rubin deployments (10341, 17621) show up as live capacity rather than bring-up milestones. Finally, look for any disclosure on whether the DDTL 3.0 commitment window was extended or allowed to lapse.

CoreWeave's capacity story has only moved upward, and the gap between contracted and energized power is widening. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 2.9 GW that November (20200), 3.1 GW at year-end (11770, 3570), 3.5 GW in May 2026 (9323) and 4.2 GW by August (17625). Active power rose more slowly, from 360 MW at end-2024 (3552) to 590 MW (20199), 850 MW at end-2025 (3553, 11769) and just over 1 GW in May 2026 (9321). The August active figure (17811) carries no number in the record. The long-range goal has also been raised without much comment. In January the company framed it as 5+ GW of AI factories with NVIDIA by 2030 (17595, 3893), and in May it restated that alongside an 8 GW active-power target for the same date (9322, 9329). Reaching 8 GW would mean multiplying today's energized base roughly eightfold in about four years. The historical run-rate of 70 to 360 to 850 MW (3551–3553) is steep, but it has not come close to that.

Date discipline is mixed. The Core Scientific merger, expected to close by October 2025 (4517), was terminated on October 30 (4166, 19825). Claim 4731 still carries a delayed April 2026 date, which suggests the record was never cleanly reconciled. On November 10 the company described facility underperformance as isolated (19827). Five weeks later it disclosed Core Scientific weather and design delays and a delay in installing OpenAI infrastructure (19828, 19829). That is the clearest case of guidance being revised without the earlier framing being acknowledged. The DDTL 3.0 commitment date (4593) and several small acquisition closings (3577, 3578, 3579) also slipped, though these matter less for physical delivery.

There is no independent corroboration in this record. It contains no permits, no interconnection queue entries and no site reporting. Only two named projects carry a site-level figure: Lancaster at 300 MW (20201) and Indonesia at 360 MW by 2028 (8758, 155). Indonesia is leased from a colocation provider (8760) rather than self-built. Niagara Falls, Richmond, Kenilworth, Plano, Stockholm and Polaris Forge 1 have no capacity or cost figures at all.

Over the next 90 days, the key test is whether third-quarter results put a number on active power and show it clearly above 1 GW. Other things to watch are PJM queue or local permitting activity for Lancaster, identification of the Indonesian colocation partner, any update on the Core Scientific sites, and whether the Vera Rubin deployments (10341, 17621) turn into energized megawatts rather than further contracted ones.

CoreWeave's headline capacity has grown steadily and without reversals. Active power rose from 420 MW in May 2025 (claim 4972) to 590 MW by November (20199), 850 MW at year-end 2025 (11769, 3569), and past 1 GW by May 2026 (9321). Contracted power grew from 1.6 GW (4973) to 2.9 GW (20200), 3.1 GW (11770, 3554), and 3.5 GW (9323). Contracted capacity has stayed at roughly three and a half times active capacity throughout. That means energization is keeping up with new contracting but not reducing the backlog. The long-range framing has also moved. The January 2026 target of 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595, repeated in 9329) was joined in May by an 8 GW active-power target for the same date (9322). The company did not reconcile the two. Reaching 8 GW would require adding roughly 1.75 GW a year, against about 490 MW added in calendar 2025 and around 150 MW in early 2026.

The record contains few site-level dates to test. The clearest slip was corporate rather than physical. The Core Scientific merger, expected by October 2025 (4517), was terminated on October 30 (4166, 20312), with a hosting partnership continuing (20313). Weeks later, claims 19829 and 19828 cite weather and design delays at Core Scientific sites and a delayed OpenAI installation. These came shortly after management described data center performance issues as isolated (19827). Those delays have not been tied to a revised delivery date anywhere on record. The Indonesia plan for 360 MW across three Greater Jakarta facilities by end-2028 (155, 8758) relies on leased colocation space (8760). That lowers construction risk but ties the timeline to a partner that has not been named.

Independent corroboration is thin. The eight projects on file are all listed as announced, with no capacity or cost fields. Lancaster's 300 MW (20201) and the £2.5 billion UK commitment (20202) are the only site-level figures. The record contains no permits, interconnection-queue entries, or on-the-ground reporting. The announced pace therefore rests entirely on company disclosure.

Over the next 90 days, the third-quarter report is the key checkpoint. Watch whether it states active and contracted figures, since the August 11 claims (17804, 17811) carry none. Watch also whether it restates or quietly drops the 8 GW target, and whether it discloses dates for the delayed Core Scientific-hosted capacity. Also watch for PJM queue or local permitting activity at Lancaster, the naming of the Jakarta colocation partner, and evidence that Vera Rubin bring-up (17621, 10341) is converting into revenue-generating capacity.

CoreWeave's capacity story has expanded steadily in headline terms, but the gap between what is contracted and what is energized has widened. Active power rose from 420 MW (claim 4972) to 590 MW (20199), then to 850 MW at year-end 2025 (3569, 11769), and passed 1 GW by May 2026 (9321). Contracted power grew faster: 1.6 GW (4973), 2.9 GW (20200), 3.1 GW (3559, 11770), 3.5 GW (9323) and 4.2 GW by August 2026 (17625, 18935). The backlog of contracted but undelivered power therefore went from roughly 1.2 GW to more than 3 GW in about fifteen months. The long-range target moved in the same direction, from 5+ GW of AI factories with NVIDIA by 2030 (17595, 9329) to an 8 GW active target (9322). Reaching 8 GW would require roughly 1.5 GW of energization per year, about three times the roughly 490 MW added in calendar 2025 (3568, 3569).

Timing discipline is mixed. The Core Scientific deal, which would have brought owned powered land, was expected to close around October 2025 (4517) and was terminated instead (4166, 19825). Weather and design delays at Core Scientific sites and a slipped OpenAI installation followed (19829, 19828), after management had described data center underperformance as isolated (19827). Several acquisition dates in the record are flagged as delayed (3587, 3588, 3589), but these look more like reconciliation artifacts than operational slippage. The DDTL 3.0 commitment date (4593) is also flagged as delayed without any stated revision.

The record contains no independent corroboration: no permits, no interconnection queue positions, and no site reporting. Most named projects carry no disclosed capacity or cost. Lancaster is the exception at 300 MW (20201), but with no delivery date. Indonesia is specified at 360 MW by end-2028 (8758, 155), and it is leased colocation (8760) rather than self-built, which lowers execution risk but also reduces control.

Over the next 90 days, the Q3 print is the key test. Watch whether active capacity grows meaningfully beyond the 1 GW mark or whether contracted power keeps outrunning it. Also watch for utility interconnection filings or local permits at Lancaster, Kenilworth and Niagara Falls, the name of the Indonesia colocation partner, the status of DDTL 3.0 after its July 2026 commitment date, and whether the Vera Rubin deployments (10341, 17621) show up as energized megawatts rather than validation milestones.

CoreWeave's capacity story has expanded steadily in every headline metric while the facility-level record stays thin. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 3.1 GW by February 2026 (claims 11770, 3570), 3.5 GW in May (claim 9323), and 4.2 GW by August (claims 18935, 17625). Active power has grown more slowly: 360 MW at end-2024, 850 MW at end-2025 (claims 3568, 3569), and past 1 GW in May 2026 (claim 9321). The 2030 ambition has been restated upward. It began as 5+ GW of AI factories with NVIDIA (claims 17595, 3893), and by May it sat alongside a separate 8 GW active-power target (claims 9322, 9329). That target was raised without any accompanying change to the disclosed site list. The gap between contracted and energized capacity is now roughly four to one, and it has widened rather than narrowed.

Timing credibility is mixed. The one hard near-term date is the 1.85 GW active target for end-2026 (claim 8666). Reaching it from just over 1 GW in May means energizing more capacity in about seven months than the company added in all of 2025. The history of this company's promised dates does not favor an easy reading. The Core Scientific merger slipped past its original date (claims 4517, 4731) and was then terminated (claims 19825, 4166). In November 2025, performance problems were characterized as isolated (claim 19827). A month later came an OpenAI installation delay and Core Scientific weather and design delays (claims 19828, 19829). The DDTL 3.0 commitment window is also flagged as delayed (claim 4593).

Independent corroboration is essentially absent from this record. None of the eight named projects, including Kenilworth, Lancaster, Plano, and Niagara Falls, carries a disclosed capacity, cost, or date. No permits, interconnection-queue entries, or site reporting are attached. The newest project, Indonesia, is 360 MW across three leased colocation sites by 2028 (claims 8758, 8760). That structure depends on a partner's build schedule, not CoreWeave's own. The aggregate figures therefore rest entirely on company statements.

Over the next 90 days, the key test is the Q3 active-power print. Anything well below roughly 1.4 GW would make the 1.85 GW year-end target hard to reach. Also watch whether the Indonesia colocation partner and specific sites are named, whether Vera Rubin deployments (claims 17621, 10341) produce customer-visible capacity, and how quickly DDTL 5.0 is drawn (claim 9121). Watch too for utility or county filings that tie any of the named US campuses to real megawatts.

CoreWeave's contracted power has grown quickly and on a steady cadence. It went from 1.6 GW in May 2025 (claim 4973) to 3.1 GW in the February–March 2026 filings (claims 11770, 3570), then 3.5 GW in May (claim 9323) and 4.2 GW by mid-August (claims 18935, 17625, 8522). Active capacity has grown more slowly. The stated footprint rose from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (claims 3551–3553), and passed 1 GW in May 2026 (claim 9321). The gap between contracted and energized power is therefore widening, from roughly 1.2 GW to more than 3 GW. The long-range target has also changed. It was framed as 5+ GW of AI factories with NVIDIA by 2030 (claims 3893, 17595, 9329), and by May it had become an 8 GW active-power target (claim 9322). CoreWeave has not explained how the two figures relate.

On timing, the facility record shows little slippage the company has acknowledged, but several signs of strain. The Core Scientific merger, which was expected to close around October 2025 (claim 4517), was terminated instead (claims 19825, 4166). Late 2025 brought an OpenAI installation delay and weather and design delays at Core Scientific sites (claims 19828, 19829). An earlier characterization of data-center underperformance as isolated (claim 19827) now reads as premature. The DDTL 3.0 commitment window is flagged as delayed (claim 4593). The 1.85 GW active-power target for 2026 (claim 8666) is the first near-term delivery commitment that can be tested. Reaching it requires energizing roughly 800 MW in the second half of the year, which is about the same as the entire 2025 build.

The record contains no independent evidence, such as permits, interconnection-queue filings, or site reporting, for any named project. All eight projects remain at announced status with no disclosed capacity or cost. The Indonesia plan for 360 MW across three Greater Jakarta sites by 2028 (claims 8758, 155) depends on leased colocation (claim 8760), which moves execution risk to a partner that has not been named. The August active-power disclosure (claim 17811) carries no figure in our record, and the record ends partway through claim 8523.

Over the next 90 days, the Q3 print is the main test. It should show whether active capacity is tracking toward 1.85 GW or whether that target is quietly re-baselined. Other things to watch are identification of the Indonesian colocation partner, evidence of Vera Rubin capacity actually energized rather than validated (claims 10341, 17621), and drawdown pace on the DDTL 5.0 facility (claims 9121, 8736).

CoreWeave's headline numbers have only moved in one direction, but contracted power has grown much faster than power actually energized. Active capacity went from 420 MW (claim 4972) to 850 MW at year-end 2025 (claims 3569, 11769), then past 1 GW by May 2026 (claim 9321). Over the same stretch, contracted power went from 1.6 GW (claim 4973) to 3.1 GW (claims 3570, 11770), 3.5 GW (claim 9323) and 4.2 GW by August (claims 18935, 8522). The ratio of contracted to active capacity has therefore widened from roughly 3.8x to about 4x, and the 2030 framing has changed as well. The January statement of 5+ GW of AI factories with NVIDIA (claims 3893, 17595) was restated in May next to a separate 8 GW active-power target for 2030 (claims 9322, 9329). The company has not publicly explained how the two figures relate.

The company's timeline record is weaker than its capacity record. The Core Scientific merger was expected to close by October 2025 (claim 4517) and was terminated on October 30 (claims 4166, 19825). That removed an owned-capacity path the company had leaned on. In November, CoreWeave described data center underperformance as isolated (claim 19827). A month later, it disclosed an OpenAI installation delay and weather and design delays at Core Scientific sites (claims 19828, 19829). The DDTL 3.0 commitment window is also flagged as delayed (claim 4593). None of these slips came with a restated site-level schedule.

The record contains no independent corroboration. None of the eight named projects, including Lancaster, Plano, Kenilworth, Niagara Falls and Polaris Forge 1, has a disclosed capacity or cost, and no permits, interconnection-queue entries or construction reporting are logged against them. The Indonesia plan of 360 MW across three sites by 2028 (claims 8758, 155) relies on leased colocation (claim 8760), so its pace depends on a partner the company has not named.

Over the next 90 days, the main test is the 1.85 GW active target for year-end 2026 (claim 8666). Going from just over 1 GW in May to that figure means adding more capacity in seven months than the company added in all of 2025. Third-quarter results should show whether that pace is holding. Other things to watch are a reconciliation of the 5 GW and 8 GW figures, identification of the Indonesian colocation partner, and any utility or queue filings that attach megawatts to the named U.S. sites.

CoreWeave's disclosed capacity has scaled consistently, and faster in contracts than in energized power. Active power rose from 70 MW at end-2023 (claim 3567) to 360 MW at end-2024 (claim 3568) and 850 MW at end-2025 (claims 3569, 11769), then passed 1 GW by May 2026 (claim 9321). Contracted power grew faster still: 1.6 GW (claim 4973), then 3.1 GW (claims 11770, 3559), 3.5 GW (claim 9323), and 4.2 GW by August (claims 18935, 17625). The gap between contracted and active capacity therefore widened from roughly 1.2 GW to more than 3 GW. The backlog is growing faster than the build. The long-range framing also shifted. The earlier target was 5+ GW of AI factories with NVIDIA by 2030 (claims 3893, 17595, 9329). The newer one is 8 GW of active power by 2030 (claims 9322, 8534, 8574). The company presents these as parallel goals, not a revision, but the larger figure is now the headline.

On dates, the facility record holds up better than the corporate-transaction record. The Core Scientific merger missed its stated October 2025 target (claim 4517) and was then terminated (claim 4166). That removed a path to owned power that the contracted figures may have partly assumed. The DDTL 3.0 commitment date is also flagged as delayed (claim 4593). No facility-level dated promise has visibly slipped. However, the 2026 year-end target of 1.85 GW (claims 8666, 8533) was raised by 150 MW mid-year (claim 8582). Hitting it requires energizing about 1 GW in a single year, roughly double 2025's increment of about 490 MW.

Independent corroboration is thin in this record. None of the eight named projects carries a disclosed capacity, cost, permit, or interconnection-queue reference. Indonesia's 360 MW across three Greater Jakarta sites by 2028 (claims 8758, 155) rests on leased colocation space (claim 8760) with no named operator or site. Polaris Forge 1's 400 MW contracted figure (claim 370) lacks a location in the project record. The supplied claim list is also truncated after claim 8574, so later entries may change this picture.

Over the next 90 days, watch the Q3 active-power figure. Anything materially below about 1.4 GW would put the 1.85 GW year-end target in doubt. Watch also for Vera Rubin deployments moving from bring-up (claim 17621) to multi-site production (claim 10341), for draw-down on the DDTL 5.0 facility (claims 9121, 8736), and for any named Indonesian colocation partner or site filing.

CoreWeave's capacity story has expanded steadily, with no retreat on the record. Contracted power has grown from 1.6 GW (claim 4973) to 3.1 GW (claims 11770, 3559), then 3.5 GW (claim 9323), and 4.2 GW by August 2026 (claims 18935, 8522). Active capacity has followed a similar path: 360 MW at end-2024, 850 MW at end-2025 (claims 3568, 3569), and more than 1 GW by May 2026 (claim 9321). The year-end 2026 target now sits at 1.85 GW (claims 8533, 8666) after a 150 MW upward revision (claim 8582). The long-range ambition has grown too. The NVIDIA-linked 5+ GW by 2030 (claims 17595, 9329) now sits alongside a broader 8 GW active target (claims 9322, 8534). That 8 GW figure is roughly double today's contracted base, so most of it is not yet backed by signed power.

On timelines, the headline capacity dates have not yet come due, so they cannot be judged as held or missed. The corporate calendar is weaker. The Core Scientific merger slipped past its October 2025 target (claim 4517) and was terminated weeks later (claim 4166). Several tuck-in acquisitions are flagged as delayed against their original dates (claims 3587, 3588, 3589). The DDTL 3.0 commitment date is also marked delayed (claim 4593). The Core Scientific collapse matters most for delivery, because it removed a path to owned capacity. Yet later contracted-power figures show no visible step-down or reconciliation, which suggests the gap was filled by leases that are not itemised.

The record contains no independent corroboration: no permits, no interconnection queue positions, no on-site reporting. Every named campus except Indonesia and Polaris Forge 1 carries no disclosed capacity or cost. Indonesia's 360 MW by 2028 (claims 8758, 155) is leased from an unnamed colocation provider (claim 8760). Polaris Forge 1's 400 MW (claim 370) has no location attached. Claims 156 and 17811 reference portfolio and active-capacity figures without values. For now the pace rests entirely on company disclosure.

Over the next 90 days, watch the Q3 active-capacity print. Reaching 1.85 GW by December requires adding roughly 800 MW in about seven months from the May baseline, well above the 2025 run rate. Also watch whether the Indonesia colocation partner is named, whether the Vera Rubin multi-site deployments (claims 10341, 17621) come with site-level megawatts, and whether DDTL 5.0 drawdowns (claims 9121, 8736) map to specific campuses. The strongest confirmation would be utility or state filings for Lancaster, Kenilworth, or Plano.

CoreWeave's capacity story has grown quickly, and the company says it is bound by contracts rather than aspiration. Contracted power has nearly tripled in fifteen months, from 1.6 GW (claim 4973) to 3.1 GW (claims 11770, 3570), then 3.5 GW (claim 9323), then 4.2 GW (claims 18935, 8522). Active power has kept pace. The company reports footprint of 70 MW at end-2023, 360 MW at end-2024 and 850 MW at end-2025 (claims 3567-3569), and it passed 1 GW by May 2026 (claim 9321). Headline ambitions have also moved up. The January framing was 5+ GW of AI factories with NVIDIA by 2030 (claims 17595, 3893), and by May it was 8 GW of active power by 2030 (claims 9322, 8534). The two figures are not strictly contradictory, but the second is the one to hold the company to. The near-term 2026 target was raised 150 MW to 1.85 GW (claims 8582, 8533). That is a rare upward revision, and it is paired with a statement that near-term capacity is sold out (claim 1494).

On timelines, facility-level slippage is harder to judge than corporate slippage because almost nothing on record carries a site-specific date. None of the eight named projects has a stated capacity or cost in the project records. The only site-level figures are Indonesia's 360 MW across three leased facilities by end-2028 (claims 8758, 155, 8670, 8760) and 400 MW contracted at Polaris Forge 1 (claim 370). The most material deviation is the Core Scientific merger. It was flagged delayed (claims 4731, 4517) and then terminated (claim 4166). That removed a path to owned, powered sites and left the growth plan dependent on leased and colocation capacity. The DDTL 3.0 commitment date is also marked delayed (claim 4593), though that is a financing term rather than a build date.

Independent corroboration is absent from this record. There are no permits, no interconnection-queue entries and no on-the-ground reporting for Niagara Falls, Richmond, Lancaster, Kenilworth, Plano or Stockholm. The announced pace therefore rests on company disclosure. Note also that the claim list supplied here was truncated after claim 8574, so later entries were not assessed.

Over the next 90 days, the key test is the Q3 active-power print against the 1.85 GW year-end target. Reaching it requires adding roughly 850 MW in about eight months from the May baseline, faster than anything the company has yet shown. Also watch for the naming of the Indonesian colocation partner and a first-phase energization date, for Vera Rubin capacity actually going live beyond validation (claims 10341, 17621), and for any PJM or ERCOT queue filings that tie the Lancaster or Plano sites to real megawatts.

CoreWeave's headline numbers have only grown. Contracted power rose from 1.6 GW (4973) to 3.1 GW (11770, 3570), then 3.5 GW (9323), then 4.2 GW by August 2026 (18935, 8522). The 2030 ambition was first framed as 5+ GW of AI factories with NVIDIA (17595, 3893, 9329). By May 2026 it had become an 8 GW active-power target (9322, 8534, 8574). Active capacity is the harder number, and it is also rising. The company reports 360 MW at end-2024 and 850 MW at end-2025 (3568, 3569), and said it passed 1 GW in May 2026 (9321). The year-end 2026 target now sits at 1.85 GW (8533, 8666), after a 150 MW raise (8582). Reaching that target means adding roughly 800 MW in about seven months, well above the roughly 490 MW added in all of 2025. The 8 GW goal is nearly double today's contracted pipeline, so much of it is not yet backed by signed power.

On timelines, the clearest slips come from transactions rather than buildings. The Core Scientific merger target moved from October 2025 (4517) to April 2026 (4731) before the deal was terminated (4166). That removed a path to owned, powered sites. The DDTL 3.0 commitment date (4593) is also flagged as delayed. At the facility level, few dated promises exist to test. Niagara Falls, Digital Drive, Kenilworth, Lancaster and Sweden carry no capacity, cost or delivery date in the record. Indonesia is the exception: 360 MW across three Greater Jakarta sites by end-2028 (8758, 155, 8670). That capacity is leased from a colocation provider (8760), not self-built. Polaris Forge 1 appears only as 400 MW contracted (370), with no location or energization date.

The record contains no independent corroboration: no permits, no interconnection queue entries, no site reporting. Every capacity figure here is company-stated. The step-ups in contracted power are plausible given the Meta commitment (17654) and the Anthropic agreement (8571), but they cannot be checked against grid or construction evidence.

Over the next 90 days, watch the third-quarter active-capacity print. Anything much below about 1.4 GW would make the 1.85 GW year-end target hard to reach. Also watch whether contracted power moves beyond 4.2 GW, and whether the Indonesia colocation partner and site permits are named. Two further signals would firm up the picture: capacity figures attached to the named U.S. campuses, and confirmation that Vera Rubin deployments (10341, 17621) are energized rather than validated. The claim list supplied here was truncated after 8574, so later claims may change this read.

CoreWeave's announced capacity has grown at every disclosure. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3567, 3568, 3569). It passed 1 GW by May 2026 (9321), and the year-end 2026 target is now 1.85 GW (8533, 8666), which was itself raised by 150 MW (8582). Contracted power has grown faster: 1.6 GW (4973), then 3.1 GW (11770, 3570), 3.5 GW (9323) and 4.2 GW (17625, 8522). The 2030 ambition also changed. It started as 5+ GW of AI factories tied to NVIDIA (17595, 9329) and is now 8 GW of active power (9322, 8534). These are different metrics, and the record does not say whether the 5 GW figure sits inside the 8 GW. One definitional wrinkle stands out: 420 MW of active power "added" (4972) is hard to reconcile with a 360 MW end-2024 footprint (3568).

Few of the facility promises carry dates, so there is little to hold them against. The dated slippage is in the M&A layer, and it matters for capacity. The Core Scientific merger was due by October 2025 (4517). It was flagged delayed (4731) and then terminated (4166). That removed a path to owned power, so the build-out rests more on leases and third-party campuses. The Indonesia plan follows that model: 360 MW across three Greater Jakarta sites by end-2028 (8758, 155), in leased colocation space (8760), with no partner or per-site split disclosed. The delayed tags on the DDTL 3.0 termination date (4593) and the software acquisitions (3587, 3588, 3589) look like amendments or date-extraction artifacts rather than delivery misses.

The record contains no independent evidence of physical progress: no permits, interconnection queue positions or site reporting. All seven named projects remain at announced status with no capacity or cost attached. The only site-level figure is 400 MW contracted at Polaris Forge 1 (370), a third-party-developed campus. Financing (11225, 9121, 8736) and customer commitments (17654, 8571, 4352) show demand and funding, not energized megawatts.

The next 90 days center on the Q3 print. Reaching 1.85 GW means adding roughly 850 MW in 2026, about the same as all of 2025. The key signals are whether the active-capacity number keeps that pace, whether the target is restated, and whether the contracted-to-active gap keeps widening. Also watch for a named Indonesian colocation partner, and for utility filings or interconnection entries at Lancaster, Kenilworth, Niagara Falls and Digital Drive. The Vera Rubin deployments (10341, 17621) should show up as disclosed megawatts, not just validation milestones.

CoreWeave's capacity story has grown on every axis, and the upward revisions are real. Contracted power has more than doubled in fifteen months, from 1.6 GW (4973) to 3.1 GW (11770, 3554), then 3.5 GW (9323), then 4.2 GW (8522, 17625). Several of those entries restate the same filing, so the true number of step-ups is four, not a dozen. Active power went from 70 MW to 360 MW to 850 MW across three year-ends (3551-3553) and passed 1 GW by May 2026 (9321). The 2030 ambition moved from 5+ GW of NVIDIA-linked AI factories (17595, 3893, 9329) to 8 GW of total active power (9322, 8534). These measure different things, so the jump is partly a change of denominator rather than a pure raise. On the near term, the year-end 2026 target of 1.85 GW (8533, 8666) was lifted by 150 MW (8582). Hitting it requires adding roughly a gigawatt in calendar 2026, about double the 2025 increment.

Date discipline is harder to judge because most capacity claims carry no deadline. Where dates exist, the record is mixed. The Core Scientific merger slipped past its expected close (4517) and was then terminated (4166). That removed an owned-site pathway the company had presented as capacity-accretive, and later guidance did not address the loss explicitly. The DDTL 3.0 commitment termination date (4593) is flagged delayed with no stated revision. The Indonesia program is the only facility disclosure with a firm horizon: 360 MW across three Greater Jakarta sites by end-2028 (8758, 155). Even there, the sites are leased from an unnamed colocation provider (8760).

Independent corroboration is thin. All seven named projects remain at announced status with no disclosed capacity or cost. The record contains no permits, interconnection queue positions, or site reporting that tie the 4.2 GW contracted figure to specific energization schedules. The Polaris Forge 1 figure of 400 MW contracted (370) is the only site-attributed megawatt number outside Indonesia. The model is leasing capacity and buying demand, and demand is well evidenced: Meta's $14.2 billion through 2031 (17655) and the claim that near-term capacity is sold out (1494). Supply-side verification lags that.

Over the next 90 days, the Q3 report should show whether active power is tracking toward 1.85 GW, which would need roughly 1.4 GW or more by quarter-end. It should also show whether the undated 'active power capacity' disclosure (17811) acquires a figure. Other things to watch are identification of the Indonesian colocation partner, any site-level filings for Niagara Falls, Lancaster, Kenilworth or Digital Drive, and any amendment or explanation of the DDTL 3.0 date.

CoreWeave's capacity story has only grown, and it has grown faster in contracted than in delivered terms. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 3.1 GW by the 2025 annual filing (claims 11770, 3554), 3.5 GW in May 2026 (claim 9323), and 4.2 GW by mid-August (claims 17625, 8522). Active power also rose, from 70 MW at end-2023 (claim 3551) to 360 MW at end-2024 (claim 3552), 850 MW at end-2025 (claim 3553), and past 1 GW by May 2026 (claim 9321). The 2030 ambition moved from a 5+ GW figure framed around NVIDIA (claims 17595, 3893, 9329) to an 8 GW active-power target (claims 9322, 8534). The gap between contracted and energized capacity is now about four to one. That makes conversion, not demand, the variable that matters.

On dates, the facility targets have mostly not reached a test yet. The corporate calendar has been less tidy. The Core Scientific merger, promised by October 2025 (claim 4517), was terminated weeks later (claim 4166). Marimo, OpenPipe and Monolith all closed past their indicated dates (claims 3587 to 3589). The DDTL 3.0 commitment window is also flagged as delayed (claim 4593). None of these were facility slips. But the pattern of targets being quietly overtaken is worth keeping in mind. The near-term test is the 1.85 GW year-end 2026 target (claims 8666, 8533), which was itself raised by 150 MW (claim 8582). Getting there from roughly 1 GW in May means adding about 850 MW in eight months. All of 2025 added roughly 490 MW.

The record contains no independent corroboration of any site. There are no permits, no interconnection queue positions, and no on-the-ground reporting. The named campuses at Niagara Falls, Kenilworth, Lancaster and Stockholm carry no disclosed capacity or cost. Polaris Forge 1's 400 MW figure is contracted, not energized (claim 370). The Indonesia plan of 360 MW across three Greater Jakarta sites by 2028 (claims 8758, 8670) rests on leased colocation space (claim 8760). That means the schedule belongs partly to an unnamed landlord.

Over the next 90 days, watch the third-quarter active-power print. Anything materially below about 1.4 GW would make 1.85 GW by December hard to reach. Also watch for the Indonesian colocation partner to be named, for utility or grid filings tied to Polaris Forge 1 and the New York and Pennsylvania sites, and for whether the 8 GW figure is restated or quietly reframed. Note that the supplied claim list ends mid-entry after claim 8466, so later disclosures may not be reflected here.

CoreWeave's announced capacity has only moved upward, and the pace has quickened. Contracted power roughly doubled from 1.6 GW (4973) to 3.1 GW (11770, 3554) within ten months, then reached 3.5 GW (9323) and 4.2 GW (17625, 8522) within six more. Active power followed a steeper curve. It went from 70 MW (3551) to 360 MW (3552) to 850 MW (3553) across three year-ends and passed 1 GW by May 2026 (9321). The 2030 ambition has also grown, from the NVIDIA-linked 5+ GW (17595, 3893) to 8 GW active (9322, 8534). That 8 GW target now sits nearly 4 GW above everything contracted, so roughly half of the 2030 goal has no identified power behind it on this record.

On dates, the facility targets have held so far. The year-end 2026 active target of 1.85 GW (8666, 8533) was raised by 150 MW (8582) rather than cut. The slippage shows up in transactions and financing. The Core Scientific merger missed its expected close (4517) and was terminated (4166), yet claim 4731 still carries an April 2026 date, which looks like a stale record rather than a live commitment. The Marimo, OpenPipe and Monolith deals are each flagged as late against their original targets (3577 through 3589). The DDTL 3.0 commitment termination date (4593) is also marked delayed, and the record contains no acknowledgement of an extension.

Independent corroboration is thin. All six named projects are listed as announced, and none has a capacity or cost field attached. The dataset contains no permits, interconnection queue positions or site reporting for Niagara Falls, Kenilworth, Lancaster or Stockholm. The 400 MW at Polaris Forge 1 (370) rests on a third-party developer's buildout. The Indonesia plan for 360 MW by 2028 (8758, 155) depends on a colocation lessor who is not named (8760). The most material outside signal is lender behaviour: successive DDTL facilities (11225, 9121, 8736) indicate that banks are underwriting the pipeline, but they do not verify physical delivery.

Over the next 90 days, the key test is the third-quarter print. Reaching 1.85 GW requires adding about 1 GW during 2026, which is more than the entire 2025 build, so active capacity in the high-1.4 GW range or above would suggest the target is on pace. Other things to watch are whether contracted capacity rises again past 4.2 GW, the naming of the Indonesian colocation partner, any permit or utility filings at Lancaster or Kenilworth, the Vera Rubin deployments (17621, 10341) moving from validation to revenue, and clarification of the DDTL 3.0 terms. Note that the claims record supplied here was cut off at claim 8466, so later entries are not reflected.

CoreWeave's announced capacity has scaled consistently upward. Contracted power roughly doubled from 1.6 GW in May 2025 (claim 4973) to 3.1 GW by the 2025 annual filing (claims 11770, 3570), then rose to 3.5 GW in May 2026 (claim 9323) and 4.2 GW by mid-August (claims 17625, 8522). Active capacity has kept pace on paper. The filing-reported footprint grew from 70 MW (claim 3567) to 360 MW (claim 3568) to 850 MW at end-2025 (claim 3569), and the company said it passed 1 GW by May 2026 (claim 9321). The long-range framing has also grown. The 5+ GW NVIDIA-linked AI factory goal (claims 3893, 17595, 9329) now sits alongside a broader 8 GW active-power target for 2030 (claims 9322, 8534, 8574). The year-end 2026 target of 1.85 GW (claims 8666, 8533) was itself raised by 150 MW in August (claim 8582). That target implies adding about as much capacity in the second half of 2026 as the company held in total at the end of 2025. This is the most demanding near-term commitment on record.

The record on dates is weaker than the record on capacity. The Core Scientific merger missed its expected close (claim 4517) and was terminated in October 2025 (claim 4166). The termination removed a path to owned, powered sites, and nothing has been announced that visibly replaces it. The tuck-in acquisitions of Marimo, OpenPipe and Monolith carry dates flagged as delayed (claims 3587, 3588, 3589). Those deals nonetheless closed and were later reported without comment on timing (claims 11775, 11776), so the slippage was never acknowledged. The DDTL 3.0 commitment window is also flagged as delayed (claim 4593).

Independent corroboration is thin. Niagara Falls, Kenilworth, Lancaster and Stockholm remain in announced status with no disclosed capacity or cost, and the record contains no permits, interconnection-queue positions or on-site reporting for them. The only site-level figures are third-party or leased capacity. Polaris Forge 1 carries 400 MW contracted (claim 370), and the three Greater Jakarta facilities are leased from a colocation provider (claims 8758, 8760) for 360 MW by 2028 (claim 155). The growth claims therefore rest mostly on company-reported aggregates rather than on verifiable project milestones.

Over the next 90 days, the third-quarter print is the key test. Watch whether active power is tracking toward 1.85 GW or whether the year-end target is quietly revised. Watch for named colocation partners and energization dates in Indonesia, and for permit or utility filings at Lancaster and Kenilworth. Watch also for evidence that Vera Rubin deployments (claims 10341, 17621) are landing in energized halls rather than in contracted capacity that has not yet been built.

CoreWeave's disclosed footprint has grown quickly and consistently. Active power rose from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553). It then passed 1 GW by May 2026 (9321). Contracted power has risen faster. It went from 1.6 GW (4973) to 3.1 GW (11770, 3559), then 3.5 GW (9323), then 4.2 GW by August (17625, 8522). The long-range targets have also grown. The January framing of 5+ GW of AI factories with NVIDIA by 2030 (3893, 17595) now sits beside a broader 8 GW active target (9322, 8534, 8574). The two figures describe different things and should not be conflated. Nothing on record has shrunk. What has widened is the gap between contracted and energized power: about 4.2 GW signed against roughly 1 GW live.

On dates, the facility record is thin, so there is little to test. The clearest reversal is corporate, not physical. The Core Scientific merger was restated through late September 2025 (4313, 4294) and then terminated on 2025-10-30 (4166). Its earlier promised dates (4517, 4731) were never met. Marimo, Monolith and OpenPipe carry 2025 target dates flagged delayed (3577, 3578, 3579), though later filings describe them as completed. The DDTL 3.0 commitment date (4593) is also flagged delayed without a stated revision. The 2026 active-power target of 1.85 GW (8666, 8533) was raised by 150 MW in August (8582). Delivering it means adding roughly as much capacity in 2026 as in all prior years combined.

Independent corroboration is absent from this record. There are no permits, interconnection queue positions or site reporting for Niagara Falls, Kenilworth, Lancaster or Sweden, and none carries a capacity or cost figure. Indonesia is the only project with numbers: 360 MW across three leased colocation sites by 2028 (8758, 155, 8760, 8670). Polaris Forge 1 is a 400 MW contracted position at a third-party campus (370), not owned build. The August active-power disclosure (17811) and the March portfolio claim (156) carry no values. The record provided also ends mid-list, so later claims may be missing.

Over the next 90 days, the third-quarter report is the key test. It should show whether active power is tracking toward 1.85 GW, which implies well over 200 MW energized per quarter. Also watch for the Indonesia colocation partner and sites to be named, for NYISO or PJM filings tied to the New York, New Jersey and Pennsylvania sites, for energization milestones at Polaris Forge, and for any restated status on DDTL 3.0.

CoreWeave's capacity story has only grown. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (claims 3551, 3552, 3553). The company then said it had passed 1 GW in May 2026 (claim 9321). Contracted power grew even faster, from 1.6 GW in May 2025 (claim 4973) to 3.1 GW (claim 11770), 3.5 GW (claim 9323) and 4.2 GW by August 2026 (claims 17625, 8522). The long-range target changed shape along the way. In January it was 5+ GW of AI factories with NVIDIA by 2030 (claims 3893, 17595). By May it had become 8 GW of active power by 2030 (claims 9322, 8534, 8574), a larger number measured differently. That change has not been reconciled on the record. The near-term guide was raised by 150 MW to 1.85 GW for year-end 2026 (claims 8582, 8533), which reads as confidence rather than retreat.

The facility dates hold up better than the corporate ones. The Core Scientific merger, which carried an October 2025 target (claim 4517), was terminated instead (claim 4166). The DDTL 3.0 commitment termination date is flagged as delayed (claim 4593). None of the site-level timelines on record has visibly slipped, but that is partly because few have dates at all. Indonesia is the exception: 360 MW across three leased Greater Jakarta facilities by end-2028 (claims 8758, 155, 8760).

Independent corroboration is the weak point. All six named projects, including Niagara Falls, Kenilworth, Lancaster and Stockholm, sit at announced status with no disclosed capacity or cost. The record holds no permits, interconnection queue positions or on-the-ground reporting for any of them. Only Polaris Forge 1 carries a site-specific figure, 400 MW contracted (claim 370). Much of the reported growth therefore rests on company aggregates and on third-party or leased capacity, not on self-built sites a reader could check.

The first thing to watch in the next 90 days is the third-quarter disclosure. Hitting 1.85 GW means adding roughly 850 MW in the second half, about what the whole of 2025 delivered. Second, whether the August active-capacity figure (claim 17811) is published with a number attached. Third, whether the Indonesia colocation provider is named. Fourth, whether any NYISO, PJM or Swedish grid filings appear for the U.S. and Nordic sites. Finally, whether management explains how the 5 GW and 8 GW 2030 targets relate.

CoreWeave's capacity story is one of steady upward revision, told almost entirely in portfolio-level aggregates. Active power grew from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553), then passed 1 GW by May 2026 (9321). Contracted power rose faster, from 1.6 GW (4973) to 3.1 GW (11770, 3554), 3.5 GW (9323) and 4.2 GW by August (8522, 17625). The long-range target has also grown. It was framed as 5+ GW of AI factories with NVIDIA by 2030 (17595, 3893), and later became 8 GW of active power by 2030 (9322, 8534, 8574). The company has not clearly reconciled the two, and they may measure different things. The 2026 year-end target was raised by 150 MW to 1.85 GW (8582, 8533, 8666). The company moves its targets up, not down.

On dates, facility claims rarely carry firm promise dates, so few can be tested. The slippage on record sits in adjacent commitments. The Core Scientific deal missed its 2025-10-01 target (4517) and was terminated weeks later (4166), yet claim 4731 still carries a live delayed status that should be retired. The Marimo, OpenPipe and Monolith closings (3587, 3588, 3589) also ran past their tagged dates. The DDTL 3.0 commitment window (4593) is flagged delayed too. None of these is a facility, but they show that announced timetables have not always held.

There is no independent evidence in this record. It contains no permits, no interconnection queue entries, and no site reporting. The named US and Swedish projects (Niagara Falls, Kenilworth, Lancaster, Stockholm) have no capacity or cost attached. Indonesia's 360 MW by 2028 (8758, 155) will be leased from an unnamed colocation provider (8760). Polaris Forge 1's 400 MW (370) is contracted third-party capacity. The pace is therefore supported by company disclosure and customer backlog, such as the Meta commitment (17654), rather than by physical milestones outside CoreWeave's own reporting.

Over the next 90 days, the Q3 print is the key test. Reaching 1.85 GW means adding roughly 1 GW in a single year, about double 2025's increment. The August active-capacity figure (17811) was recorded without a value. Also watch for the Indonesia colocation partner and site to be named, for energization evidence at Polaris Forge, for site-level MW on the named campuses, and for how Vera Rubin deployments (17621, 10341) and DDTL 5.0 funding (9121, 8736) map onto specific buildings.

The capacity story here scales cleanly on paper and thins out considerably when you ask what has actually been energized. Active power moved from 420 MW added against 1.6 GW contracted (claims 4972, 4973) to 850 MW active against 3.1 GW contracted at the end of 2025 (claims 11769, 11770, 3553), past 1 GW active with 3.5 GW contracted in May 2026 (claims 9321, 9323), and to 4.2 GW contracted by August (claims 17625, 8522). Note that contracted capacity has grown roughly three times faster than active capacity across that span. The 2030 goalpost also changed shape rather than simply rising: the 5+ GW figure (claims 3893, 17595, 9329) is an NVIDIA co-branded subset, while the 8 GW figure (claims 9322, 8534, 8574) is company-wide, and the two are presented interchangeably in a way that flatters the trajectory.

On dates, the clearest failure is the Core Scientific merger, promised for October 2025 (claim 4517) and reaffirmed through late September (claims 4331, 4313, 4294) before termination on 30 October (claim 4166). Contracted-capacity guidance rose afterward without any visible re-derivation of where that capacity now sits, shifting the base from owned sites toward leased and greenfield. The tuck-ins — OpenPipe, Marimo, Monolith (claims 3578, 3577, 3579) — are minor but all surfaced late in the annual filing, which says something about disclosure cadence. The DDTL 3.0 commitment termination slipping (claim 4593) into DDTL 4.0 and 5.0 (claims 11225, 9121) shows expansion financed on a rolling, roughly quarterly refinancing cycle.

Independent corroboration is where this record is weakest. All six named sites remain at announced status with no capacity or cost attached except Polaris Forge 1 at 400 MW contracted (claim 370). Nothing in the file points to a permit, an interconnection queue position, or a utility filing. Indonesia is explicitly a colocation lease (claim 8760) for 360 MW by end-2028 (claims 8758, 155), so that schedule belongs to a partner.

Over the next ninety days, the binding test is the 1.85 GW year-end target (claims 8666, 8533, 8582), which requires energizing in seven months roughly what took the prior three years. Watch the Q3 print for a quiet re-basing, watch for Vera Rubin moving from bring-up (claim 17621) to revenue, and watch Kenilworth, Lancaster and Niagara Falls for any filing that converts announced into shovels.

CoreWeave's announced capacity has scaled monotonically and without a single downward revision, which is itself the thing to interrogate. Active power moved from 420 MW (4972) to 850 MW (11769, 3553) to above 1 GW (9321), while contracted power went 1.6 GW (4973) to 3.1 GW (11770) to 3.5 GW (9323) to 4.2 GW (17625, 8522). The 2030 ambition ratcheted from a 5 GW figure framed around NVIDIA (17595, 3893, 9329) to an 8 GW active target (9322, 8534, 8574), and the two numbers now coexist in the record without reconciliation — one is a partnership envelope, one a company-wide goal, and they are not additive. The near-term target was also revised upward, not down: 1.85 GW by year-end 2026 (8666, 8533) including an explicit 150 MW increase (8582).

On dates, the clean failure is Core Scientific. Agreed with an October 2025 close (4517), re-affirmed at 0.1235 across five separate filings (4729, 4731, 4331, 4314, 4294), then terminated (4166). Notably, contracted capacity kept climbing afterward, which means the 3.1 to 4.2 GW step rests on third-party leases rather than owned sites — consistent with the colocation structure disclosed for Indonesia (8760). The tuck-in software deals flagged as delayed (3578, 3577, 3589) appear as completed in later disclosures (11775, 11776, 10329); that reads as a filing artifact rather than genuine slippage, but the record never acknowledges the reset.

Independent corroboration is thin where it matters most. The strongest external signals are financing and demand, not construction: successive DDTL facilities (4593, 11225, 9121, 8736), the Meta commitment (17654, 17655), Anthropic (8571), and the NVIDIA backstop (4352), plus hardware validation (17621, 10341). What is absent is permitting or interconnection evidence for Niagara Falls, Kenilworth, Lancaster, or Stockholm — all still carry announced status with no capacity, cost, or dated milestone. Polaris Forge 1 has a 400 MW contracted figure (370) and no energization date. Indonesia's 360 MW by end-2028 (8758, 155, 8670) has no interim milestone either.

Over the next 90 days, the arithmetic is the test: reaching 1.85 GW from roughly 1 GW in May requires the Q3 print to land near 1.4 GW. Watch also for a first date attached to any of the four dormant campuses, DDTL 5.0 sizing, and whether the 5 GW and 8 GW targets are ever squared. The claim record thins after mid-August 2026, so Q3 disclosure is the first real checkpoint.

The capacity story here is one of consistent upward revision rather than retreat, which is itself the thing to scrutinize. Contracted power has roughly tripled inside sixteen months (4973 to 11770 to 9323 to 17625), and active power has followed a similar curve off a small base (4972, 3553, 9321). Nothing has been quietly shrunk. The pressure point is the year-end 2026 figure: reaching 1.85 GW active (8533, 8666, including the 150 MW raise in 8582) requires adding roughly as much live capacity in a single half-year as the company brought online in its entire history through 2025. That is not impossible given the Vera Rubin bring-up cadence (17621), but it is the claim most likely to be settled by energization schedules rather than intent.

On dates, one promise clearly failed: the Core Scientific merger was announced with a targeted close, reaffirmed on identical terms at least four times through September, and then terminated (4729, 4517, 4294, 4166). The significance is structural, not cosmetic — the path to owned substations and owned interconnection was abandoned, so today's contracted gigawatts rest on leases and colocation. The Indonesia build is explicit about this (8760). Separately, the 2030 framing has shifted from 5+ GW of NVIDIA AI factories to 8 GW of active power (3893, 9322, 8574) without the earlier number being retired (9329); these measure different things, and the substitution has not been acknowledged as a change.

Independent corroboration is where the record is thinnest. Five named campuses remain at announced status with no disclosed capacity, and only Polaris Forge 1 carries a site-level figure (370). The corporate totals therefore cannot be reconciled bottom-up from anything on record — no permits, no queue positions, no construction reporting. The repeatedly flagged delays on small software acquisitions (3577, 3588) look like filing artifacts, not operational slippage, and should not be read as such.

For the next ninety days: whether the 1.85 GW exit target is reaffirmed or softened at Q3 results, whether contracted capacity moves past 4.2 GW while the delivered number lags, site-level permitting in Greater Jakarta against the 2028 date (8758), and whether a successor to the DDTL 5.0 facility (9121) appears, since the buildout is debt-paced.

The shape of CoreWeave's disclosure has changed more than its delivery has. Contracted power has been the headline number since the 1.6 GW figure in May 2025 (claim 4973), and it has roughly tripled through 3.1 GW (claims 11770, 3554), 3.5 GW (claim 9323) and 4.2 GW (claims 17625, 8522). Active power, the number that actually bills, moved from 360 MW to 850 MW to just past 1 GW over the same window (claims 3552, 3553, 9321). The gap between the two has widened from under fourfold to better than fourfold on a much larger base, which means the contracting book is outrunning energization rather than tracking it. That is not inherently a failure, but it converts the story from procurement to construction and interconnection, where CoreWeave has less control.

The firmest near-term commitment is 1.85 GW active by year-end 2026 (claims 8666, 8533), which was raised by 150 MW mid-year (claim 8582). Reaching it requires energizing in seven months roughly what the company energized cumulatively through 2025. The 2030 picture carries two unreconciled figures, 5+ GW of NVIDIA-linked AI factories (claims 17595, 3893, 9329) and 8 GW of active power (claims 9322, 8534, 8574), presented alongside one another without a stated relationship.

On dates, the clearest break is Core Scientific. It was repeatedly affirmed through September 2025 (claims 4729, 4331, 4314, 4294) against an October 2025 close (claim 4517), then terminated (claim 4166) — and capacity guidance rose afterward without public accounting for the owned-site pipeline that walked away. The delayed flags on OpenPipe, Marimo and Monolith (claims 3578, 3577, 3579) look like filing-date artifacts on software deals, not facility slippage, and should not be read as evidence of construction delay.

Independent corroboration is thinnest exactly where it matters. Niagara Falls, Kenilworth, Lancaster and Stockholm remain announced with no capacity or cost on record. Indonesia is leased from an unnamed colocation provider (claims 8760, 8758, 155), so its 360 MW by 2028 rests on a third party's schedule. Watch the successive DDTL facilities (claims 11225, 9121, 8736) as the real cadence signal; watch whether the 1.85 GW target and the 150 MW raise survive the next print; watch PJM and NYISO queue postings for the four unquantified sites; and note that the NVIDIA purchase commitment (claim 4352) makes "sold out" (claim 1494) weaker evidence than it appears.

The shape of this record is a contracted-capacity curve that has roughly doubled every two quarters while the delivered curve has advanced at a steadier, slower pace. Contracted power moves from 1.6 GW (4973) through 3.1 GW (11770, 3554) and 3.5 GW (9323) to 4.2 GW (8522, 8268) in about fifteen months; active power moves 360 to 850 to just past 1 GW (3552, 3553, 9321). Both are real growth, but the gap between them widened rather than closed, which is the normal signature of a company selling forward faster than it energizes. The 2030 framing also changed metric mid-stream: the 5+ GW figure originally attached to NVIDIA AI factories (17595, 3893, 9329) sits alongside an 8 GW active-power goal announced four months later (9322, 8534, 8574) without an explicit reconciliation of the two.

On dates, the largest revision is the Core Scientific transaction, which moved from an October 2025 close (4517) to an April 2026 outside date (4731) and then terminated entirely (4166). That matters less as a deal failure than as a sourcing change: the owned-site, self-interconnected path was replaced by leased and colocation capacity in Indonesia (8760) and Stockholm (7871), and the 3.1-to-4.2 GW step happened after that pivot. The smaller acquisition slips flagged for OpenPipe, Marimo and Monolith (3588, 3577, 3579) appear to be filing-date artifacts rather than execution misses and should not be read as delivery signal.

Independent corroboration is the weak link. Every capacity figure here traces to company filings or releases. Niagara Falls, Kenilworth, Lancaster, Sweden and Polaris Forge 1 remain at announced status with capacity and cost fields empty; only Polaris Forge 1 carries a number at all (370), and that is contracted rather than energized. No permit records, interconnection queue positions or site reporting appear in the record to date, so the announced pace is currently unfalsifiable from outside.

Through year-end, the checkable item is the 1.85 GW active target (8533, 8281, 8666) including the 150 MW raise (8582), which implies more than doubling active power within one calendar year. Watch third-quarter disclosure for whether that number is reaffirmed or quietly re-based, whether a further DDTL generation follows 5.0 (9121), and whether Indonesia's 2028 360 MW commitment (155, 8758) acquires any site-level or utility-side documentation.

The capacity line has moved in one direction and moved fast. Contracted power went from 1.6 GW in May 2025 (4973) to 3.1 GW by the FY2025 filing (3554, 11770), 3.5 GW in May 2026 (9323), and 4.2 GW by mid-August 2026 (8522, 8262, 8268). Active power followed the same slope (3552, 3553, 9321). What deserves attention is that the largest single step-up occurred across a period in which the company lost its biggest inorganic capacity source: the Core Scientific merger was signed in July 2025 with an October 2025 outside date (4517), extended to April 2026 (4731), and terminated in October 2025 (4166). Contracted capacity roughly doubled anyway, which means the 3.1 GW figure rests on leases and greenfield development rather than acquired megawatts, and should be read as a weaker commitment than the earlier 1.6 GW.

The 2030 goalpost was also redefined rather than missed. The January 2026 framing was 5+ GW of AI factories with NVIDIA (3893, restated at 9329), which by May 2026 had become 8 GW of active power (9322, 8534, 8574). Those are different quantities on different meters, and the substitution was not flagged as a revision. The delayed markers on Marimo, OpenPipe and Monolith (3577, 3578, 3579) look like artifacts of retroactive 10-K dating against earlier close dates rather than genuine slippage; all three appear as completed in the same filing (3548, 3549, 3550).

Independent corroboration is thin in this record. There are no permits, interconnection queue entries, or site reporting here. The only external validation is financing syndication, and it is steady: DDTL 3.0 (4593, 4594), 4.0 (11225), 5.0 (9121), then JPMorgan and MUFG arranging in August 2026 (8736). Capital is arriving on schedule; energization dates are not. Niagara Falls, Kenilworth at 250 MW (7869), Lancaster, Stockholm (7871, 7875) and Polaris Forge 1 at 400 MW (370) all remain announced with no commissioning date disclosed. Only Indonesia carries one, 360 MW by year-end 2028 (8758, 155).

Over the next ninety days the test is arithmetic. Reaching the 1.85 GW year-end 2026 target (8666, 8533, 8281) requires roughly 850 MW from the May 2026 base in seven months, against 490 MW added across all of 2025. Watch whether that target and the 150 MW raise (8582) are reaffirmed at the Q3 print or absorbed silently, and whether any named site produces a utility filing.

The capacity line has only gone one direction, and quickly. Contracted power moved from 1.6 GW (4973) to 3.1 GW (11770, 3554), then 3.5 GW (9323), then 4.2 GW (8522, 8268, 8262) inside roughly fifteen months, while active power climbed 70 to 360 to 850 MW across the 2023-2025 year-ends (3551, 3552, 3553) and crossed 1 GW in May 2026 (9321). Nothing here has shrunk. What has shifted is the framing of the terminal number: the January 2026 commitment was 5-plus GW of NVIDIA AI factories by 2030 (3893, 9329), and by May 2026 the headline had become 8 GW of active power by 2030 (9322, 8534, 8574). Those are not the same metric, and the record does not show the two being reconciled — a reader should treat the 8 GW figure as a new, larger goal rather than a restatement.

Dates have held unevenly. The clean failure is Core Scientific: an October 2025 close (4517) slipped to April 2026 (4731) and then ended in termination (4166). That matters more than the acquisition itself, because it removed the owned-power path to the very capacity curve that kept rising afterward. The subsequent build leans on leases and colocation, explicit in Indonesia (8760), which changes who controls the schedule. The smaller acquisition slippages (3577, 3578, 3579) look like disclosure-timing artifacts. The nearer-term operational promise, 1.85 GW active by year-end 2026 (8666, 8533, 8281, 7888), including a 150 MW raise (8582), implies roughly a gigawatt of additions in a single year against 490 MW the prior year.

On corroboration, the record is thin by its own construction: Niagara Falls, Kenilworth (7869, 7873), Lancaster, Stockholm and Polaris Forge 1 (370) carry announced status with no permits, interconnection queue positions, or site reporting attached. Contracted megawatts are commercial commitments, not energized ones.

Over the next 90 days, the useful checks are whether Q3 disclosure shows active power on a trajectory that can reach 1.85 GW by December 31, whether Kenilworth and Niagara produce visible local permitting and utility queue entries, whether Polaris Forge 1 reports energization against its 400 MW, whether the Indonesia 360 MW by 2028 (8758, 155) names its colocation counterparty, and whether DDTL 5.0 (9121) closes on the terms arranged.

The shape of this record is escalation, not drift. Contracted power moves from 1.6 GW (4973) through 3.1 GW (11770, 3554) and 3.5 GW (9323) to 4.2 GW (8522, 8268), while the 2030 ambition is rewritten from 5 GW (3893) to 8 GW (9322, 8534, 8574) inside roughly three months. The 5 GW figure survives alongside the 8 GW figure on the same day (9329 versus 9322), which suggests the smaller number is the NVIDIA-specific tranche rather than a retraction, but the company has not said so plainly, and readers are left to reconcile it themselves.

Against that, the single hardest-dated commitment in the file failed outright. The Core Scientific merger carried an explicit close of 2025-10-01 (4517), was extended to 2026-04-07 (4731), and was terminated on 2025-10-30 (4166) after the 0.1235 exchange ratio had been restated at least four times (4331, 4314, 4294). That matters beyond optics: Core Scientific was the powered-shell path, and its loss means the subsequent contracted-capacity growth had to be rebuilt through leases and greenfield. The smaller acquisitions show a quieter pattern — OpenPipe due 2025-09-01 (3578), Marimo 2025-10-01 (3577), Monolith 2025-11-01 (3579) all carry slipped dates, yet later disclosure presents them as done (11775, 11776, 10329) without noting the slip.

Independent corroboration is thin because the record is almost wholly self-reported. The strongest external signal is the financing cadence: DDTL 3.0 (4593, 4594), then DDTL 4.0 arrangers (11225), DDTL 5.0 (9121), and JPMorgan and MUFG as bookrunners (8736). Lenders underwriting on roughly a quarterly rhythm is real evidence of construction spend, though it is also mounting leverage. Note that Indonesia (155, 8758, 8760) and Stockholm (7875) are colocation leases, so no CoreWeave permit trail should be expected there; Kenilworth at 250 MW (7869) and Lancaster at 300 MW (7870) are greenfield and should generate municipal approvals and PJM interconnection entries. Their absence would be diagnostic.

Over the next ninety days, the test is arithmetic. Getting from just over 1 GW in May (9321) to 1.85 GW by year-end (8533, 8281), including the 150 MW raise (8582), requires adding more in seven months than in all of 2025 (3552 to 3553). Watch the Q3 print, the Kenilworth and Lancaster filings, Jakarta lease commencement against the 2028 date, and whether "sold out" (1494) and the Anthropic agreement (8571) are quantified.

The capacity line has moved in one direction and moved fast. Contracted power went from 1.6 GW (4973) to 3.1 GW (11770, 3554), then 3.5 GW (9323), then 4.2 GW (8522, 8262, 8268) inside roughly fifteen months, while the active-power series compounds cleanly at 70, 360 and 850 MW for year-ends 2023 through 2025 (3551, 3552, 3553) before crossing 1 GW in May 2026 (9321). Nothing here has shrunk. The one number worth watching for definitional drift is 2030: the 5+ GW figure (3893, 9329) is scoped to the NVIDIA AI-factory framework, while 8 GW (9322, 8574, 8534, 7889) is the total active-power ambition, and both were stated on the same day. They are not in conflict, but they are easy to conflate, and the company has not been forceful about separating them.

Dates have mostly held where they were specific, with two exceptions. The Core Scientific merger slipped from an October 2025 close (4517) to April 2026 (4731) and was then terminated outright (4166) — a material change to the capacity base that the subsequent contracted-power increases absorbed without narrative acknowledgement. Separately, the OpenPipe, Marimo and Monolith closings carry September, October and November 2025 dates (3588, 3587, 3589) that surfaced only in the March 2026 annual filing and the February 2026 announcements (11775, 11776); that is retroactive disclosure rather than slippage, and should not be scored as a missed promise.

Independent corroboration is structurally limited here because the model is increasingly leased colocation — explicitly so in Indonesia (8760) and Stockholm (7875). Interconnection queue entries and permits will sit with landlords, not CoreWeave, so Lancaster (7874) and Kenilworth are the only US sites where county filings and PJM queue positions offer a direct check. Note also that Lancaster's headline 300 MW is the option, not the commitment; the firm number is 100.

Over the next ninety days the binding test is the 1.85 GW year-end 2026 target (8533, 8281, 7888) plus its 150 MW raise (8582), which requires more additions in one half-year than all of 2025 delivered. Also watch financing cadence — DDTL 3.0 through 5.0 arrived in under a year (4593, 11225, 9121) — the naming of the Indonesian colocation partner, and any siting or date detail for Polaris Forge 1 (370) and Kenilworth, which remain capacity figures without geography or schedule.

The capacity line has moved in one direction and quickly. Contracted power ran from 1.6 GW (4973) to 3.1 GW by the FY2025 filing (3570, 11770), 3.5 GW three months later (9323), and 4.2 GW by mid-August (8522, 8262). Active power followed a similar curve, 70 MW to 360 MW to 850 MW across year-ends (3551, 3552, 3553), crossing 1 GW in May (9321) against a 1.85 GW year-end target (8533, 8281), which itself was revised up by 150 MW mid-year (8582). The 2030 anchor was rebased from 5+ GW (3893) to 8 GW (9322, 8534, 7889) without the earlier figure being retired; the NVIDIA-scoped 5 GW is still restated alongside it (9329), so the two numbers now circulate as if independent.

The date record is weaker than the capacity record. The Core Scientific merger slipped from an October 2025 close (4517) to April 2026 (4731) and was then terminated outright (4166) — yet the contracted-capacity step from 1.6 GW to 3.1 GW brackets that failure without the company reconciling how much of the earlier pipeline assumed those sites. Three tuck-ins were disclosed retroactively with completion dates already past: OpenPipe at September 2025 (3588), Marimo at October (3587), Monolith at November (3589), and the same transactions were re-announced in February and again in April (11775, 11776, 10329), which inflates apparent cadence without adding assets.

The independent trail is thin in this record. Nothing here contains a permit, interconnection position, or site visit; every capacity figure is company-sourced. That matters most where delivery is leased rather than built — Indonesia is explicitly colocation (8760), Stockholm is two colocation campuses (7875) — so the verifiable queue entries sit with landlords, not CoreWeave. Indonesia's 360 MW by end-2028 (155, 8758) equals the company's entire 2024 global footprint (3552), an unusually large first move into a new region announced only weeks ago (3013).

Over the next ninety days, the Q3 print is the test: active power against 1.85 GW, given roughly 850 MW needed in under five months. Watch whether 4.2 GW is restated upward again before the year-end number lands, whether the Indonesian colocation counterparty is named, and whether DDTL 4.0 and 5.0 close on terms (11225, 9121, 8736) — financing cadence, not demand, has been the binding constraint on this build.

CoreWeave's capacity story has moved in one direction, and quickly. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553), then passed 1 GW by May 2026 (9321). Contracted power grew faster: 1.6 GW in May 2025 (4973), 3.1 GW by February 2026 (11770, 3570), 3.5 GW in May (9323) and 4.2 GW by August (8522, 8466). The long-range target escalated in the same way. A January figure of 5+ GW of AI factories with NVIDIA by 2030 (3893) became 8 GW of active power by 2030 in May (9322) and was restated through August (8534, 7889). None of the visible targets has been cut. The year-end 2026 goal of 1.85 GW was itself nudged up by 150 MW (8582, 8666).

The date record is less tidy, though most of the slippage sits outside facilities. The Core Scientific merger was reaffirmed repeatedly across July to September 2025 (4729, 4313, 4294) and then terminated (4166). That removed an owned-power path, and the later capacity announcements lean heavily on leased and colocation capacity instead. Indonesia is explicitly leased from a colocation provider (8760), Stockholm is two colocation campuses (7875), and Polaris Forge 1 is a 400 MW contract on a third party's campus (370). The DDTL 3.0 commitment termination date carries a delayed flag (4593) without a clear public explanation in the record. The facility-level 2030 targets have not yet been tested.

Independent corroboration is the weak point. This record contains no permits, interconnection-queue entries or site reporting for any of the five named projects. Kenilworth and Lancaster appear only as announced, with no capacity, cost or date attached. The Indonesia claim of 360 MW across three facilities by end-2028 (155, 8758) names neither the landlord nor the energization schedule. The pace is backed mainly by financing activity (DDTL 4.0 and 5.0 arrangers, 11225, 9121, 8736) and customer demand claims (1494, 8571), not by physical evidence. The source list was truncated after the Stockholm entries, so later claims may change this picture.

In the next 90 days, the key test is the Q3 report. Reaching 1.85 GW by year-end means adding roughly 800 MW in about seven months, a faster rate than any prior period. Also watch whether contracted power keeps rising while active power lags, whether the Indonesia colocation partner and phasing are disclosed, and whether PJM queue or local zoning filings emerge for Lancaster and Kenilworth.

CoreWeave's capacity story has two curves, and the gap between them is the main thing to track. Contracted power has grown faster than delivered power. It went from 1.6 GW in May 2025 (4973) to 3.1 GW in the annual filing (3570, 11770), then 3.5 GW in May 2026 (9323) and 4.2 GW by mid-August (8522, 8268). Active power has also grown quickly, from 70 MW to 360 MW to 850 MW across year-ends 2023 to 2025 (3551, 3552, 3553), and it passed 1 GW by May 2026 (9321). Still, contracted power now runs roughly four times active. The long-range framing has moved too. The January 2026 NVIDIA language spoke of 5+ GW of AI factories by 2030 (3893, reiterated in 9329). A separate 8 GW active-power target for 2030 now appears alongside it (9322, 8534, 7872), and the record does not reconcile the two figures. The near-term signal points upward. The year-end 2026 active target stands at 1.85 GW (8666, 8281), and 8582 describes this as a 150 MW increase over the prior goal. Reaching it implies adding roughly 800 MW in the back half of 2026, which would be the company's fastest delivery pace yet.

Dated commitments have a mixed record, but most of the slippage is corporate rather than physical. The Core Scientific merger missed its targets (4517, 4731) and was terminated (4166). The Marimo, Monolith and OpenPipe closings are flagged as late against their 2025 dates (3577, 3579, 3578). The DDTL 3.0 commitment window also shows a delay (4593). None of these delays was acknowledged as a slip in the claims themselves. At the facility level there is little to test yet. The only dated project commitment is Indonesia's 360 MW by end-2028 (155, 8758). Lancaster (7870), Stockholm (7871, 7875) and Polaris Forge 1 (370) carry capacity figures but no in-service dates, and the Kenilworth claim (7869) is truncated in this record.

No independent corroboration exists in this dataset. There are no permits, no PJM or utility queue entries, and no site reporting. Much of the footprint is leased, including Indonesia via a colocation provider (8760) and Polaris Forge 1 from its developer, so the evidence of physical progress will sit with landlords rather than CoreWeave.

Over the next 90 days, the Q3 report is the key checkpoint. The question is whether active power is tracking toward the 1.85 GW target (8533) or whether that target quietly softens. It also matters whether the contracted figure keeps rising faster than delivery. Other items to watch are the naming of the Indonesia colocation partner and its permits, PJM queue visibility for Lancaster and Kenilworth, energization milestones at Polaris Forge 1, and drawdowns on the DDTL 5.0 facility (9121).

CoreWeave's capacity story has grown on two separate tracks, and the gap between them is what matters. Contracted power has more than doubled in fifteen months, from 1.6 GW (4973) to 3.1 GW (11770, 3554), then 3.5 GW (9323), then 4.2 GW by mid-August 2026 (8522, 8466). Active power has also grown quickly, from 70 MW to 360 MW to 850 MW across 2023-2025 (3551-3553) and past 1 GW by May 2026 (9321). But active capacity still covers only about a quarter of what is contracted. The 2030 ambition has also changed shape. It began as a 5+ GW figure tied to NVIDIA (3893, 9329) and has since been restated as 8 GW of active power (9322, 8534, 7872). No reconciling bridge between the two figures appears in the record. The year-end 2026 target of 1.85 GW (8666, 8533, 8281) was raised by 150 MW (8582). Hitting it means adding roughly 800 MW in the back half of 2026, against about 490 MW added in all of 2025.

On dates, the facility record is thin, because most capacity claims carry no promised-by date. The clearest reversal is corporate rather than physical. The Core Scientific merger was guided to an October 2025 close (4517) and was terminated instead (4166), which removed an owned-power path the company had leaned on. The delayed flags on Marimo, OpenPipe and Monolith (3577-3579) look like closing-date bookkeeping, not missed commitments, since later filings treat those deals as done (10329). The DDTL 3.0 commitment window was extended (4593), and DDTL 4.0 and 5.0 followed in quick succession (11225, 9121, 8736). That pattern suggests financing is being rolled forward to match a build schedule that keeps expanding.

Independent corroboration is largely absent from this record. The site-level figures are Indonesia's 360 MW across three leased colocation facilities by 2028 (8758, 155, 8760), Polaris Forge 1 at 400 MW (370), and Lancaster at 300 MW (7870). None of them is paired with permits, utility interconnection filings or on-the-ground reporting. The Stockholm claims (7871, 7875) carry no capacity figure at all, and the Kenilworth entry (7869) is truncated in the data provided. Because much of the Asian and European footprint is leased, delivery depends on the landlords' schedules rather than CoreWeave's own.

Over the next 90 days, the third-quarter print is the key test. It should show whether active power is on a path to 1.85 GW or whether the target is quietly reset. Beyond that, watch for PJM queue and local zoning activity at Lancaster and Kenilworth, for identification of the Jakarta colocation partner and its permits, and for whether the drawdown pace on DDTL 5.0 tracks the energization the company is claiming.

CoreWeave's capacity story has grown faster on paper than on the ground, though the ground-level numbers are still moving quickly. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553), and passed 1 GW by May 2026 (9321). Contracted power has grown faster. It stood at 1.6 GW in May 2025 (4973), reached 3.1 GW by early 2026 (11770, 3554), 3.5 GW in May (9323), and 4.2 GW by August (8522, 8466). The gap between contracted and energized capacity is therefore widening, not closing, from roughly 1.2 GW to more than 3 GW. The 2030 framing has also shifted. A 5+ GW figure tied to NVIDIA (3893, 9329) now sits beside a broader 8 GW active target (9322, 8534, 7872). These are not strictly contradictory, but the headline number has changed without a reconciliation.

On dates, the facility record offers little to test, because most capacity claims carry no promised-by date. The firm near-term marker is 1.85 GW active by year-end 2026 (8666, 8533, 8281). Claim 8582 describes this as a 150 MW upward revision. Hitting it requires energizing roughly 850 MW between May and December, which is more than the entire 2025 increment in about two-thirds of the time. The slippage that is documented sits on the corporate side. The Core Scientific merger was flagged against an October 2025 date (4517) and then terminated (4166), removing a path to owned, powered sites. Several tuck-in acquisitions also carry delayed flags (3577, 3578, 3579), although those look like close-timing artifacts rather than capacity misses.

The dataset contains no independent corroboration of any kind: no permits, interconnection queue positions, or site reporting. The newer projects also lean on third parties. Indonesia is a leased colocation arrangement for 360 MW by end-2028 (8760, 8758). Stockholm is colocation as well (7875). Polaris Forge 1's 400 MW is framed as contracted rather than built (370). Lancaster's 300 MW (7870) and the Kenilworth greenfield campus (7869, whose record is truncated here) are the projects where public permitting and PJM queue filings should eventually appear.

Over the next 90 days, the third-quarter print is the key checkpoint. Active capacity would need to be well above 1.3 GW to keep 1.85 GW credible, and any quiet softening of that target would matter more than another contracted-capacity increase. Other things to watch are the naming of the Indonesian colocation partner, municipal or utility filings for Lancaster and Kenilworth, and drawdown activity under the DDTL 5.0 facility (9121) as a proxy for actual build spend.

CoreWeave's announced capacity has scaled steadily and in one direction. Contracted power went from 1.6 GW in May 2025 (claim 4973) to 3.1 GW by early 2026 (claims 11770, 3559), 3.5 GW in May (9323) and 4.2 GW by mid-August (8522, 8268). Active power followed a consistent reported path: 70 MW at end-2023, 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3569), then past 1 GW by May 2026 (9321). The long-range framing also grew. The earlier 5+ GW of AI factories with NVIDIA by 2030 (3893, 9329) now sits alongside a broader 8 GW active-power target (9322, 8534, 7868). The two are not strictly comparable, since the first is partner-scoped, but the headline number the company leads with has risen by more than half. The one clear contraction was the terminated Core Scientific merger (4166), which removed a path to owned capacity that had been repeated across five filings.

On dates, most facility claims carry no delivery date, which limits accountability. The year-end 2026 target of 1.85 GW (8666, 8533, 8281) is the only near-term dated commitment. Claim 8582 describes a 150 MW increase to it, which implies an earlier, lower figure that does not appear in this record. Meeting 1.85 GW means adding roughly 800 MW in about seven months, more than the roughly 490 MW brought online in all of 2025. The DDTL 3.0 commitment date (4593) and the Core Scientific timeline (4731, 4517) are both flagged delayed. Neither was framed publicly as a slip.

This record contains no independent corroboration. There are no permits, no interconnection-queue entries and no site reporting for Lancaster (300 MW, 7870), Kenilworth (250 MW, 7869), Stockholm (7871, unsized) or the three leased Indonesia sites (8758, 8760), which have a 2028 horizon. Polaris Forge 1 (370) is contracted capacity at a third party's campus, so energization evidence would come from that operator. The claim list as supplied also appears truncated at 7872.

Over the next 90 days, the Q3 report is the key checkpoint: whether active power has moved meaningfully above 1 GW, and whether 1.85 GW is reaffirmed or quietly softened. Other items to watch are PJM queue positions and local land-use filings for Lancaster and Kenilworth, identification of the Indonesian colocation partner, and energization milestones at Polaris Forge 1.

CoreWeave's announced capacity has only grown. Contracted power was 1.6 GW in May 2025 (claim 4973), 3.1 GW by early 2026 (claims 11770, 3559), 3.5 GW in May (claim 9323) and 4.2 GW by mid-August (claims 8522, 8268). That is roughly 2.6x in fifteen months. Active capacity has also grown, from 360 MW at end-2024 (claim 3552) to 850 MW at end-2025 (claim 3569) and past 1 GW by May (claim 9321). The long-range framing has shifted, though. The 5+ GW NVIDIA-linked figure for 2030 (claims 3893, 9329) now sits beside an 8 GW active-power goal (claims 9322, 8534, 7868). The newer goal is almost double the contracted base. Treat it as an aspiration rather than a pipeline.

The company's own timelines are uneven. The year-end 2026 target of 1.85 GW (claims 8666, 8533, 8281) was restated consistently through August. Claim 8582 describes a 150 MW increase to that target, which suggests it was raised rather than cut. The slips in the record are corporate, not physical. The Core Scientific merger, first slated for October 2025 (claim 4517), was terminated (claim 4166). That removed what would have been a large owned-power pathway. Filings later labelled the Marimo, OpenPipe and Monolith deals as delayed against their original dates (claims 3577 through 3589). Those revisions were recorded without much narrative explanation.

Independent evidence is thin. The site-level figures of 400 MW at Polaris Forge 1 (claim 370), 300 MW at Lancaster (claim 7870) and 250 MW at Kenilworth (claim 7869) carry no delivery dates or costs. Stockholm carries no capacity at all (claim 7871). No permits, utility agreements or interconnection-queue entries appear in the record. Indonesia's 360 MW by 2028 (claims 8758, 155) depends on leased colocation (claim 8760) with an unnamed provider. That dependency moves execution risk onto a third party.

For the next 90 days, the Q3 print is the key test. Reaching 1.85 GW means adding roughly 800 MW of active power in about seven months, a faster pace than 2025 delivered. Watch whether reported active capacity is on that line or the year-end figure gets quietly reframed. Watch for PJM queue or municipal filings for Kenilworth and Lancaster, since the greenfield New Jersey site is the most exposed to permitting drag. Also watch for disclosure of the Indonesia colocation partner and energization schedule. The claim record supplied here ends mid-entry at claim 7872, so any later restatement is not reflected in this read.

CoreWeave's capacity story has grown on every axis at once, and the growth in commitments is outpacing the growth in delivered power. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (claims 3551-3553, 11769). The company then reported crossing 1 GW in May 2026 (9321) and set a year-end 2026 target of 1.85 GW (8666, 8533, 8281). Claim 8582 describes that target as a 150 MW increase, so management raised the bar mid-year rather than hedging it. Contracted power has compounded faster: 1.6 GW (4973), then 3.1 GW (11770, 3554), 3.5 GW (9323), and 4.2 GW by August 2026 (8522, 8466). The 2030 figure moved from 5+ GW of NVIDIA-linked AI factories (3893, 9329) to 8 GW of active power (9322, 8574, 7868). The filings do not reconcile the two numbers, and readers should not assume one replaced the other.

Facility dates are hard to test because most capacity claims carry no deadline. The misses on record are transactional rather than physical. The Core Scientific merger slipped past its 2025-10-01 target (4517) and was terminated (4166), and the DDTL 3.0 commitment window is flagged as delayed (4593). That termination matters for facilities, because acquiring Core Scientific would have brought owned powered sites. Without it, CoreWeave remains dependent on leases, as the Indonesia arrangement shows: 360 MW across three Greater Jakarta sites by 2028, leased from a colocation provider (8758, 155, 8760, 8670). The 400 MW Polaris Forge 1 figure (370) is likewise contracted capacity at a third-party campus, not a CoreWeave build.

The record contains no independent evidence for any named project: no permits, no interconnection queue positions, and no site reporting. The 250 MW Kenilworth greenfield campus (7873, 7869) is the claim most in need of it, because a greenfield site in PJM territory is where timelines usually slip. Lancaster and the Stockholm colocation (7871) have no stated capacity or dates. The claim list supplied here ends at 7872, so later entries may change this picture.

The arithmetic is demanding. Reaching 1.85 GW means adding roughly 850 MW in the second half of 2026, more than CoreWeave delivered in all of 2025. Over the next 90 days, watch the Q3 print for the active-power figure and whether the 1.85 GW target is reaffirmed or quietly reframed. Watch Kenilworth for local land-use filings and a PJM queue entry, and look for the Indonesia landlord to be named. Watch Applied Digital's energization disclosures for Polaris Forge 1. Finally, look for any statement explaining how the 5 GW and 8 GW 2030 figures relate.

CoreWeave's capacity story has only moved in one direction, and it has moved fast. Active power grew from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553). The company said it passed 1 GW in May 2026 (9321). Contracted power grew faster still: 1.6 GW in May 2025 (4973), 3.1 GW in February 2026 (11770, 3554), 3.5 GW in May (9323) and 4.2 GW by mid-August (8522, 8268). The long-range target has also grown. It was framed as 5+ GW of NVIDIA-linked AI factories by 2030 (3893, 9329) and is now 8 GW of active power by 2030 (9322, 8534, 7868). The two figures have been cited side by side, which blurs whether 8 GW replaces the earlier number or includes it. The year-end 2026 target of 1.85 GW (8666, 8533) came with a 150 MW upward revision (8582). So the guidance has been raised, not cut.

The record shows no outright facility-date slips, but that is partly because few site-level claims carry dates at all. The one significant reversal was the Core Scientific merger. It was expected by October 2025 (4517), and the termination was announced on 2025-10-30 (4166). That removed a self-owned capacity pathway without any restated capacity plan following it. The DDTL 3.0 commitment date is flagged as delayed (4593) with no explanation on record. Of the named projects, only Indonesia has a deadline: 360 MW across three leased Greater Jakarta sites by end-2028 (8758, 155, 8670, 8760). Kenilworth's 250 MW (7873), Stockholm (7871) and Polaris Forge 1's 400 MW (370) have no dates, and Lancaster has no quantified claim in the visible record.

No independent corroboration appears in this file: no permits, no interconnection-queue positions, no site reporting. The supplied claim list is also truncated after 7872. The near-term math is demanding. Reaching 1.85 GW means adding roughly 850 MW in the second half of 2026, far more than the 490 MW added in all of 2025. Contracted capacity also covers only about half of the 8 GW goal.

Over the next 90 days, watch whether the Q3 report shows active power on a path toward 1.85 GW or quietly redefines the metric. For Kenilworth, look for municipal site-plan approvals and PSE&G or PJM interconnection filings. For Indonesia, watch for the colocation partner to be named and a phase schedule. Watch too for any capacity figure for Stockholm or Lancaster.

CoreWeave's capacity story has grown faster on paper than on the ground, though the ground numbers are also moving quickly. Contracted power roughly went from 1.6 GW in May 2025 (claim 4973) to 3.1 GW in February 2026 (11770), 3.5 GW in May (9323) and 4.2 GW by mid-August (8522). That is more than a 2.5x increase in fifteen months, and none of the steps was ever revised down. Active power has also risen. The company reported 360 MW at year-end 2024 (3552) and 850 MW at year-end 2025 (3553). It said active power passed 1 GW in May 2026 (9321), and the 2026 year-end target now stands at 1.85 GW (8666). That target was raised by 150 MW in August (8582). The longer-range goal has changed in a way the company has not reconciled. The NVIDIA-linked figure of 5+ GW by 2030 (3893, 9329) now appears alongside an 8 GW active-power target for the same year (9322, 8534, 7868), and CoreWeave has not explained how the two relate. Many of these entries repeat the same filing figures and should not be read as separate confirmations.

On dates, no facility-capacity deadline has visibly slipped, largely because almost none carry one. The exceptions are the 2026 year-end target and Indonesia's 360 MW by end-2028 (8758, 155). The clearest miss is structural rather than a slipped schedule. The Core Scientific merger was flagged as delayed (4517, 4731) and then terminated (4166). That removed an owned-capacity route, and the company has not publicly tied the termination to any change in its targets. The 'delayed' flags on OpenPipe, Marimo and Monolith (3577–3589) concern software deals and say nothing about data-center delivery.

Independent corroboration is thin. The record contains no permits, interconnection-queue positions or site reporting for Kenilworth (7873), Stockholm (7871), Lancaster or Indonesia. Indonesia and Stockholm are leased colocation arrangements (8760, 7871), so their timing depends on the landlords. Polaris Forge 1's 400 MW (370) is likewise contracted capacity at a third party's campus. Kenilworth, a 250 MW greenfield site in New Jersey, is the project where public planning and PJM grid-queue records should exist, and so far they have not been cited.

Over the next 90 days, the third-quarter print is the main test. Reaching 1.85 GW means adding roughly another 800 MW of active power in the second half of the year, which would be a pace above any prior half-year. Beyond that, the items to track are whether the 2030 target settles on 5 GW or 8 GW, whether Kenilworth zoning or interconnection filings appear, whether the Indonesia colocation provider is named, and how much of the DDTL 4.0 and 5.0 financings (11225, 9121, 8736) is actually drawn against specific sites.

CoreWeave's capacity record has only grown. Active power went from 70 MW at end-2023 to 360 MW at end-2024 and 850 MW at end-2025 (3551, 3552, 3553). The company said it passed 1 GW by May 2026 (9321). Contracted power has grown faster: 1.6 GW in May 2025 (4973), 3.1 GW in February 2026 (11770), 3.5 GW in May (9323) and 4.2 GW by mid-August (8522). The 2030 goal has also moved. It was framed as 5+ GW of AI factories under the NVIDIA framework (3893, 9329). It is now 8 GW of active power (9322, 8574). These may be different scopes rather than a revision, but the company has not reconciled them. As contracted capacity grows faster than energized capacity, more of the story depends on future delivery.

The near-term dates hold up better than the M&A dates. The year-end 2026 active target of 1.85 GW (8666, 8533) was raised by 150 MW in August (8582), which points to confidence rather than slippage. Reaching it requires roughly 850 MW of additions in 2026, about equal to all of 2025's growth. The clearer failure is Core Scientific. It was promised for October 2025 (4517), and the agreement was terminated (4166). That removed an owned-power path the capacity plan had seemed to lean on. The delayed flags on Marimo, OpenPipe and Monolith (3577 through 3589) sit awkwardly beside later statements that treat those deals as done (10329). They look more like date-parsing artifacts than real misses. The DDTL 3.0 commitment termination date (4593) is also flagged delayed and deserves a check against the latest credit filings.

The project-level claims are thin on specifics. Indonesia is described as 360 MW across three leased Greater Jakarta sites by end-2028 (8758, 155, 8670). The colocation provider is not named (8760). Polaris Forge 1 carries 400 MW contracted (370) with no energization date. Kenilworth is a 250 MW greenfield campus (7873), and Stockholm (7871) and Lancaster have no capacity or timing on record. The record contains no permits, no interconnection queue positions and no site reporting. The announced pace is therefore unverified, not contradicted.

Over the next 90 days, the Q3 print is the key test. Active capacity near 1.4 GW would keep 1.85 GW believable, and a flat figure with a reaffirmed target would suggest a back-loaded fourth quarter. Look for Kenilworth and Lancaster in the PJM queue and in local site-plan approvals, a named Indonesian colocation partner, and any restatement that aligns the 5 GW and 8 GW figures. This review was built on a truncated claim list, so later entries may change these findings.

CoreWeave's capacity numbers have only moved up, and the reported history is internally consistent. Active power went from 70 MW at end-2023 (3551) to 360 MW at end-2024 (3552), 850 MW at end-2025 (3553, 11769), and more than 1 GW by May 2026 (9321). The year-end 2026 target is 1.85 GW (8666, 8533, 8281), which includes a 150 MW upward revision (8582). Contracted power rose faster, from 1.6 GW in May 2025 (4973) to 3.1 GW (11770, 3554), then 3.5 GW (9323), then 4.2 GW by August 2026 (8522, 8466). The 2030 framing has drifted. A 5+ GW figure tied to NVIDIA (3893, 9329) now sits beside an 8 GW active-power goal (9322, 8534, 7868). The two may measure different things, but the company has not reconciled them in the record. Reaching 8 GW from 1.85 GW requires about 1.5 GW of energization a year, and current contracts cover only about half of that endpoint.

On dates, the capacity milestones that can be checked have held. The slips are corporate rather than physical. The Core Scientific deal missed its October 2025 target (4517) and was terminated openly (4166). The Marimo, OpenPipe and Monolith closings ran a few months behind their flagged dates (3577-3579) before being confirmed (11776, 11775). The DDTL 3.0 commitment window passed its flagged date (4593), and DDTL 4.0 and 5.0 followed (11225, 9121, 8736). We found no silent revision of a capacity date.

The weakness is at the project level. None of the five named projects has a capacity or cost field populated. The site figures in the record are 360 MW across three leased Indonesian facilities by end-2028 (8758, 155, 8760), a 250 MW greenfield campus at Kenilworth (7873, 7869), and 400 MW contracted at Polaris Forge 1 (370). Stockholm (7871) and Lancaster have no figures attached. The record contains no permits, interconnection queue positions or on-the-ground reporting, so the announced pace rests entirely on company disclosure. Most of the pipeline is also leased or third-party developed, which moves execution risk onto partners the record does not name.

Over the next 90 days, the Q3 print should show whether active capacity is on a credible path to 1.85 GW. It should also show whether contracted power moves past 4.2 GW and whether the 5 GW and 8 GW targets are reconciled. The strongest independent tests are local zoning and utility or PJM interconnection filings for Kenilworth and Lancaster, and naming of the Indonesian colocation partner with a first-phase energization date. Draw activity under DDTL 5.0 would confirm that the claim of sold-out near-term capacity (1494) is turning into funded builds.

CoreWeave's announced capacity has expanded substantially while execution discipline has wavered. Contracted power capacity grew from 1.6 GW in May 2025 (claim 4973) to 4.2 GW by August 2026 (claims 8522, 8262)—a 163 percent increase in fifteen months. Active power capacity has followed a more measured path: 70 MW end-2023 (claim 3551), 360 MW end-2024 (claim 3552), 850 MW end-2025 (claim 3553), and surpassing 1 GW by May 2026 (claim 9321). This trajectory aligns with publicly stated quarterly targets, but the gap between contracted and active capacity widened throughout 2026, suggesting either customer demand outpaced deployment or timelines shifted without formal revision.

M&A plans faltered materially. The Core Scientific merger, promised closure by April 2026 (claim 4731), terminated in October 2025 (claim 4166). Acquisitions of Marimo, OpenPipe, and Monolith all missed Q4 2025 targets (claims 3577, 3588, 3589). Instead, capacity growth now relies on facility partnerships: Indonesia's three-facility, 360 MW commitment targets end-2028 (claims 8758, 155) via leased colocation (claim 8760), not acquisition. Kenilworth's 250 MW greenfield campus (claim 7869) and Swedish sites (claim 7871) follow similar models, with recent financing from JPMorgan and MUFG (claim 8736) supporting deployment.

The 1.85 GW active power target for year-end 2026 (claims 8533, 8666, 8281) arrives in approximately 3.5 months and offers near-term verification. Interconnection queue status with regional grid operators, executed power purchase agreements, and permit filings for Kenilworth and Indonesian facilities should confirm whether claimed capacity is funded and under construction. The Indonesia project's 2028 deadline requires demonstration of land access, local power authority agreements, and colocation vendor commitments. Watch for whether CoreWeave reaches 1.85 GW active by year-end, how capacity is distributed across announced geographies, and whether any 2027-2028 deliverables are revised downward without acknowledgement.

CoreWeave has demonstrated aggressive capacity expansion targets while managing persistent execution delays. The company's announced active power capacity nearly doubled from 420 MW in May 2025 to 850 MW by year-end 2025, a credible growth rate that continued to surpass 1 GW by May 2026 and targeted 1.85 GW by December 2026 (claims 4972, 11769, 9321, 8666). However, the trajectory reveals a consistent pattern: contracted capacity outpaces deployed capacity by widening margins. Contracted power jumped from 1.6 GW (May 2025) to 4.2 GW (August 2026)—a 2.6x expansion—while active capacity grew roughly 3.5x in the same window (claims 4973, 8522). This gap signals either accelerating deployment in later phases or optimistic contracted projections that may not materialize.

The company's merger strategy merits scrutiny. The Core Scientific acquisition consumed strategic attention throughout 2025, with promised close dates slipping from October to April 2026 before outright termination in October (claims 4731, 4517, 4166). Simultaneously, three smaller acquisitions—Marimo, OpenPipe, and Monolith—missed their original 2025 deadlines and resurfaced in March 2026 announcements without revised timelines (claims 3577, 3578, 3579). These execution setbacks raise questions about CoreWeave's M&A integration capacity when pursuing transformative partnerships like the NVIDIA capacity commitment and Anthropic's multi-year agreement.

On the facility side, only the Indonesia expansion carries a specific near-term commitment: 360 MW across three Greater Jakarta sites by December 2028 (claims 8758, 155). This stands as one of few timeline-bound projects with named capacity. Polaris Forge 1 carries a 400 MW contracted label but no deployment date (claim 370). Kenilworth Campus shows 250 MW without timeline specificity (claim 7869). The 8 GW target by 2030 appears anchored primarily to these unspecified facilities rather than detailed project schedules (claims 8534, 9322).

The recent DDTL 3.0 termination date slipped from July 2026 without public acknowledgement (claim 4593). Over the next 90 days, CoreWeave's credibility hinges on delivering the 1.85 GW year-end 2026 target and clarifying whether Indonesia facilities remain on track. Permit filings, interconnection queue positions, and infrastructure reporting will reveal whether the gap between contracted and active capacity narrows as promised.

CoreWeave has aggressively expanded its capacity announcements over 16 months. Contracted power capacity increased from 1.6 GW (May 2025, claim 4973) to 4.2 GW (August 2026, claim 8522), while active operational capacity grew from 420 MW (May 2025, claim 4972) through 850 MW (December 2025, claim 3553) to a target of 1.85 GW by year-end 2026 (claim 8533). The announced 2030 capacity target has also expanded: initial "5+ GW" guidance in January 2026 (claim 3893) shifted to 8 GW active capacity by August (claim 8534), suggesting ambition has accelerated faster than near-term buildout.

Execution risk is evident in the acquisition record. The Core Scientific merger, promised for October 2025 (claim 4517), was terminated by October 2025 with no disclosed rationale (claim 4166). Three software acquisitions—Marimo, OpenPipe, and Monolith—all slipped from 2025 targets to "stated" status by March 2026 without revised dates. Only Weights & Biases appears to have closed. These delays suggest the company is stretched across parallel hardware buildout and M&A, raising questions about execution bandwidth.

The gap between contracted and active capacity is widening. CoreWeave claims 4.2 GW contracted but surpassed only 1.0 GW active by May 2026 (claim 9321), implying a 7–12 month deployment lag. The company stated "near-term capacity sold out" in August 2026 (claim 1494) while simultaneously targeting 1.85 GW active by year-end (claim 8533), requiring 0.85 GW in 120 days—approximately 2× the 2025 deployment rate. No permits, grid interconnection filings, or third-party construction milestones have been disclosed; all statements are corporate announcements.

Geographic expansion appears nascent. Indonesia's 360 MW facility is promised for end-2028 (claim 155) but was announced only in August 2026 with no disclosed permitting or closed funding. Kenilworth (250 MW, claim 7869) and Lancaster (300 MW, claim 7870) sites are named but lack timelines.

Watch for: Q4 2026 commissioning data to validate the 1.85 GW target, public permits or grid operator filings for any of the five geographies, interconnection queue positions, and disclosure of which contracted capacity is backed by signed customer agreements.

CoreWeave's power capacity trajectory shows acceleration in contracted commitments alongside execution gaps. The company reported 70 MW deployed at end-2023, scaling to 850 MW by end-2025, representing a 12-fold increase in eighteen months. However, the gap between contracted and active capacity widened significantly. By May 2026, CoreWeave claimed to have surpassed 1 GW active power (claim 9321) while contracted capacity had grown to 3.5 GW (claim 9323)—a three-to-one ratio. By August 2026, contracted capacity reached 4.2 GW (claim 8522) against an end-of-year 2026 active deployment target of 1.85 GW (claims 8666, 8533, 8281), suggesting the contracted-to-active ratio remains roughly 2.3:1 as of today.

Timeline adherence has been inconsistent. The Core Scientific merger, announced in May 2025 and promised for October 2025 (claim 4517), was terminated outright by October 30, 2025 (claim 4166), representing a material strategic reversal rather than a simple delay. Three smaller acquisitions—Marimo, OpenPipe, and Monolith—were promised for September through November 2025 (claims 3577-3579) but appear unresolved as of March 2026 claims (claims 3587-3589), with no subsequent confirmation through August 2026. The DDTL 3.0 commitment termination, promised by July 31, 2026 (claim 4593), is marked delayed.

Recent messaging shows CoreWeave has shifted toward geographically distributed build-out. The August 2026 Indonesia announcement (claims 8758, 155) commits 360 MW across three facilities with a 2028 year-end deadline. U.S. campus announcements include Kenilworth, New Jersey (250 MW, claim 7873) and Lancaster, Pennsylvania (300 MW potential, claim 7874). The 8 GW target by 2030 (claims 8574, 8534, 7868, 8534, 7889) has become consistent messaging across multiple recent announcements, though the path from current 1+ GW active to that goal remains heavily contingent on execution at five announced locations simultaneously.

Watch for: (1) quarterly active capacity disclosures through Q4 2026 to confirm the 1.85 GW year-end target; (2) Indonesia facility construction permits and interconnection queue entries to validate the 2028 timeline; (3) any further delays or terminations in the DDTL facility financing, given the July miss; (4) whether contracted capacity growth moderates or continues outpacing deployment.

CoreWeave's operational capacity footprint has grown from 70 megawatts at year-end 2023 to 1.0 gigawatt by May 2026, a 14-fold expansion over two years. Year-end 2025 brought 850 megawatts of claimed active capacity (claims 3551–3553). However, the company's announced contracted capacity has scaled more aggressively: 1.6 gigawatts in May 2025, 3.1 gigawatts by March 2026, expanding to 4.2 gigawatts by August 2026 (claims 4973, 3554, 8522). The ratio of contracted-to-active capacity has widened from roughly 1.9:1 in May 2025 to 4.2:1 in August 2026, indicating either aggressive forward contracting, deployment constraints, or timing mismatches between customer commitments and power availability. Timeline execution has been uneven. Multiple software acquisitions—OpenPipe, Marimo, and Monolith—carried specific 2025 completion dates that shifted without acknowledgment (claims 3577–3579, 3587–3589). The Core Scientific merger, a more material expansion attempt, initially targeted October 1, 2025 but was terminated in October after delays (claims 4517, 4166). Facility financing arrangements show similar delay patterns through mid-2026, suggesting operational execution or financial constraints. The company's December 31, 2026 active capacity target of 1.85 gigawatts requires 850 megawatts of new deployment in the final three quarters of 2026 (claims 8666, 8281). This depends primarily on facilities lacking announced construction dates or regulatory filings: Kenilworth, New Jersey (250 megawatts), Lancaster, Pennsylvania (300 megawatts), three Indonesian facilities (360 megawatts by 2028), and Stockholm campuses (claims 7869, 7870, 8758, 7875). No permits, interconnection queue positions, or on-the-ground reports are cited. The pattern combines ambitious future targets—8.0 gigawatts by 2030 (claim 7889)—with near-term execution gaps. Over the next 90 days, watch whether CoreWeave delivers the 1.85 gigawatt year-end result; a material miss would indicate systematic overestimation of deployment velocity and undermine confidence in 2028–2030 timelines. Monitor financing closures for Indonesia and Sweden expansions and track quarterly capacity additions to determine whether the widening contract-to-active ratio reflects customer phasing or execution friction.

CoreWeave's capacity narrative has expanded dramatically over eighteen months, but deployment velocity remains uncertain. The company announced 420 MW of active power (claim 4972) and 1.6 GW contracted (claim 4973) in May 2025. By March 2026, it retroactively reported 850 MW as of year-end 2025 (claim 3553) against 3.1 GW contracted (claim 3554)—a 5.2x increase in contracted pipeline. Current posture targets 1.85 GW active by December 2026 (claims 8666, 8533, 8281) and 4.2 GW contracted (claims 8522, 8268), with longer-term ambitions of 8 GW by 2030 (claims 8534, 8574). This represents scaling upward, though the company has not revised earlier targets downward in explicit statements.

Promised dates tell a different story. The Core Scientific merger, initially promised for April 2026 (claim 4731), slipped to October 2025 (claim 4517) before termination was announced in October 2025 (claim 4166). Separately, three software acquisitions—Marimo, OpenPipe, and Monolith—were each promised for fall 2025 (claims 3577, 3578, 3579) but appear delayed as of March 2026 (claims 3587, 3588, 3589). These slippages cluster in the M&A domain and predate the March 2026 filing, suggesting the company did not acknowledge delays contemporaneously.

The gap between contracted and active capacity warrants scrutiny. By May 2026, CoreWeave claimed 1 GW active (claim 9321) against 3.5 GW contracted (claim 9323)—a 29 percent deployment ratio. The dataset contains no reference to power interconnection queue positions, state permitting milestones, or third-party infrastructure audits. Indonesia facilities are promised at 360 MW across three locations by end-2028 (claims 8758, 155), but no ground reporting or construction timelines are evident. Kenilworth, New Jersey and Lancaster, Pennsylvania projects are noted (claims 7869, 7873, 7874) without delivery dates.

In the next 90 days, watch whether active capacity reaches the stated 1.85 GW target by December 2026. Observe whether Indonesia development announcements provide specific construction milestones or power procurement commitments. Monitor whether the contracted-to-deployed ratio improves or narrows. Track whether the company revises 2030 capacity goals upward or establishes binding 2027–2028 intermediate targets. Any announcement of facility delays or power procurement setbacks would signal execution stress.

CoreWeave has scaled active capacity 14-fold in 30 months, from 70 MW in late 2023 to 1 GW by May 2026, demonstrating execution at the infrastructure level. Contracted capacity has grown more aggressively, from 1.6 GW announced in May 2025 (claim 4973) to 4.2 GW by August 2026 (claims 8522, 8466), creating a 4:1 ratio between contracted and active capacity that suggests deployment velocity lags deal pace. This gap raises questions about power infrastructure readiness and construction timelines.

The long-term target trajectory reveals unacknowledged scope expansion. Claims from January 2026 specified 5+ GW by 2030 (claim 3893). By May-August 2026, this had shifted to 8 GW (claims 9322, 8534, 7889)—a 60 percent uplift presented without explicit walk-back of prior guidance. The company has not publicly acknowledged this revision.

M&A execution has faltered. The Core Scientific merger faced multiple delays before termination in October 2025 (claim 4166). Smaller acquisitions slipped 2-6 months: OpenPipe and Marimo delayed from September and October 2025 respectively (claims 3588, 3577). These delays suggest capital constraints or integration challenges emerged during rapid growth.

Facility announcements remain light on independent corroboration. Indonesia's three facilities (360 MW by end-2028, claims 8758, 8670) operate under lease arrangements (claim 8760), reducing construction risk but introducing landlord dependencies. Kenilworth (250 MW) and Lancaster (300 MW) campus announcements lack interconnection queue positions, power agreements, or permitting documentation in the record. The Sweden footprint cites colocation partnerships (claims 7871, 7875) without specific capacity or timing.

The immediate test arrives year-end 2026. CoreWeave has committed to 1.85 GW active capacity by December 31 (claims 8666, 8281, 8533). With 1.0 GW confirmed in May, the company must add 850 MW in the final four months—a pace 15 percent faster than the prior quarter. If this target slips, it signals the first explicit miss on an announced active-capacity deadline and may indicate power or deployment bottlenecks. Watch for Indonesia site commencement announcements, permitting releases, Weights & Biases integration progress, and any revised year-end guidance.

CoreWeave's announced capacity has grown substantially in absolute terms—from 70 MW at end-2023 (claim 3551) to 1 GW active by mid-2026 (claim 9321)—but the company has consistently inflated and revised its contracted power pipeline. Total contracted capacity claims escalated from 1.6 GW (claim 4973, May 2025) to 3.1 GW (claims 3554, 3570, March 2026) to 3.5 GW (claim 9323, May 2026) to 4.2 GW (claims 8522, 8262, August 2026), each revision nominally justified by new partnerships or facility announcements but lacking contemporaneous third-party validation. The company's acquisition strategy—repeatedly telegraphed but repeatedly delayed—undermines confidence in execution velocity. The Core Scientific merger was announced four times with different deal terms (claims 4313, 4314, 4294, 4331) before termination in October 2025 (claim 4166). Three smaller acquisitions (Monolith, Marimo, OpenPipe) announced in March 2026 with closure dates in September-November 2025 remain marked as delayed through August 2026, suggesting post-announcement lag between public statements and actual closings. The active-to-contracted ratio has tightened: CoreWeave is now claiming 1 GW deployed against 4.2 GW contracted, a 23% realization rate roughly consistent with industry deployment lags but requiring visible progress on the near-term pipeline. The year-end 2026 active power target of 1.85 GW (claims 8533, 8281, 7888) implies adding 850 MW in roughly 3.5 months, demanding demonstrated construction momentum by September-October. Recent facility announcements clustered in mid-August 2026—Indonesia's 360 MW, three-facility Greater Jakarta campus targeting 2028 completion (claims 155, 8758); Kenilworth's 250 MW greenfield (claim 7869); Lancaster's 300 MW (claim 7870); Stockholm colocation (claim 7875)—are concrete in naming but extremely long in timeline, with only Indonesia's carrying a specific delivery date. Over the next 90 days, watch for: permit filings for Kenilworth and Lancaster; any Indonesia construction commencement notices or interconnection queue progression; achieved active power capacity by November 15 relative to the 1.85 GW target; and disclosure of any further acquisition delays or deal terminations. Partnership announcements with Anthropic, NVIDIA, and Leidos provide revenue underpins but do not reduce execution risk on the physical capacity side.

CoreWeave's announced capacity has scaled sharply, growing from 70 megawatts in late 2023 to 850 megawatts by year-end 2025, with a stated 1.85-gigawatt target for December 2026 (claims 3551, 3553, 8533). Contracted power commitments have more than doubled within eighteen months, rising from 1.6 gigawatts in May 2025 to 4.2 gigawatts by August 2026 (claims 4973, 8522). This trajectory reflects genuine momentum, yet the company's acquisition playbook reveals persistent execution friction. Monolith, Marimo, and OpenPipe were each promised by late 2025, then disclosed as pending through March 2026, representing four-to-six-month delays without public timeline revision (claims 3589, 3588, 3579). The Core Scientific merger, announced with fixed closing dates through September 2025, was terminated in October 2026, suggesting either operational constraints or financial reassessment (claims 4731, 4166). The DDTL 3.0 commitment termination also slipped past its July 2026 target without acknowledged revision (claim 4593).

CoreWeave's capacity growth from 2023 through 2025 is substantive—year-on-year doubling is credible for a specialized AI infrastructure platform. The May 2026 active-power milestone of 1 gigawatt and corresponding 3.5-gigawatt contracted base anchor near-term reality (claims 9321, 9323). However, the 1.85-gigawatt year-end 2026 target represents an 118-percent eight-month increase that lacks independent corroboration. No cross-checks via permits, utility interconnection filings, or third-party audits are provided to validate this pace.

New facility announcements—Indonesia (360 megawatts across three leased colocations by 2028-12-31), Kenilworth (250 megawatts), Lancaster (300 megawatts), and Sweden—remain light on near-term deployment specifics (claims 155, 7869, 7870, 7875). Strategic partnerships with Anthropic and NVIDIA underpin demand signals, but the acquisition slip pattern compounds uncertainty around integration velocity and operational bandwidth.

Monitor over the next ninety days: actual deployed capacity trajectories toward the 1.85-gigawatt year-end target; granular per-facility breakdowns to reconcile the 4.2-gigawatt contracted pool; any public revision to the 2030 8-gigawatt goal (claim 8534) should near-term growth falter; and whether Indonesia phase-one facilities advance from announcement to signed lease or permit filing.

CoreWeave's operational footprint has expanded substantially, growing from 70 MW of active power at end-2023 (claim 3551) to 1+ GW by May 2026 (claim 9321), validating the rough trajectory implied by its December 2025 capacity statement of 850 MW (claim 3569). This near-doubling-annually pace through 2026 appears supported by the timing of announced facilities: Lancaster (300 MW potential, claim 7870), Kenilworth (250 MW, claim 7869), and Polaris Forge 1 (400 MW contracted, claim 370). However, the company has built a significant gap between contracted and operational capacity. By August 2026, CoreWeave claimed 4.2 GW of contracted power capacity (claim 8522) against a stated target of 1.85 GW active by year-end 2026 (claim 8281)—a 2.3x overhang that suggests either aggressive procurement ahead of actual deployment or risk of slower-than-announced ramp. Contracted capacity claims have themselves escalated consistently: from 1.6 GW in May 2025 (claim 4973) to 3.1 GW by March 2026 (claim 3554) to 4.2 GW by August 2026 (claim 8466). This acceleration reflects Indonesia's entry into the pipeline (360 MW by end-2028, claim 8758) and expanded customer commitments, but no independent evidence of grid-side or power-purchase agreements has been provided to verify takeability of that contracted power. CoreWeave's M&A narrative has been chaotic. The Core Scientific merger, promised for October 2025 (claim 4731), was terminated outright by late October (claim 4166). Acquisitions of Marimo, OpenPipe, and Monolith, each promised for September–November 2025 (claims 3577–3589), were still listed as delayed in March 2026 (claim 3589) with no subsequent closure announcement. The Weights & Biases acquisition ($1 billion, claim 4516) has not been updated since March 2026 claims (claim 3576), leaving its status unclear. The 2030 capacity target has been restated at 8 GW (claims 8534, 8574) and separately as 5+ GW (claim 3893), a range that reflects either uncertainty or audience segmentation. Watch whether the 1.85 GW year-end 2026 target is met, whether any of the stalled acquisitions close or are formally abandoned, and whether Indonesia's 2028 deadline stays firm as a credibility signal for longer-horizon commitments.

CoreWeave has demonstrated aggressive capacity acceleration, with active power growing from 70 MW at end-2023 to 1 GW by May 2026—a 14-fold increase in roughly 18 months (claim 3551, claim 9321). The company is targeting 1.85 GW in active power by year-end 2026, matching its disclosed guidance (claims 8533, 8666). Contracted capacity has scaled even more sharply, expanding from 1.6 GW in May 2025 (claim 4973) to 4.2 GW as of August 2026 (claim 8522), suggesting a two-to-three-year deployment pipeline filling with customer commitments. The 850 MW footprint announced for end-2025 (claim 3569) aligns with disclosed milestones, though third-party validation of quarterly capacity additions remains sparse.

On the M&A front, CoreWeave's execution record is uneven. The Core Scientific merger, promised to close by October 2025 (claim 4517), was terminated in October 2026 after repeated delays (claim 4166)—a nine-month slip. The Weights & Biases acquisition, announced in May 2025 with a $1 billion valuation (claim 4516), appears to have closed and is now generating platform updates (claim 9328). Three smaller acquisitions—Marimo, OpenPipe, and Monolith—all missed their original close windows (promised September-November 2025, claims 3588–3589) and have not been updated as closed.

The company's eight-year roadmap to 8 GW by 2030 (claims 9322, 8534) demands reaching that target from a projected 1.85 GW run-rate, implying 4.3x growth over 3.3 years. Recent facility announcements cluster in the latter half of 2026: 250 MW in Kenilworth, NJ (claim 7869), 300 MW in Lancaster, PA (claim 7870), Stockholm colocation facilities (claim 7875), and Indonesia expansion targeting 360 MW by end-2028 (claim 8758). These totals easily exceed 1 GW in newly announced capacity, but granular deployment schedules for each remain unspecified, raising questions about deployment cadence relative to the contracted pipeline.

Over the next 90 days, monitor whether CoreWeave sustains its active-power trajectory toward the 1.85 GW year-end target and whether the delayed M&A transactions (Marimo, OpenPipe, Monolith) receive closure statements. Watch also for Indonesia infrastructure filings or announcements that could validate the 2028-2030 deployment window and clarify whether this represents genuine geographic diversification or pre-signed capacity ahead of firm, disclosed customer wins.

CoreWeave has demonstrated aggressive capacity commitments alongside operational deployment lags that merit close monitoring. Active power capacity grew from 70 MW (end 2023, claim 3551) to 360 MW (end 2024, claim 3552) to 850 MW (end 2025, claim 3553), reaching 1+ GW by May 2026 (claim 9321). However, contracted capacity has scaled faster: from 1.6 GW (May 2025, claim 4973) to 3.1 GW (March 2026, claim 3582) to 4.2 GW (August 2026, claim 8522). This widening gap—contracted at 4.2 GW versus active at 1 GW as of early summer—suggests either accelerating deployment ahead or overpromised contract book. The company targets 1.85 GW active by year-end 2026 (claims 8666, 8281), requiring 850 MW deployment in four months from today's date.

Acquisition execution has faltered. The Core Scientific merger, initially promised for October 2025 (claim 4517) then April 2026 (claim 4731), was terminated by October 2025 (claim 4166). Three software acquisitions—OpenPipe (promised September 2025, claim 3588), Marimo (promised October 2025, claim 3587), and Monolith (promised November 2025, claim 3579)—all show delays as announced March 2026. This pattern suggests integration challenges or deal fatigue, though none have been formally abandoned as of the latest claims in August 2026.

Partnerships appear substantive. NVIDIA designated CoreWeave an Exemplar Cloud partner (claim 9324) and committed $6.3 billion in capacity purchases through 2032 (claim 4352). Anthropic multi-year agreement (claim 8571) and JPMorgan/Mitsubishi UFJ as DDTL 5.0 arrangers (claim 8736) provide financial and commercial validation. The Indonesia facility announcement (August 2026, claim 8758) targeting three sites with 360 MW in Greater Jakarta by end 2028 represents Asia-Pacific entry, though the 2028 timeline sits far enough out to absorb delays.

The 8 GW target by 2030 (claims 8574, 8534, 8872) requires roughly 3× the active capacity in 3.3 years. Over the next 90 days, watch whether the company closes the contracted-to-active gap through accelerated deployments at Polaris Forge 1 (400 MW, claim 370), Kenilworth (250 MW, claim 7869), or Lancaster (300 MW, claim 7870). Permitting status and grid interconnection queue positions for Indonesia and greenfield U.S. sites remain undisclosed. Any acquisition closures, particularly Weights & Biases (repeatedly cited, claims 3562, 3576), would signal execution capability recovery.

CoreWeave's announced capacity trajectory reveals aggressive expansion paired with significant execution risk. In May 2025, the company claimed 420 MW active and 1.6 GW contracted (claims 4972, 4973). By March 2026, three months into operations as a public company, it had revised contracted capacity upward to 3.1 GW (claim 3554) while reporting 850 MW active as of December 31, 2025 (claim 3553). By August 2026, contracted capacity escalated further to 4.2 GW (claim 8522), but the company was only half way to its stated 1.85 GW year-end active target (claims 8533, 8281). This 40 percent deployment ratio—active-to-contracted—has remained consistent across multiple reporting cycles, suggesting either aggressive contracting, phased power availability from suppliers, or capacity not yet drawing load.

Promised timelines show acute slippage on strategic acquisitions. The Weights & Biases acquisition was announced twice (May 2025 claim 4974, March 2026 as core vertical integration play), but its completion date has never been disclosed. The Core Scientific merger, originally promised to close by October 2025 (claim 4517), was terminated in October 2026 (claim 4166)—a complete reversal announced without prior disclosure of material obstacles. Marimo, OpenPipe, and Monolith were each promised by September-November 2025 but remain unresolved as of March 2026 (claims 3577, 3578, 3579), nine months overdue.

Capacity targets have escalated without acknowledgement of prior commitments. January 2026 stated a 5 GW target by 2030 (claim 3893). Seven months later, August 2026 announcements introduced an 8 GW target for the same deadline (claims 8534, 7868)—a 60 percent increase in ambition in under a year. Simultaneously, new geographies entered public record in August: Indonesia (360 MW by 2028), Lancaster (300 MW potential), Kenilworth (250 MW greenfield), and Stockholm (two campuses)—projects that were not named in prior filings and whose power sourcing remains unclear.

Over 90 days, observe whether CoreWeave publishes Q3 2026 active capacity trending toward 1.85 GW, whether Indonesia facilities secure formal power contracts with PT PLN, and whether August's newly announced projects receive filed interconnection queue positions. The gap between contracted and deployed capacity will signal whether the company is constrained by supply-chain, power availability, or customer off-take.

CoreWeave's capacity roadmap has undergone dramatic expansion in the nine months since May 2025, yet the trajectory reveals persistent gaps between announced ambitions and tangible progress. Contracted power capacity claims have escalated from 1.6 GW (claim 4973, May 2025) to 4.2 GW (claim 8522, August 2026), while active deployments have crossed 1 GW (claim 9321, May 2026) against a year-end 2026 target of 1.85 GW. The company reported 70 MW operational as of end-2023 (claim 3567), 360 MW by end-2024 (claim 3552), and 850 MW by end-2025 (claim 3569)—a credible progression if verified. However, announced capacity has outpaced visible execution: the four facility announcements (Indonesia, Kenilworth, Lancaster, Stockholm) remain undated for operational launch, with only the Indonesia cluster carrying a delivery promise (360 MW by December 2028, claim 8758).

The acquisition strategy demonstrates material execution risk. Three attempted purchases—Marimo (claim 3577), OpenPipe (claim 3588), and Monolith (claim 3589)—all missed their late-2025 close dates. The Core Scientific merger, positioned as transformative, was terminated in October 2025 (claim 4166) after slipping repeatedly from its original September 2025 close. Weights & Biases acquisition announcements recurred without closure indication (claims 4974, 3547, 3562, 3576, 3586). This pattern suggests either integration challenges or shifts in strategic priorities that have not been publicly acknowledged as reversals.

No claims cite independent validation: permits filings, regional interconnection queue entries, or equipment procurement notices. All facility and partnership claims carry [stated] status, indicating announced commitments rather than third-party corroboration. The partnership announcements—IBM Granite (claim 4971), NVIDIA collaboration (claim 3892 and subsequent NVIDIA-tied claims), Anthropic (claim 8571), and Hudson River Trading (claim 1234)—lack specificity on capacity allocation or delivery timing. The NVIDIA Exemplar Cloud designation (claim 9324) and joint 5+ GW commitment by 2030 (claim 9329) are significant but carry no enforcement mechanism disclosed.

In the next 90 days, watch whether CoreWeave reports Q3 2026 active power results (expected near October earnings) and whether the 1.85 GW year-end target remains intact (claim 8281, claim 8666). Monitor Indonesia facility announcements for permit filings or colocation provider confirmations (claim 8760). Track whether any of the three stalled acquisitions relaunch or formally dissolve. Finally, observe whether the 4.2 GW contracted capacity shows signs of actual build-out through procurement announcements or customer SLA confirmations.

CoreWeave has disclosed a consistent upward capacity curve: 70 megawatts at year-end 2023 (claim 3551), 360 MW by end of 2024 (claim 3552), and 850 MW at end of 2025 (claim 3553). The company announced it surpassed 1 gigawatt of active power in May 2026 (claim 9321), with a stated year-end 2026 target of 1.85 GW (claims 8281, 8533, 8666). This trajectory—roughly 5x growth over three years—shows disciplined scaling without announced reductions. Contracted capacity claims have also shifted upward: from 1.6 GW stated in May 2025 (claim 4973) to 3.5 GW by May 2026 (claim 9323) to 4.2 GW by August 2026 (claim 8522). The company projects 8 GW of active power by end of 2030 (claims 9322, 8534, 8574), implying a near-doubling requirement from mid-2026 levels over four years.

The acquisition program shows material slippage. The Core Scientific merger, promised by April 2026 (claim 4731) and then October 2025 (claim 4517), was terminated entirely (claim 4166). Marimo, OpenPipe, and Monolith acquisitions, each promised between September and November 2025 (claims 3578, 3579, 3588, 3589), remain unresolved as of September 2026. Weights & Biases has been announced multiple times without a stated close (claims 4974, 3562, 3576) and carries a $1 billion reported price (claim 4516). One commitment—DDTL 3.0's termination date of July 31, 2026 (claim 4593)—has already passed without public acknowledgement of impact.

Facility announcements arrived en masse in August 2026: Indonesia (360 MW by end 2028, claims 155, 8758), Kenilworth greenfield (250 MW, claim 7869), Lancaster expansion (300 MW, claim 7874), and Sweden co-location (claims 7871, 7875). These are geographically diversified but timeline-distant. No independent evidence from permits, interconnection queue entries, or construction reports appears in the claims record.

Watch the next 90 days for end-of-year capacity delivery against the 1.85 GW target. Any Indonesia or Kenilworth project updates, clarifications on delayed acquisitions, or revised timelines will signal whether the 8 GW 2030 path rests on binding commitments or optimistic guidance.

CoreWeave's capacity narrative shows aggressive scaling promises paired with notable execution gaps. Active power capacity has grown from 70 MW (end-2023) to a 1.85 GW target for end-2026 (claims 3551, 8281)—a compounded rate that would place the company on track for its 8 GW ambition by 2030 (claims 9322, 8534). Yet contracted capacity now stands at 4.2 GW (claim 8522), nearly 2.3x the active power target, widening the gap between commitments and live deployment. This divergence signals either aggressive forward bookings ahead of build-out or execution risk that customers won't take allocated capacity on schedule.

Acquisition activity reveals operational strain. The Core Scientific merger, initially pledged by April 2026 (claim 4731), then October 2025 (claim 4517), was terminated outright by late October (claim 4166)—a significant strategic reversal not explicitly acknowledged in later announcements. The Weights & Biases acquisition, announced May 2025 (claim 4974) at $1 billion (claim 4516), appears stalled; March 2026 filings list it without completion dates, signaling delay. Marimo, OpenPipe, and Monolith acquisitions all slipped past their 2025 close dates (claims 3577-3589). This clustering suggests CoreWeave is deprioritizing M&A in favor of capital deployment and operational execution.

Geographic footprint is expanding rapidly. Indonesia's 360 MW commitment by end-2028 (claim 8758) marks Asia-Pacific entry; Kenilworth, New Jersey targets 250 MW greenfield (claim 7873); Lancaster, Pennsylvania 300 MW (claim 7874); Stockholm colocation arrangements (claim 7871). These August announcements add geographic credibility, though no independent permit filings, construction updates, or interconnection queue entries appear in the disclosed record. Partnership announcements—Anthropic (claim 8571), Leidos (claim 8669), Hudson River Trading (claim 958)—signal demand strength but not capacity proof.

The critical near-term test arrives in four months: whether active power reaches 1.85 GW by December 2026. Watch whether the 4.2 GW contracted pipeline begins converting to live deployment or whether revised guidance follows. Indonesia project visibility—site acquisition, permitting, power interconnection queue movement—will establish credibility for the 2028 timeline. Any further contraction in the 8 GW 2030 target or public acknowledgment of the capacity overhang would signal material execution challenge.

CoreWeave has articulated one of the industry's most aggressive power expansion roadmaps: 70 MW in 2023 rising to 850 MW by year-end 2025 and targeting 1.85 GW by December 2026 (claims 3551, 3553, 8666), with an ultimate 8 GW vision by 2030 (claim 8574). The capacity trajectory reflects near-exponential growth, but execution has proven uneven and independent verification remains limited.

The most significant slippage involves the Core Scientific merger, originally promised to close by October 2025 (claim 4517) but terminated entirely in October 2026 (claim 4166). This was CoreWeave's marquee infrastructure acquisition and its failure suggests overconfidence in earlier deal timelines. Smaller acquisitions targeting software—OpenPipe, Monolith, and Marimo—also each missed promised closeout dates (claims 3578, 3589, 3577). The Weights & Biases acquisition, announced mid-2025 for $1 billion, did close successfully, but the acquisition tempo has noticeably slowed.

CoreWeave has pivoted strategy, anchoring growth through long-term partnerships rather than asset purchases. NVIDIA committed $6.3 billion in capacity purchases through 2032 (claim 4352); Anthropic, Hudson River Trading, Backblaze, and Leidos signed multi-year agreements (claims 8571, 1234, 7645, 8669). These partnerships offset capacity risk by pre-selling power, but they also lock in unfavorable pricing if demand outpaces supply.

The 4.2 GW of contracted power capacity (claim 8522) substantially exceeds deployed MW, signaling either legitimate growth confidence or exposure to non-completion risk. Indonesia's first Asia-Pacific deployment targets 360 MW by end of 2028 (claim 8758); Kenilworth, New Jersey and Lancaster, Pennsylvania greenfield sites are also underway (claims 7869, 7870). What's notably absent from the record: permits, interconnection queue entries, or public filings from colocation partners that would independently validate these timelines. Reliance on announcement cadence alone leaves material risk unpriced.

CoreWeave must hit 1.85 GW by December 31, 2026—roughly 1 GW of new capacity in four months (claim 8281). If this target slips, the 8 GW 2030 vision becomes questionable, especially given Core Scientific's abandonment and acquisition delays. Watch also for updates to the DDTL 3.0 debt facility (claim 4593), which saw its termination date pushed beyond original guidance—financial flexibility underpins execution.

CoreWeave's infrastructure trajectory shows aggressive growth commitments paired with execution headwinds. From December 2023 to December 2025, the company deployed from 70 MW to 850 MW of active power capacity, a compound 2.4x annual growth rate (claims 3551, 3553). That growth continues by design: claims 8533 and 8281 target 1.85 GW active capacity by year-end 2026. But a critical signal emerged in August: claim 8666 set the year-end 2026 target at 1.85 GW on August 13, yet eight days later claim 7652 revised it downward to 1.7 GW without public acknowledgement. This 150 MW contraction, precisely matching the incremental increase cited in claim 8582, suggests internal pacing delays or pushed-out facility timelines.

The contracted-to-active capacity gap widens considerably. While active power is projected at 1.7–1.85 GW by year-end 2026, contracted capacity now sits at 4.2 GW (claims 8522, 8262). The company is signing customer commitments faster than deploying hardware—a pattern that intensifies execution risk if customer timelines harden against the announced schedule.

Announced facilities lack delivery precision. Kenilworth (250 MW), Lancaster (300 MW), Sweden (unspecified), and Indonesia (360 MW by end-2028) carry no granular construction or permitting milestones (claims 7869, 7870, 155). Indonesia, CoreWeave's stated first Asia-Pacific facility, targets three Greater Jakarta sites with colocation via 360 MW by December 2028 (claim 8758)—ambitious given typical land-power-interconnection lead times in that region.

Acquisition delays compound integration uncertainty. The Core Scientific merger, promised for October 2025, terminated the following month (claim 4166). Marimo, OpenPipe, Monolith, and Weights & Biases acquisitions all slipped past 2025 windows (claims 3577–3589), suggesting financing friction or due-diligence friction.

Watch for: (1) whether CoreWeave reaches the revised 1.7 GW target by December 31, requiring 850 MW deployment in four months; (2) filing or public permitting news on Kenilworth greenfield or Indonesia grid interconnection; (3) customer renegotiation signals in the 4.2 GW contracted backlog as lead times compress against the 2030 8 GW target (claim 8534).

CoreWeave has demonstrated substantial capacity growth from 70 megawatts at year-end 2023 to 850 megawatts by year-end 2025, representing a twelve-fold increase in deployed infrastructure (claims 3551, 3569). The company has announced an 8-gigawatt target by 2030, supported by 4.2 gigawatts of contracted capacity currently in various stages of deployment (claims 8522, 7629). However, this trajectory masks significant execution challenges and ambiguities.

The core issue is the gap between contracted and deployed capacity. As of mid-2026, CoreWeave has contracted 4.2 GW of power but will have deployed roughly 1.85 GW by year-end 2026 if it meets its target (claims 8533, 8522). This 2.4x multiple suggests either aggressive purchasing ahead of actual deployment needs or confidence in acceleration. The slight variance between claimed year-end 2026 targets—1.85 GW in claims 8533 and 8281 versus 1.7 GW in claim 7652—suggests internal planning uncertainty.

Execution risk is evident in the company's acquisition history. The Core Scientific merger, announced with a 0.1235 share exchange ratio (claims 4331, 4314), was terminated outright on 2025-10-30, only two months past its promised close date (claim 4166). Smaller acquisitions of Marimo, OpenPipe, and Monolith, each promised for Q3-Q4 2025, remain unresolved as of March 2026 (claims 3577, 3578, 3579). The Weights & Biases acquisition for 1 billion dollars was stated without a date.

Geographic expansion plans—Indonesia's 360-megawatt facility by 2028-12-31, Lancaster's 300 megawatts, Kenilworth's 250 megawatts, and Stockholm colocation deals—appear stated but lack dated deadlines for most (claims 155, 7870, 7869, 7875). Indonesia is the only facility with a firm timeline.

Independent verification is absent from this record. No permits, grid interconnection queue entries, or third-party construction reports are referenced. The Anthropic and Hudson River Trading partnerships are named but lack capacity or duration specifics (claims 8571, 7605).

Watch whether CoreWeave hits its 1.85 gigawatt target by year-end 2026, only four months away. Any miss would signal execution slowing below the 12-month doubling pace shown through 2025. Track Indonesia's site status and grid connection progress toward the 2028-12-31 deadline, and monitor whether the 4.2 gigawatts of contracted capacity begins materializing into deployment schedules with quarterly milestones.

CoreWeave's deployment trajectory reveals substantial contracted capacity paired with more modest active deployment, creating a widening gap between announcement and execution. The company claimed 70 MW active capacity at end-2023, reaching 360 MW by end-2024 and 850 MW by end-2025—maintaining roughly 5x to 2.4x annual growth rates through those periods. However, claims 8533 and 8666 project 1.85 GW active capacity by year-end 2026, representing just 2.2x growth and falling meaningfully short of the 4.2 GW total contracted capacity now claimed across claims 8522, 8262, and 8268. This suggests CoreWeave is contracting power faster than deploying it, a pattern typical of infrastructure companies in expansion mode but worth monitoring for demand-supply misalignment.

Timeline adherence has been inconsistent. CoreWeave announced multiple acquisitions with specific promised dates—OpenPipe by September 2025 (claims 3578, 3588), Marimo by October 2025 (claims 3577, 3587), and Monolith by November 2025 (claims 3579, 3589)—all marked delayed. More significantly, the Core Scientific merger was announced repeatedly with evolving promised dates, culminating in a April 2026 target (claim 4731), before termination on October 30, 2025 (claim 4166). These misses suggest execution challenges in both M&A and facility delivery, though the company has consistently reiterated an 8 GW target by 2030 across claims 8534, 7872, and 7888.

Recent announcements reveal geographic diversification: Indonesia's first Asia-Pacific facility targeting 360 MW by end-2028 (claim 155, 8758), Polaris Forge 1 at 400 MW (claim 370), Kenilworth Campus at 250 MW (claim 7869), and Lancaster expansion to 300 MW (claim 7870), plus two Stockholm colocation campuses (claims 7871, 7875). August 2026 brought major partnerships—JPMorgan and Mitsubishi UFJ as lead arrangers (claim 8736), NVIDIA's 6.3 billion USD commitment through 2032 (claim 4352), and Anthropic's multi-year deal (claim 8571)—suggesting serious financial and customer backing despite M&A stumbles.

For the next 90 days, watch whether active capacity actually reaches 1.85 GW by December 2026. CoreWeave must add roughly 1 GW in four months—achievable given announced sites but dependent on power interconnection readiness, which independent queue data will clarify. Also track whether delayed acquisitions (Marimo, OpenPipe, Monolith) finally close or vanish from discourse entirely.

CoreWeave's announced capacity has followed an aggressive expansion trajectory, growing from 70 MW at end-2023 to a stated 850 MW by end-2025, with a 1.85 GW target for end-2026. Contracted capacity tells a more expansive story: rising from 1.6 GW in May 2025 to 3.1 GW by March 2026 to 4.2 GW by August 2026. This gap between contracted and active capacity—currently 2.35 GW—reflects typical lag between commitment and deployment but suggests execution risk if operational ramp-up falters. Timeline performance has been mixed. The Core Scientific merger, promised by October 2025, was terminated entirely by October 30, 2025. Three smaller acquisitions (Marimo, OpenPipe, Monolith) each promised by September–November 2025 remained delayed as of March 1, 2026. A DDTL 3.0 commitment termination also slipped past its July 31, 2026 date. These misses represent strategic repositioning rather than pure failure—the Core Scientific deal abandonment coincided with partnerships announced six months later—but they signal execution friction on the M&A side. What supports the pace: major institutional backing materialized. JPMorgan and Mitsubishi UFJ jointly arranged financing by August 9, 2026; NVIDIA committed $6.3 billion across partnership frameworks and capacity purchases; Anthropic, Hudson River Trading, and Leidos signed multi-year agreements. The speed of recent partnership announcements (all August 2026) suggests operational readiness. Indonesia's three-facility footprint, announced August 6, 2026 and targeting 360 MW by end-2028, indicates geographic expansion timing but relies on colocation partnerships, not greenfield construction. Domestic projects (Kenilworth, 250 MW; Lancaster, 300 MW; Polaris Forge 1, 400 MW) suggest real pipeline activity. Watch whether CoreWeave hits the 1.85 GW active capacity target by December 31, 2026—a miss would signal either demand absorption issues or sustained construction delays. Monitor Indonesia site activation timing, which could compress if third-party colocation partners accelerate provisioning, or extend if power and networking infrastructure lags. Track whether the delayed acquisitions close and whether their integration actually adds compute-adjacent capabilities or remains aspirational.

CoreWeave's capacity trajectory shows aggressive acceleration alongside delayed acquisitions. From May 2025 through March 2026, the company more than doubled stated active power capacity from 420 MW (claim 4972) to 850 MW (claim 3553), while contracted capacity grew from 1.6 GW (claim 4973) to 3.1 GW (claim 3570). However, year-end 2026 targets show inconsistency: August announcements cite both 1.85 GW (claims 8666, 8533, 8281) and 1.7 GW (claim 7652) as the active power target. Reaching either would require 1–1.2 GW of deployment in nine months—a meaningful acceleration from the 490 MW added across 2025.

The company's acquisition track record has been mixed. Core Scientific merger, initially promised for October 2025 (claim 4517), slipped and was terminated by October 30 (claim 4166). Marimo, OpenPipe, and Monolith acquisitions, all promised for September–November 2025 and flagged in the March 2026 filing as delayed (claims 3577, 3578, 3579), show no completion status. The Weights & Biases acquisition was stated without a deadline (claims 4974, 3547).

Geographically, CoreWeave is expanding rapidly but with limited near-term specificity. Indonesia's first Asia-Pacific facility targets 360 MW by end-2028 (claims 155, 8758)—a longer timeline than domestic projects. Lancaster, Pennsylvania and Kenilworth, New Jersey facilities are announced with 300 MW and 250 MW potential respectively (claims 7874, 7873), but no completion dates. Sweden's Stockholm facilities are described in general terms (claims 7871, 7875). Polaris Forge 1 carries 400 MW of contracted capacity (claim 370) with no timeline attached.

Recent activity shows renewed momentum. August announcements brought partnerships with JPMorgan and Mitsubishi UFJ (claim 8736), Leidos, Anthropic, Hudson River Trading, and Backblaze. The 8 GW by 2030 target (claims 8534, 7872, 7889, 7629) appears standardized across recent communications.

Watch whether the company closes the year-end 2026 target gap, clarifies the 1.85/1.7 GW variance, and demonstrates facility progress beyond announcements. The pending M&A completions and Indonesia facility operational timeline matter; if domestic capacity targets slip again while international facilities lag, the 2030 8 GW goal becomes harder to credibly defend.

CoreWeave's announced capacity has grown materially from 70 MW at end-2023 to 850 MW at end-2025, with targets reannounced at 1.85 GW for year-end 2026 and 8 GW by 2030. The historical doubling-per-year trajectory is notable, though the 2026 target (1.85 GW per claims 8533, 8666, 8281, 7888) faced internal revision downward to 1.7 GW as recently as August 20th (claim 7652), suggesting forecast confidence may have wavered. The company has restated its 2030 target identically across six separate announcements since January (claims 3893, 8534, 8574, 7872, 7629, 7889) without adjustment despite nine months of operational history, implying either unchanged planning or unrevised modeling.

Promised acquisition dates have consistently slipped or terminated. The Core Scientific merger moved from an April 2026 close to October 2025 (claims 4731, 4517) and was terminated entirely by October 30th (claim 4166). Software acquisitions of Marimo, OpenPipe, and Monolith all exceeded their target close windows of September through November 2025 (claims 3577, 3578, 3579). The DDTL 3.0 commitment termination date promised for July 2026 now appears delayed (claim 4593). Only the Weights & Biases acquisition received restatement without explicit date revision, though no close date was promised in the record provided.

Independent evidence from permitting or grid interconnection queues is not visible in the claim record. The portfolio includes three Indonesia facilities (360 MW by 2028-12-31) using colocation lease arrangements rather than owned infrastructure, reducing capital risk but introducing third-party landlord and permitting dependencies (claims 155, 8758, 8760). Announced greenfield builds in New Jersey (250 MW) and Pennsylvania (300 MW potential) carry longer lead times. Contracted capacity across all projects stands at 4.2 GW (claims 8262, 8268, 8466) against deployed power of 850 MW—a ratio suggesting either aggressive bookings or execution risk.

The company stated near-term capacity sold out by mid-August (claim 1494), indicating demand elasticity, but the next ninety days will test whether this translates to builds. Reaching 1.85 GW by year-end requires approximately 1 GW of deployed capacity between now and December 31st—a quarterly pace roughly four times historical ramps. Indonesia deployment schedules remain unvalidated against local permitting timelines. Watch for actual commissioning announcements and any further revision to the 2026 active capacity target.

CoreWeave's announced infrastructure capacity has expanded dramatically and repeatedly. In May 2025, contracted power stood at 1.6 gigawatts (claim 4973). By March 2026, that figure had roughly doubled to 3.1 GW (claim 3554), and by August 2026 it climbed to 4.2 GW (claim 8522)—a 2.6-fold increase in fifteen months. The trajectory resembles a company aggressively acquiring commitments, yet active capacity—the compute already running—tells a different story. Year-end 2025 active capacity was claimed at 850 MW (claim 3553), with a 2026 target of 1.85 GW (claim 8533). The gap between 4.2 GW contracted and 1.85 GW targeted for operational deployment at year-end suggests either that contracts are outpacing buildout capability, or that near-term construction pipelines face headwinds.

Acquisition discipline has slipped. The Core Scientific merger, promised for April 2026, was terminated by October 2025 (claim 4166) after missing its original October 2025 deadline (claim 4517). Smaller bolt-ons—Marimo, OpenPipe, Monolith—each missed late-2025 close dates (claims 3577, 3588, 3589). These delays suggest integration complexity or shifting strategic priorities.

The Indonesia announcement in August 2026 introduces both promise and calendar risk. CoreWeave committed to three facilities in Greater Jakarta totaling 360 MW by end of 2028 (claims 8758, 155). The geographic diversification is strategic, but the 2028 delivery window is two years out, making near-term verification difficult. Domestically, three projects anchor the next phase: Polaris Forge 1 (400 MW, claim 370), Kenilworth (250 MW, claim 7873), and Lancaster (300 MW, claim 7870). CoreWeave has stated near-term capacity is sold out (claim 1494), suggesting demand, not supply constraints—but converting contracted megawatts into operational capacity has proven harder than anticipated.

By December 2026, CoreWeave must deliver the 1.85 GW active target to sustain momentum and investor confidence. The next ninety days will reveal execution velocity against major domestic projects and validate whether 2030's 8 GW goal remains credible or drifts further.

CoreWeave has announced an aggressive expansion trajectory, progressing from 70 MW of operational capacity at end-2023 to 850 MW by end-2025, with a stated target of 1.7–1.85 GW active capacity by year-end 2026 and 8 GW by 2030 (claims 3551–3553, 7652, 7889). This implies roughly doubling deployed capacity in four months from the August 2026 announcement window, followed by sustained annual growth of ~1.5 GW through 2030. Contracted power capacity stands at 4.2 GW (claim 8522), meaning only a fraction of long-term commitments are operationalized. The company's acquisition roadmap has proven volatile: Core Scientific was announced for merger in May 2025, re-confirmed through September 2025, then terminated by late October 2025 (claims 4731, 4517, 4166). Three software acquisitions (OpenPipe, Monolith, Marimo) initially targeted for September–November 2025 close-out appear delayed as of the March 2026 reporting cycle (claims 3577–3589). On timeline execution, a DDTL 3.0 commitment termination was promised for July 31, 2026 (claim 4593), yet no public closure has been documented past that deadline. Named projects suffer from uneven specificity: the Indonesia facility (first Asia-Pacific deployment) has a single operational milestone pegged to 360 MW by end-2028 (claim 155), while Polaris Forge 1 (400 MW contracted; claim 370) and Sweden expansion (claim 7644) carry no delivery dates. Recent partnership announcements—Anthropic, Hudson River Trading (including alleged $22.4 billion in aggregate commitments), Backblaze, and JPMorgan/MUFJ financing roles—arrived in mid-to-late August 2026 and remain unvalidated on deployment. Notably, August claims contain contradictory year-end 2026 targets: multiple sources cite 1.85 GW (claims 8666, 8281, 7888) while claim 7652 specifies 1.7 GW, suggesting either mid-campaign recalibration or conflicting internal estimates. No permits, interconnection queue positions, or third-party construction verification are cited in disclosed materials. Watch whether the 1.7–1.85 GW year-end target materializes, conversion velocity from contracted to active megawattage, and whether the major partnership announcements translate into firm capacity bookings or remain aspirational commitments.

CoreWeave's announced capacity has experienced significant expansion with a pattern of aggressive target-setting followed by partial realization. Historical capacity grew from 70 MW (end 2023) to 360 MW (end 2024) to 850 MW (end 2025), representing an acceleration trajectory. However, contracted capacity outpaces active deployments substantially: as of March 2026, CoreWeave held 3.1 GW in contracts against 850 MW operational, a 27% conversion rate (3551, 3552, 3553, 3554). Most recent guidance targets 1.85 GW active by year-end 2026 (8533, 8281), yet claim 7652 issued four days later revises this downward to 1.7 GW without explanation, suggesting internal uncertainty or miss on interim milestones.

Timeline discipline remains the clearest vulnerability. The Core Scientific merger, announced in May 2025, was promised by April 2026 (4731), then October 2025 (4517), before termination in October without public acknowledgment of failure (4166). Smaller acquisitions—Marimo, OpenPipe, Monolith—each slipped three to six months from initial September-November 2025 windows, now buried in March 2026 filings (3577, 3588, 3589) without revision commentary. Weights & Biases acquisition claims repeat across multiple filings with no closure announcement despite a stated 1 billion USD value (4516).

Independent verification is absent from the record. No permit filings, interconnection queue entries, or third-party construction milestones substantiate the announced pace. Recent facility announcements—Kenilworth, NJ greenfield (250 MW), Lancaster, PA expansion (300 MW), Stockholm colocation (undisclosed MW), Indonesia (360 MW by 2028-12-31, claim 155)—are early-stage and carry execution risk typical of greenfield and international projects. The Indonesia timeline particularly warrants scrutiny: 2.5 years to 360 MW in a new jurisdiction with grid, permitting, and supply-chain dependencies.

Capacity growth from 850 MW to claimed 1.85 GW in 12 months would require roughly 2x historical annual deployment rates. Watch for Q3 and Q4 2026 earnings reports detailing actual active capacity adds; resolution of the 1.85 GW versus 1.7 GW target discrepancy; and meaningful progress on Kenilworth and Lancaster projects (permit filings, land acquisition). Customer ramp data from Anthropic, Hudson River, and OpenAI partnerships will signal demand-pull validity.

CoreWeave has scaled announced power capacity substantially but faces execution questions. Operational capacity grew from 70 megawatts at end-2023 (claim 3551) to 850 megawatts by year-end 2025 (claim 3553), roughly a tenfold gain. Contracted capacity tells a different story: 1.6 gigawatts in May 2025 (claim 4973) expanded to 3.1 gigawatts by March 2026 (claim 3559) and reached 4.2 gigawatts by mid-August 2026 (claim 8522). This gap between contracted and operational—currently a 2.35 gigawatt spread—reflects announcement velocity outpacing deployment.

The company's timeline track record is uneven. The Core Scientific merger, announced mid-2025 with a targeted October close (claim 4517) and detailed exchange ratios across multiple announcements, terminated entirely by late October (claim 4166). Three concurrent acquisitions announced March 2026—Marimo, OpenPipe, and Monolith—carried promised completion dates in autumn 2025, already months past when they were announced (claims 3577, 3588, 3589). CoreWeave's commitment termination deadline for DDTL 3.0 was promised for July 31, 2026, and marked delayed (claim 4593). These reversals and missed dates suggest either aggressive initial guidance or strategic pivots not transparently communicated.

Facility anchors provide some grounding. Named projects include Kenilworth, New Jersey (250 megawatts, claim 7873) and Lancaster, Pennsylvania (300 megawatts potential, claim 7874), with Swedish colocation campuses and an Indonesia facility promised for 360 megawatts by December 2028 (claim 155). The Indonesia deployment carries particular risk given geographic complexity and 25-month window. CoreWeave's 2026 year-end active capacity target appears inconsistent across filings—flagged at both 1.85 gigawatts (claim 8533) and 1.7 gigawatts (claim 7652)—suggesting imprecision in forward guidance or mid-course corrections.

Watch for Q4 2026 delivery against the 1.7–1.85 gigawatt target, Indonesia's permitting trajectory, and whether the 4.2 gigawatt contracted capacity begins converting to active deployment at promised rates. The 8 gigawatt 2030 ambition requires roughly 6-7x current active capacity in four years.

CoreWeave's capacity trajectory shows aggressive expansion, growing from 70 MW at end-2023 to 850 MW by end-2025 before targeting 1.85 GW by year-end 2026 (claims 3551, 3553, 8533). To hit that intermediate target from current levels represents a sustained doubling pace, consistent with prior growth but arriving just four months out. The company's contracted power capacity sits at 4.2 GW (claim 8522), more than double its 2026 active target and half its 2030 ambition of 8 GW (claims 8534, 8574). This gap warrants scrutiny: either deployment velocity will accelerate significantly or contracted commitments face pressure. Recent capacity announcements (late August) reference specific sites—Kenilworth, New Jersey at 250 MW and Lancaster, Pennsylvania at 300 MW, plus Indonesia at 360 MW by end-2028 (claims 7869, 7870, 155)—providing geographic diversification but fragmenting execution focus across multiple construction and grid-connection timelines.

CoreWeave's M&A discipline shows cracks. The Core Scientific merger, announced multiple times from May through September 2025, terminated in late October (claim 4166) after a promised close date of October 1 slipped (claim 4517). Smaller acquisitions of Marimo, OpenPipe, and Monolith, all announced in March 2026, missed their promised September-through-November 2025 dates by months (claims 3577, 3578, 3579). The company has not publicly acknowledged these reversals as formal cancellations; the record shows delays without explicit withdrawal language. More concerning, the DDTL 3.0 commitment termination date promised for July 31, 2026 (claim 4593) is now past; no updates clarify whether the commitment stands, was extended, or was quietly abandoned.

Partnership announcements accelerated sharply in August 2026—multi-year agreements with Anthropic, Hudson River Trading, and Backblaze, alongside reported OpenAI commitments exceeding $22 billion (claims 8571, 1242, 7448, 7598). Partnership velocity suggests strong demand capture but creates execution dependencies external to CoreWeave's control. Watch whether the 1.85 GW active target is hit on schedule by December 31; whether Indonesia begins major construction milestones; and whether delayed acquisition closures are formally updated or quietly shelved. The next quarterly filing should clarify M&A status and DDTL 3.0 standing. Slippage on the 2026 capacity target, coupled with further unacknowledged timeline slides on smaller deals, would signal execution constraints beneath the surface.

CoreWeave's expansion narrative reveals escalating commitments alongside unresolved dependencies. Historical deployment data (claims 3551–3553) shows genuine momentum: 70 megawatts at end-2023, 360 megawatts at end-2024, and 850 megawatts at end-2025. However, contracted capacity—1.6 gigawatts announced May 2025—jumped to 4.2 gigawatts by August 2026 (claims 4973, 8522), outpacing deployed volumes by five times. This gap suggests either accelerating deployment velocity or unexecuted commitments.

Near-term targets introduced numerical inconsistency. Year-end 2026 guidance ranged from 1.7 to 1.85 gigawatts across four separate August announcements (claims 8533, 8281, 7652, 1489). Even at the high end, 1.85 gigawatts represents only 2.2 times 2025 deployed capacity—modest compared to historical growth rates and the gulf to 4.2 gigawatts contracted. This suggests deployment velocity is decelerating.

Strategic acquisitions introduced execution risk with no resolution. Core Scientific merger, promised October 2025 close (claim 4731), terminated without stated cause by January 2026 (claim 4166). Marimo, OpenPipe, and Monolith, each promised to close in September–November 2025, remained delayed in March 2026 filings (claims 3577–3589). Weights & Biases acquisition remains undated despite $1 billion valuation (claim 4516). These delays matter if acquisition targets were capacity enablers; their absence may explain the deployment gap.

Long-term guidance revised without acknowledgement. January 2026 communications projected five-plus gigawatts by 2030; August 2026 messaging escalated to eight-plus gigawatts (claims 3893, 8534, 7872). The sixty-percent increase coincides with acquisition failures, raising questions about planning basis.

Geographic diversification lacks transparent foundation. Indonesia targets 360 megawatts by end-2028 (claim 155); Kenilworth and Lancaster projects are named but lack visible permits or construction milestones. Sweden mentions colocation without timeline. NVIDIA co-engineering (claim 1501) and OpenAI contract ($22.4 billion, claim 1242) confirm customer demand but not deployment certainty.

Monitor Q4 2026 results for actual year-end capacity versus stated targets. Track acquisition closure dates, Indonesia progress milestones, and Kenilworth/Lancaster permit filings. Watch whether 2027 guidance addresses contraction between contracted and deployed capacity or explains the 2030 target escalation.

CoreWeave has plotted an aggressive capacity expansion trajectory that shows significant scaling velocity but faces execution credibility questions. Reported active capacity grew from 70 MW (end 2023, claim 3551) through 360 MW (2024, claim 3552) to 850 MW (2025, claim 3553), reaching approximately 1.5 GW by late August 2026 (claim 1455). This represents compound growth, but contracted capacity has outpaced active deployment: claims 3554 and 3582 from March 2026 cited 3.1 GW total contracted capacity, while claims 8262 and 1456 from mid-to-late August now place contracted capacity at 4.2 GW. This 35 percent jump in contracted commitments over five months suggests either aggressive customer wins or revised internal assumptions about deployment feasibility.

Promised timelines reveal execution risk. The Core Scientific merger, initially promised by April 2026 (claim 4731), faced delays and ultimately terminated (claim 4166). Parallel acquisitions of Marimo, OpenPipe, and Monolith promised for September through November 2025 were reported as incomplete by March 2026 (claims 3587–3589). These setbacks suggest integration capacity constraints that may affect infrastructure buildout.

Year-end 2026 targets now present a critical test. Claims 8281 and 7652 project 1.85 GW and 1.7 GW of active capacity respectively by December 31, 2026—roughly 200 to 350 MW above current levels in approximately four months. This timing is tight but not implausible given industry deployment cadences. The long-term 2030 target of 8 GW (claims 7872, 7889, 7629) requires sustained delivery and implies roughly 1.5 GW per year growth over the next four years.

Third-party indicators offer partial validation. Recent partnership announcements—NVIDIA frameworks (claim 3892), multi-year deals with Hudson River Trading (multiple August 2026 claims), Anthropic, and Backblaze storage commitment (claim 7448)—suggest customer confidence and revenue locked in. The Indonesia facility commitment (claim 155) targets 360 MW by end 2028, a realistic timeline for greenfield deployment in Southeast Asia.

The gap between active and contracted capacity remains the primary concern. CoreWeave must convert 2.7 GW of contracted commitments into operational infrastructure while maintaining execution on M&A. Monitor fourth-quarter 2026 capacity reporting, Indonesia project permitting status, and any guidance revisions on the 8 GW by 2030 target. Merger failures and acquisition delays suggest the company may face capital allocation or integration headwinds that could pressure timelines.

CoreWeave's facility roadmap shows acceleration in headline targets alongside execution inconsistencies. Organic capacity growth from 2023 to 2025 tracked steadily—70 MW (claim 3551) to 360 MW (claim 3552) to 850 MW (claim 3553)—but the 2030 ambition jumped sharply from 5+ gigawatts (claim 3893, January 2026) to 8+ gigawatts (claim 7629, August 2026) within seven months. The 4.2 GW contracted pipeline announced in August (claim 8268) provides nominal justification, yet the jump itself suggests either opportunistic expansion or revised capacity counting rather than organically validated demand. Year-end 2026 guidance muddied in a single week: August 17 targeted 1.85 GW active (claim 8281), August 20 restated 1.7 GW (claim 7652), and August 24 claimed 1.5 GW already live (claim 1445). These figures contradict or collapse into one another depending on whether they reference targets or actuals, creating ambiguity about real-time deployment status. Acquisition execution stumbled on smaller infrastructure bets. The Marimo, OpenPipe, and Monolith acquisitions promised for September through November 2025 (claims 3577, 3578, 3579) were marked delayed by March 2026's quarterly filing. The Core Scientific merger, revised multiple times—April 2026 (claim 4731), October 2025 (claim 4517)—ultimately terminated in January (claim 4166). These delays suggest either integration complexity or capital discipline, but they pile up without formal explanation. Geographic diversification accelerated late: Indonesia's 360 MW (claim 155, promised by December 2028), Lancaster's 300 MW (claim 7874), Kenilworth's 250 MW (claim 7873), and Stockholm (claim 7875) were all disclosed in August 2026, giving little runway to validate site readiness. The Indonesia timeline provides the longest sightline—a 2.5-year window carries real risk given typical build cycles. Over the next 90 days, watch whether CoreWeave crosses 1.7–1.85 GW active capacity. Reaching mid-2 gigawatt range by year-end from 1.5 GW in early September requires adding 200–350 MW quarterly, a step change from the 122 MW average quarterly addition in 2025. Permits and utility interconnection agreements for Indonesian and Lancaster sites should begin surfacing if timelines hold. Partnership deployment rates—the Hudson River Trading (claim 7463) and Anthropic (claim 7630) agreements signed in August—will signal whether bookings translate to utilization pressure or contractual optionality.

CoreWeave's capacity trajectory shows steep acceleration from 70 MW in December 2023 to 850 MW by December 2025, with contracted capacity now standing at 4.2 GW (claims 3551, 3553, 8262). The company targets 1.7 to 1.85 GW active by year-end 2026 and 8 GW by 2030, but delivery credibility faces headwinds from several timeline slippages. The Core Scientific merger announced in May 2025 with a promised October 2025 close (claims 4517, 4331, 4313) was terminated in January 2026 without prior disclosure of execution risk (claim 4166). Three platform acquisitions—Marimo, OpenPipe, and Monolith—promised for September through November 2025 remained undisclosed as delayed until the March 2026 filing (claims 3577, 3578, 3579). The DDTL 3.0 commitment termination date shifted from July 2026 without explanation (claim 4593). These reversals suggest internal capacity constraints or strategic pivots occurring faster than public communication cycles.

Recent August announcements reveal aggressive customer lock-in: a $22.4 billion OpenAI contract, multi-year agreements with Anthropic and Hudson River Trading, and $335 million from Backblaze (claims 1242, 7630, 1820, 7448). CoreWeave claims near-term capacity is fully booked (claims 1494, 6721), creating a critical dependency on new facility activation. The 1.7–1.85 GW target requires roughly doubling utilization in six months, relying on Kenilworth (250 MW), Lancaster (300 MW), and overseas expansion. Indonesia's first Asia-Pacific site is promised for end-2028 at 360 MW (claim 155), a 2.5-year deployment window that lacks visible permitting evidence or ground-truth updates. The gap between contracted (4.2 GW) and active capacity (0.85 GW as of early 2026) is substantial; meeting the 1.85 GW year-end target would still leave 2.4 GW of signed capacity unlit. Over the next 90 days, watch for actual power draw and energy procurement announcements, first-phase completion milestones at Kenilworth and Lancaster, any disclosure of customer attrition or postponement, and quarterly commentary on whether the contracted backlog remains firm or is being renegotiated.

CoreWeave's capacity ambitions have escalated sharply since mid-year. In May 2025, the company cited 1.6 GW of contracted power (claim 4973); by March 2026, it announced 3.1 GW to deploy (claim 3554); by August 2026, that target doubled to 8 GW by 2030 (claims 7868, 7629). Near-term deployment paints a different picture: 70 MW (end 2023), 360 MW (end 2024), 850 MW (end 2025), with August forecasts targeting 1.7–1.8 GW by year-end 2026 (claims 7652, 1489). That rate of growth—roughly 450 MW over one year—would require 6–7 years to reach the 8 GW target, assuming linear expansion.

Acquisition momentum has stalled. OpenPipe, Marimo, and Monolith were announced in March 2026 with promised close dates (September, October, and November 2025, respectively) already past, flagged as delayed at announcement (claims 3588, 3587, 3589). Weights & Biases, announced May 2025 and priced at $1 billion in August (claim 4516), shows no completion confirmation across subsequent filings. The Core Scientific merger, delayed multiple times beginning July 2025 (claim 4731, promised April 2026), was terminated outright by October (claim 4166). DDTL 3.0's commitment termination date (promised July 31, 2026) is marked delayed (claim 4593).

Partnership depth is notable: Anthropic (claim 7630), Hudson River Trading (multiple references, August claims 7598–7605), Backblaze ($335M, claim 7448), and NVIDIA (co-engineering, claims 1501, 1778). OpenAI is cited for $22.4 billion in contract value (claim 1242, August 2026). Yet these agreements remain unverified by independent infrastructure signals—permit filings, interconnection queue entries, or third-party on-the-ground reporting are absent from the record.

Geographic footprint announcements are recent (Kenilworth, NJ 250 MW; Lancaster, PA 300 MW; Stockholm colocation; Indonesia 360 MW targeting 2028–end claim 155) but lack corroborating permit or construction evidence. Watch through December for three markers: (1) whether actual deployed capacity meets the 1.7–1.8 GW year-end forecast, now just four months out; (2) public confirmation of Weights & Biases integration or announcement of deal termination; (3) any permit or interconnection queue filings for Indonesia, Kenilworth, or Lancaster that signal material progress on the 8 GW roadmap.

CoreWeave has demonstrated significant year-over-year capacity growth, expanding from 70 MW (Dec 2023) to 360 MW (Dec 2024) to 850 MW (Dec 2025)—a 12x increase over two years. However, recent capacity guidance shows moderation: claims from August 2026 set end-of-year 2026 targets between 1.7 and 1.85 GW (claims 7652, 7888, 1489), representing 2–2.2x growth rather than the higher multiples of prior periods. The company's announced 8 GW target by 2030 remains consistent across claims 7868, 7872, 7889, 7629, 943, and 1490 but requires sustained 25%+ annual growth from a modestly expanding base.

A pattern of slipped M&A milestones suggests operational constraints. The Core Scientific merger (initially claim 4731, due 2026-04-07) was terminated entirely by October 2025 (claim 4166). Monolith, OpenPipe, and Marimo—all promised in Q3 2025 (claims 3579, 3578, 3577)—showed [delayed] status by March 2026 (claims 3589, 3588, 3587). Weights & Biases acquisition appears repeatedly (claims 4516, 3562, 3547, 3586, 3576) without closure signals. These delays contrast sharply with locked customer commitments: OpenAI ($22.4B, claim 1242), NVIDIA ($6.3B through 2032, claim 4352), Backblaze ($335M over five years, claim 7448), and Hudson River Trading multiyear agreements (claims 1820, 7463). Near-term capacity is described as sold out (claim 1494).

Geographic expansion announcements lack specificity. Indonesia's 360 MW deployment extends to 2028-12-31 (claim 155); New Jersey (250 MW, claim 7873), Pennsylvania (300 MW, claim 7870), and Sweden colocation sites (claims 7871, 7875) have no stated delivery dates. Current active capacity stands around 1.5 GW (claim 1445), requiring 200–350 MW deployment within four months to hit year-end 2026 targets. The firmware maturity of NVIDIA Vera Rubin and Spectrum-X networking (claims 1773, 7600) and the 128,000 GPU multi-plane claim (claim 1497) lack independent validation.

Watch for: (1) December 2026 capacity realization—whether CoreWeave closes the gap from 1.5 GW to stated 1.7–1.85 GW targets, signaling execution credibility; (2) Indonesia site progress and interconnection queue entries, which would evidence the 2028 timeline; (3) any further target revision downward, indicating demand or infrastructure constraints.

CoreWeave has demonstrated rapid capacity growth but faces execution uncertainty across multiple vectors. From December 2023 through year-end 2025, operational footprint expanded nearly six-fold, from 70 MW (claim 3551) to 850 MW (claim 3553). The company's stated trajectory targets 1.7 to 1.85 GW active capacity by 2026 year-end (claims 7652, 7888), implying roughly doubling from the 2025 baseline. However, this growth masks deeper complications in the M&A pipeline. CoreWeave's October 2025 Core Scientific merger terminated (claim 4166) after a promised October 1 close date (claim 4517) slipped and failed entirely. Additionally, claims filed March 1, 2026 indicate delayed acquisition closures: OpenPipe and the Monolith deal both missed original 2025 deadlines (claims 3588, 3589), suggesting integration velocity challenges.

The company has escalated its 2030 target significantly. January 2026 communications referenced 5+ gigawatts by 2030 (claim 3893), but August announcements pushed this to 8 GW (claims 7872, 7629, 7889). This 60 percent upward revision in seven months warrants scrutiny—growth targets often retract rather than expand if underlying unit economics deteriorate.

Recent partnership announcements, concentrated in late August 2026, included major commitments: a multi-year compute deal with Anthropic (claim 7630), a $22.4 billion OpenAI contract (claim 1242), and a $335 million storage partnership with Backblaze (claim 7448). These are substantive, though claim entries reference Hudson River Trading repeatedly without clear delineation of distinct deals versus redundant reporting. The Indonesia Data Center is promised for December 31, 2028 at 360 MW (claim 155), representing geographic expansion but offering no near-term relief.

Critical near-term metrics: CoreWeave must demonstrate 1.7+ GW in active capacity by year-end 2026 to maintain credibility. The Polaris Forge 1 facility (400 MW contracted, claim 370) timeline remains unspecified. Acquisition closures of Marimo, OpenPipe, and Monolith warrant monitoring for further delays. Watch whether the 8 GW 2030 target gets revised downward as integration pressures mount or whether independent evidence—permit filings, power purchase agreements, utility interconnection queue positions—validates the accelerating deployment pace.

CoreWeave's announced capacity has accelerated sharply, growing from 70 MW at end-2023 (claims 3551, 3567) to 850 MW by end-2025 (claims 3553, 3569), yet the 2026 trajectory shows deceleration. Mid-year targets for 2026 year-end ranged from 1.7 GW (claim 7652) to 1.85 GW (claim 7888), representing a two-fold gain rather than the five-fold growth of the prior two years. This suggests either capacity constraints or revised internal projections, though the company has not explicitly walked back the 8 GW by 2030 commitment (claims 7868, 7872, 7629, 1490).

Promised acquisition timelines have not held. The Core Scientific merger, initially promised by October 2025 (claim 4517), was terminated in October (claim 4166). OpenPipe, Marimo, and Monolith acquisitions all missed original delivery windows—September, October, and November 2025 respectively (claims 3588, 3587, 3589)—yet appear ongoing as of the most recent filings without explicit revised dates. Weights & Biases acquisition claims span from May 2025 through March 2026 (claims 4974, 4516, 3586, 3576), suggesting protracted negotiations rather than failure.

Ground-level evidence partially supports claimed pace. Polaris Forge 1 claims 400 MW contracted capacity (claim 370); Kenilworth, New Jersey and Lancaster, Pennsylvania announce 250 MW and 300 MW respectively (claims 7869, 7874). Indonesia's Greater Jakarta deployment commits 360 MW by end-2028 (claim 155), a substantial but distant milestone. Partnership announcements—NVIDIA's $6.3 billion commitment through 2032 (claim 4352), Anthropic's multi-year deal (claim 7630), and OpenAI's stated $22.4 billion contract value (claim 1242)—indicate demand support, though execution risk remains. The company noted near-term capacity as sold out (claim 1494), a bullish signal if actual deliveries materialize.

In the next 90 days, track whether CoreWeave closes the Weights & Biases acquisition and provides completion dates for OpenPipe, Marimo, and Monolith. Assess whether 2026 year-end capacity lands within stated ranges or undershoots. Monitor Indonesia's first Asia-Pacific facility for site acquisition and initial build notices. Watch for revised guidance on the 2030 8 GW target, especially if H2 2026 deployment lags the stated path.

CoreWeave presents an accelerating capacity deployment trajectory that has scaled meaningfully since mid-2025, though with notable execution friction on the acquisition front. On the infrastructure side, the company reported 70 MW of active capacity at end-2023, grew to 360 MW by end-2024, and reached 850 MW by end-2025 (claims 3551–3553). The contracted pipeline has expanded correspondingly: from 1.6 GW announced in May 2025 (claim 4973) to 3.1 GW by March 2026 (claim 3559) to 4.2 GW by August 2026 (claim 1456). The company is targeting 1.7–1.8 GW of active capacity by year-end 2026 (claims 7652, 1489), then 8 GW by 2030 (claims 1490, 6738)—roughly a 4.7x multiplier over four years.

Date slippage has been persistent on acquisitions rather than infrastructure. The Core Scientific merger, promised via multiple announcements with closing dates through October 2025, was terminated entirely in October (claim 4166). Three acquisitions—OpenPipe, Marimo, and Monolith—each slipped from original Q3 or Q4 2025 closure windows with no announced revised dates or withdrawal as of March 2026 (claims 3577–3579, 3588–3589). By contrast, infrastructure timelines have held or revised cleanly: the Indonesia facility is confirmed for 360 MW deployment by end-2028 (claim 155), and near-term capacity is marked fully booked (claim 6721).

Capacity figures have been revised without explicit acknowledgement of the change. Contracted power jumped from 1.6 GW to 3.1 GW in an eight-week window spanning March 2026, suggesting either delayed disclosure of earlier contracts or material new wins. The active-to-contracted ratio remained roughly 20–36 percent through August 2026 (1.5 GW active, 4.2 GW contracted), implying significant deployment lag ahead even at current ramp rates.

No independent corroboration is evident from permits, interconnection queues, or third-party facility surveys in the data set. In the next 90 days, watch whether CoreWeave closes the three delayed software acquisitions or formally abandons them, whether active capacity reaches the stated 1.7–1.8 GW run-rate by December, and whether any of the major partnerships (NVIDIA's 6.3 billion commitment, Anthropic, OpenAI's 22.4 billion contract value) show signs of meaningful utilization across live deployments.

CoreWeave's capacity targets have scaled dramatically and repeatedly within months. The company grew from 70 megawatts at the end of 2023 to 850 megawatts by end-2025, and now targets 1.7 to 1.8 gigawatts by December 2026 (claims 3551, 3553, 1489, 6737). The 2030 target jumped from 5 gigawatts announced in January to 8 gigawatts announced in August—a 60 percent upward revision in under seven months (claims 3893, 7629). This pattern reflects real customer demand acceleration, but the frequency of target increases also suggests initial planning was conservative.

Current active capacity sits at 1.5 gigawatts against 4.2 gigawatts contracted, with the company stating near-term capacity is fully booked (claims 1455, 1456, 6721). The 200 to 300 megawatt gap between current deployed and end-2026 target represents a compressed deployment schedule—feasible but tight. This timing concentration creates execution risk if any single facility slips.

Timeline performance shows mixed results. The DDTL 3.0 commitment termination originally promised for July 2026 now carries a delayed flag (claim 4593). Core Scientific merger negotiations collapsed entirely after running from May through October 2025 (claims 4731, 4166). The company then announced two new major facilities—Indonesia and Polaris Forge 1—both in August 2026, with only Indonesia carrying an explicit delivery date: December 2028 for 360 megawatts in Greater Jakarta (claims 3013, 155). Polaris Forge 1 lists 400 megawatts contracted but no timeline (claim 370).

Public statements anchor expectations to booked capacity rather than deployed capacity, which creates a measurement problem: if 4.2 gigawatts is contracted against a 2030 target of 8 gigawatts, the company is discussing sales success rather than delivery progress. The data provided contains no independent verification—no permit records, interconnection queue positions, or engineering milestones to calibrate announced timelines against real-world constraints.

Watch over the next 90 days whether active capacity reaches 1.7 to 1.8 gigawatts or undergoes further revision, whether Indonesia or Polaris files for grid interconnection, and whether 2030 targets shift again as new customer contracts land.

CoreWeave's announced capacity trajectory shows aggressive near-term ambitions paired with softer long-term commitments. From 70 MW at end-2023 to 1.5 GW active capacity by mid-2026, the company has demonstrated roughly 20x growth in two years (claim 1794). This acceleration is material, yet the gap between active (1.5 GW) and contracted capacity (4.2 GW) suggests announced demand significantly outpaces delivered infrastructure—a pattern common in AI compute but worth monitoring for execution risk.

Near-term forecasts remain unchanged. CoreWeave restated its end-of-2026 target at 1.8 GW as recently as August 25 (claim 1795), implying only 300 MW of new active capacity needed in the final four months of the year. Given mid-year stood already at 1.5 GW, this represents a material deceleration from the 2023–2026 pace and warrants independent verification as year-end approaches. The company's August 24 manufacturing change notification (claim 2525) signals active infrastructure work but provides no capacity update.

Partnership announcements clustered heavily in late August, particularly the repeated Hudson River Trading deal (claims 958, 1234, 1790, 1820, 1502, 1809 across five days) and the 22.4-billion-dollar OpenAI contract (claim 1242). The Hudson River Trading repetition raises questions about whether these represent distinct phases of a single partnership or overwording of a single commitment. Two partnerships withdrawn—Backblaze (claim 1243) and Anyscale (claim 1769)—hint at deals that either failed to close or proved immaterial enough to de-announce.

Long-term targets show a consistent 8 GW by 2030 goal across multiple statements (claims 943, 1490). The intermediate 3.2 GW target for end-2027 (claim 1796) extends the slowing growth trajectory and suggests CoreWeave may be normalizing expansion rates as it scales.

Watch for: achievement of the 1.8 GW EOY target in Q4 2026; third-party evidence of Indonesia's 360 MW deployment timeline (due 2028); NVIDIA partnership deliverables and validation timeline on Vera Rubin hardware (claim 1773); whether the contracted-to-active capacity ratio begins closing, signaling actual customer workload migration. The next 90 days will clarify whether CoreWeave's mid-year pace was sustainable or an anomaly.