jain.com — data centers, connectivity and security, analyzed daily

Tag: convertible notes

  • Everyone’s Watching New AI Silicon. CoreWeave Found 19.8% in Racks It Already Had.

    Everyone’s Watching New AI Silicon. CoreWeave Found 19.8% in Racks It Already Had.

    TL;DR · 30-second read

    The Short Version

    CoreWeave rents out computers that run artificial intelligence. On September 16 it published results from an independent industry test and said its machines handled AI requests faster than those of rival rental companies.

    The more interesting result is quieter: on machines it installed months ago, it got about 20 percent more work out of each chip purely by improving its own software. No new hardware.

    That matters because those chips cost a fortune. Squeezing more out of the ones you own, instead of buying more, is where the profit in this business actually sits.

    CoreWeave on September 16, 2026 published its submissions to MLPerf Inference v6.1, the industry benchmark round released the same day by MLCommons. Writing on the company blog, Shadi Saba said CoreWeave entered the Datacenter Closed division’s Available category across four NVIDIA platforms — HGX B200, HGX B300, GB200 NVL72 and GB300 NVL72 — and four model families: Qwen3-VL-235B-A22B, DeepSeek-R1-671B, GPT-OSS-120B and Llama 2 70B. Headline figures include 16,635 tokens per second per GPU on GPT-OSS-120B in the offline scenario, more than 1.19 million tokens per second from a single GB300 NVL72 rack, and 944,902 tokens per second on Llama 2 70B in the server scenario. NVIDIA, in a blog post of its own, announced that its next-generation Vera Rubin NVL72 made its MLPerf debut in the same round.

    One day later, on September 17, CoreWeave filed an 8-K disclosing a proposed $3.0 billion offering of convertible senior notes due 2033, with an option for a further $500 million, alongside an Equity Distribution Agreement covering at-the-market share sales and collared forward sales through eleven banks. The two disclosures, a day apart, describe the same business from opposite ends: what the fleet produces, and what it costs to keep building.

    Executive Summary

    CoreWeave’s MLPerf claims break into two categories, and they are not equally interesting. The absolute throughput records — 1,196 queries per second on the multimodal Qwen3-VL model, over 1.16 million tokens per second on GPT-OSS-120B from one rack — are a function of running NVIDIA’s newest shipping silicon, GB300 NVL72, at scale. Any operator with the same racks and the same power envelope is in that fight.

    The second category is harder to replicate. CoreWeave says that comparing its 72-GPU v6.1 submission against its 64-GPU v6.0 submission on GB200 NVL72, derived per-GPU server throughput on DeepSeek-R1-671B rose 19.8% in the five months between rounds. That gain came from serving-stack tuning, scheduling and operations, not from a new chip. If it holds in production, it changes the revenue a depreciating asset can generate across its life — which is the number that underwrites everything else.

    That includes the financing. The 8-K filed September 17 shows CoreWeave raising $3.0 billion in converts on top of an existing stack of senior notes carrying coupons from 8.500% to 9.750%. Capital that expensive is serviced by output per installed accelerator, and software gains on hardware already paid for are the cheapest output there is.

    The 19.8% That Didn’t Require a New Chip

    Strip the benchmark round down and one number does real work. CoreWeave reports that its DeepSeek-R1-671B per-GPU server throughput on NVIDIA GB200 NVL72 rose 19.8% between MLPerf Inference v6.0 and v6.1 — five months apart — through changes to its serving stack, scheduling and operations rather than new hardware. The mechanisms it names are specific: topology-aware scheduling through SUNK, which pins an inference workload inside a single high-bandwidth NVLink domain so that tokens moving between accelerators never leave the fastest interconnect in the rack, and fleet-level health monitoring through Mission Control. For a mixture-of-experts model like DeepSeek-R1 — an architecture that routes each token through a subset of specialised sub-networks, so traffic between chips is constant — placement is not a detail. It is the workload.

    Who this affects is everyone financing GPUs on a multi-year schedule. An accelerator bought in early 2026 has a fixed capital cost and a depreciation clock that does not care how efficiently it runs. If per-GPU output rises roughly a fifth without capex, the cost per token served falls by a comparable amount across the remaining life of that asset, and the residual economics of the installed base improve rather than decay. That is a materially different asset story from the one where value arrives only with the next generation of silicon.

    The claim deserves its caveats stated plainly, and CoreWeave states some of them itself. Per-GPU throughput is a derived metric — total system throughput divided by accelerator count — and the company notes it is not verified by MLCommons. The comparison also spans two differently sized submissions, 64 accelerators in v6.0 against 72 in v6.1, so scaling behaviour and configuration changes are folded into the same percentage as the software work. The direction is credible and the mechanism is named; the precision of the figure is CoreWeave’s own arithmetic, not the benchmark’s finding.

    The Unit of Account Moved From the Chip to the Rack

    The GB300 NVL72 is not a server. It is a rack in which 72 accelerators share one NVLink domain, meaning they address each other at memory-like speeds rather than over conventional networking. CoreWeave’s results read accordingly: over 1.16 million tokens per second in the server scenario and over 1.19 million offline on GPT-OSS-120B from a single rack, 944,902 and 1,136,100 tokens per second on Llama 2 70B, and 1,196 queries per second on Qwen3-VL-235B-A22B. The company’s GB200 NVL72 submission reached 901,058 tokens per second on GPT-OSS-120B in the server scenario — the same architecture, one generation back, and a visible step down.

    The consequence lands on the people who build and power facilities rather than the people who buy models. When the smallest sensible deployment unit is a rack drawing far more than a legacy hall was designed for, throughput per rack becomes a proxy for how much revenue a given square metre and a given megawatt can carry. Operators comparing a rack-scale NVL72 deployment against clusters of HGX B200 or B300 servers — which CoreWeave also submitted, and which it says led all HGX entries from cloud providers — are really comparing two different power-density and liquid-cooling commitments, not two spec sheets.

    Available Is the Word Doing the Most Work

    CoreWeave’s submissions sit in the Datacenter Closed division’s Available category. Closed means every entrant runs the same reference implementation, so results are comparable rather than a contest in custom optimisation; Available means the system can actually be obtained. That second qualifier matters this round, because NVIDIA used the same release to announce that its next-generation Vera Rubin NVL72 made its MLPerf debut. A benchmark headline generated by a platform that is debuting is not a price a buyer can procure against today; CoreWeave’s four platforms are all Blackwell and Blackwell Ultra, the generations in production now.

    For anyone running a procurement process, the practical read is to check three fields before treating any v6.1 result as a quote: which silicon generation produced it, whether the entry sits in the Available category, and whether the metric quoted is a submitted number or a derived per-GPU figure. CoreWeave also stresses that it ran on the same production clusters and images customers get rather than a benchmark-tuned rig — a genuinely meaningful distinction, though one the benchmark itself does not audit.

    Throughput Is the Collateral Story

    The day after the benchmark post, CoreWeave’s 8-K disclosed a proposed $3.0 billion offering of convertible senior notes due April 1, 2033, sold under Rule 144A to qualified institutional buyers, with an option for up to $500 million more and capped call transactions to limit dilution on conversion. A concurrent Equity Distribution Agreement adds at-the-market share sales and collared forward sales through a syndicate that includes Deutsche Bank, Goldman Sachs, J.P. Morgan, Morgan Stanley and Citigroup. Interest and conversion terms, the release states, are set at pricing.

    The comparison worth making is internal. CoreWeave’s own press release lists the existing obligations the new notes will rank alongside: senior notes at 9.250%, 9.000%, 9.750%, 9.625% and 8.500%, and earlier convertibles at 1.75%. Converts price far below straight debt because the buyer is paid partly in equity optionality rather than cash coupon — cheaper carry now, potential dilution later, which the capped calls are bought to blunt. For a business whose costs are dominated by accelerators, the servicing capacity for any of it comes back to tokens per installed GPU.

    That is the honest link between two documents filed a day apart, and it is a structural one rather than a causal one: a benchmark result is a performance claim, not a revenue disclosure, and nothing in the filings ties the offering to the MLPerf round. But the neocloud model runs on both halves at once — demonstrable output per accelerator to justify the capital, and access to capital to buy more accelerators. Software gains of the kind CoreWeave reported on already-deployed GB200 racks are the half that does not require either.

    Background

    CoreWeave, based in Livingston, New Jersey and listed on Nasdaq as CRWV, is a specialist cloud provider built around NVIDIA accelerators for AI training and inference — a category often called a “neocloud,” distinguished from general-purpose hyperscalers by its narrow focus and its heavy reliance on debt and equity markets to fund hardware. The company was recently named a Visionary in the Gartner Magic Quadrant for Cloud AI Infrastructure.

    MLPerf, administered by the MLCommons consortium, has become the closest thing the industry has to an audited comparison of AI serving performance, which is why vendors submit and why the categories matter. Inference — running a trained model to answer requests — is now the larger and more cost-sensitive half of the AI workload, because a model is trained once but served continuously. That shift is what makes tokens per accelerator, rather than raw chip specifications, the metric operators and their lenders increasingly watch.

    Sources

    Source: MLPerf® Inference v6.1 Results: CoreWeave Leads Providers — CoreWeave’s September 16, 2026 post detailing its Datacenter Closed, Available submissions across four NVIDIA Blackwell platforms. See also NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut, NVIDIA’s account of the same benchmark round.

    Primary sources: CoreWeave, Inc. Form 8-K filed September 17, 2026 (Items 7.01 and 8.01, disclosing the proposed convertible notes offering and the Equity Distribution Agreement); Exhibit 99.1 — CoreWeave Announces Proposed $3.0 Billion Convertible Senior Notes Offering; and Exhibit 99.2 — CoreWeave Investor Presentation, September 2026. Full benchmark tables are published by MLCommons at mlcommons.org.

  • WhiteFiber Buys 2 North Carolina Sites as $865M Nscale Lease Begins

    WhiteFiber Buys 2 North Carolina Sites as $865M Nscale Lease Begins

    TL;DR · 30-second read

    The Short Version

    WhiteFiber, a company that rents out computing power for artificial intelligence, is spending $60 million on two properties in North Carolina to build more data centers, which are warehouses full of computers. The sites are expected to come with enough electricity for tens of thousands of homes, and that much power is increasingly hard to get.

    It already has customers. A 10-year deal worth about $865 million has started paying, and the company has signed more than $550 million in other contracts since May.

    The catch: it still loses money, $15 million last quarter, and it issued new shares, so each existing share now owns a smaller slice of the company.

    WhiteFiber, Inc. (WYFI), an operator of AI data centers and GPU cloud services, attached an investor presentation to an 8-K filed with the SEC on September 14, 2026. Minichart’s summary of the presentation reported that on August 16 the company signed a definitive agreement to buy two North Carolina industrial properties, designated NC-2 and NC-3, for $60 million in cash. The sites are expected to provide at least 60 MW of initial gross electrical capacity, with potential for roughly 198 MW over time. Confirmation of that capacity is a closing condition, and closing is expected in the fourth quarter of 2026.

    The presentation also reported more than $550 million of new cloud contract value booked since May 2026. Billing has begun at the NC-1 campus under Nscale’s 10-year, 40 MW lease, which is worth about $865 million. WhiteFiber described a pipeline of roughly 1.5 GW, completed an August exchange that retired about 86% of its 2031 convertible notes, and reported a second-quarter net loss of $15.0 million.

    Executive Summary

    WhiteFiber’s September investor update shows a company reorganizing around contracted, long-duration revenue. At its core is the Nscale anchor lease at NC-1. That single contract averages about $86.5 million a year over its term, which is close to the company’s entire $94.5 million of combined revenue for the twelve months ended June 30. Around it sit five disclosed cloud contracts and a sixth undisclosed one, spread across Canada, Iceland, Paris and WhiteFiber’s own U.S. fleet.

    The $60 million North Carolina purchase is the physical follow-through. WhiteFiber is using its backlog as the case for buying powered land before it has binding tenants, and it says non-binding letters of intent backed by investment-grade credit are in hand. The strategy is coherent. The open questions are about execution. Revenue depends on construction, GPU deliveries and customer acceptance. Net losses widened year over year, and the note exchange added about 6.3 million shares.

    From GPU Renter to Landlord

    WhiteFiber has two businesses. One rents GPU computing capacity by the hour or by contract, which is its cloud services business. The other leases data center space and power to other AI companies, which is its data center business. Cloud is still the larger segment, with $79.6 million of trailing-twelve-month revenue against $15.0 million from data centers. The data center line, however, grew 10.7 times from FY2024, and the Nscale lease is set to change the mix sharply once it ramps.

    The economics of that shift are the story. A 10-year lease for 40 MW of IT load gives visibility that hourly GPU rental cannot. IT load is the power available to customers’ servers. Gross capacity, by contrast, also covers cooling and other building overhead. WhiteFiber says Phase 1 of NC-1 is fully contracted and that Duke Energy has completed work to deliver the initial 54 gross MW. Management believes the site could eventually reach 200 gross MW. If it does, NC-1’s current tenant would occupy only part of a campus that already has a utility relationship, which is a meaningful advantage when grid connections are the industry’s main bottleneck.

    The presentation says initial billing began in the third quarter and that full revenue contribution is expected in the same quarter as the facility reaches contractual capacity. That timing is still subject to deployment, commissioning and customer acceptance. Whether the ramp lands on schedule is the near-term test investors will watch in the next quarterly report.

    Buying Power, Not Just Buildings

    At $60 million for a minimum of 60 MW of initial gross capacity, WhiteFiber is paying roughly $1 million per initial megawatt for NC-2 and NC-3. That falls to about $300,000 per megawatt if the full 198 MW materializes. The price covers the properties, not the fit-out. Turning industrial buildings into AI-ready halls with high-density power and liquid or advanced cooling costs far more than the land. The acquisition is best read as an option on scarce, electrified sites rather than as the full cost of new capacity.

    Two structural protections stand out. First, the power capacity has to be confirmed as a condition of closing, so WhiteFiber should not end up owning sites without the electricity it is paying for. Second, the sites are about 55 miles from NC-1, which could let the company reuse staff, contractors and regional know-how. The target of initial ready-for-service capacity in the third quarter of 2027 is aggressive for conversions of this kind, and it depends on the deal closing on time in late 2026.

    Demand support is real but not yet binding. WhiteFiber cites non-binding letters of intent backed by investment-grade credit from prospective customers. That is a stronger signal than a speculative build. It is still short of a signed lease like Nscale’s at NC-1.

    How Solid Is $550 Million of Cloud Backlog?

    The five disclosed contracts total about $538 million, and the unpresented July win of roughly $16 million brings the figure above $550 million. The contracts differ considerably. Terms run from two years (Hyperbolic / Modal Labs, about $17 million) to five years (the Paris deployment, over $160 million, and a Canada inference contract, about $88 million). Prime Intellect’s roughly $165 million three-year deal is the largest single item. Averaged evenly across their terms, the five contracts would add up to around $150 million a year at full run rate. That figure is simple arithmetic, not company guidance.

    Most of this backlog requires new hardware: NVIDIA B300 systems in Canada and VR200 systems for WhiteFiber’s first Vera Rubin deployment in the U.S. The capital cost of those GPUs is the central economic question. The Paris contract is the most de-risked on the page, with an investment-grade customer, customer prepayments and project-level financing. The Hyperbolic / Modal Labs deal is the least capital-intensive because it uses H200s already in the fleet and needs no new GPU spending.

    Contract length matters because GPU generations turn over quickly. A three-year contract on current-generation hardware can recover a large share of the equipment cost within the term. What the systems earn afterward is harder to predict. The revenue-sharing upside on the Canada inference contract adds potential gains, and it also means that part of that contract’s value depends on the customer’s success.

    The Balance Sheet Behind the Build

    The August note exchange raised about $298.5 million in net proceeds from new $310 million 2032 convertible notes. WhiteFiber spent roughly $118.5 million in cash to retire about 86% of its 2031 notes, leaving around $180 million net. It also issued about 6.3 million shares, roughly 16% of the 38.85 million basic shares outstanding afterward. The conversion price on about 91% of the principal involved is set at a 25% premium to a $27.07 reference price. Separately, WhiteFiber holds a zero-strike call on 5.91 million shares, bought in January 2026 for about $120 million. Arrangements of that kind are generally used to offset dilution.

    The operating picture shows growth ahead of profit. Second-quarter revenue rose about 54% to $28.8 million, and adjusted EBITDA increased to $5.5 million from $3.3 million. Over the same period the net loss widened to $15.0 million from $8.8 million, which included a $5.0 million impairment of capitalized software. That pattern is common for infrastructure companies spending ahead of contracted revenue. It does mean the company’s expansion depends on continued access to capital markets and project financing.

    Background

    WhiteFiber is a New York-headquartered AI infrastructure company that went public in 2025 after being carved out of Bit Digital. It runs two lines of business: GPU cloud services, which rent high-end NVIDIA computing capacity to AI developers, and data center services, which lease powered space to other AI operators. Its footprint spans Canada, Iceland, the United States and a Paris region deployment, with its flagship owned campus, NC-1, in North Carolina.

    WhiteFiber belongs to a group of so-called neoclouds, specialist GPU providers that have grown alongside demand for AI training and inference. Across the sector, the limiting factor has shifted from chips to electricity, because utility grid connections can take years to secure. Companies that control powered sites are therefore increasingly able to sign long-term contracts. WhiteFiber’s 10-K for 2025 was filed on March 26, 2026, and its 10-Q for the second quarter on August 12, 2026.

    Sources

    Source: WhiteFiber Signs $60 Million Deal for Two North Carolina Data Center Sites (Minichart), a summary of WhiteFiber’s September 2026 investor presentation. Related coverage: WhiteFiber unveils $865M Nscale anchor deal, >$550M cloud wins and 1.5GW pipeline (TradingView) and WhiteFiber Updates Investors on AI Data Center Strategy (TipRanks).

    Primary sources: WhiteFiber, Inc. Form 8-K Current Report, filed September 14, 2026; Exhibit 99.1: Investor Presentation of WhiteFiber, Inc. as of September 2026.

  • IREN Closes $3 Billion Convertible Notes Offering to Fund AI Infrastructure Buildout

    IREN Closes $3 Billion Convertible Notes Offering to Fund AI Infrastructure Buildout

    IREN, the publicly traded bitcoin miner repositioning itself as an AI infrastructure company, has closed a $3 billion convertible notes offering, according to a report from The Block dated May 16, 2026. The raise ranks among the largest capital events yet for a company making the miner-to-AI transition.

    Convertible notes are debt instruments that can later be exchanged for shares, letting companies borrow at lower interest rates in exchange for potential future dilution. For IREN, the proceeds arrive as the company accelerates its push into AI compute and data center capacity.

    Executive Summary

    The headline fact is simple: $3 billion in fresh capital, closed, for a company that began life mining bitcoin and now markets itself as an AI infrastructure provider. Capital at that scale is not raised to sustain a mining operation — it is raised to build data centers, buy GPUs, and sign the power and construction commitments that AI compute demands. The offering’s closure, rather than mere announcement, means the money is in hand.

    Why it matters: the miner-to-AI pivot has been the dominant strategic story in the bitcoin mining sector for over two years, but most pivots have been announced in press releases rather than financed in capital markets. A closed $3 billion convertible offering is a market verdict of sorts — institutional buyers were willing to lend against IREN’s AI story at convertible terms. It suggests the pivot narrative, at least for the largest and most credible miners, has graduated from concept to bankable strategy.

    That said, the report is brief, and the substantive details that determine whether this is cheap or expensive capital — coupon, conversion premium, hedging arrangements, and specific use of proceeds — are not spelled out in the source. Readers should treat the raise as a strong signal of momentum while withholding judgment on its economics.

    From Mining Rigs to GPU Halls: Why the Pivot Attracts Capital

    Bitcoin miners and AI data center operators need the same scarce ingredients: large blocks of grid power, industrial land, cooling, and the operational muscle to run energy-dense facilities. Miners spent a decade securing exactly those assets, often in power-rich regions where capacity was cheap. When AI demand exploded and grid interconnection queues stretched to five years or more in many markets, energized megawatts became the bottleneck — and miners suddenly held an asset the AI industry desperately wants.

    The pivot is not automatic, however. A mining facility is engineered for cheap, interruptible, low-redundancy compute; an AI data center serving enterprise or hyperscale customers typically requires far higher reliability, denser networking, and liquid cooling. Converting one into the other is a genuine construction project, not a rebranding exercise. That is precisely why a raise of this magnitude is the tell: $3 billion is conversion-and-buildout money.

    The Economics of Convertible Debt in an AI Land Rush

    Convertible notes have become the financing instrument of choice for capital-hungry compute companies. The logic is straightforward: a company with a volatile, high-momentum stock can borrow at a much lower cash interest cost than straight debt would demand, because lenders are partly paid in the option to convert into equity if the stock rises. For shareholders, the trade-off is potential dilution down the road.

    For a company straddling bitcoin mining and AI — two of the most volatility-prone narratives in public markets — convertibles are arguably the only large-scale debt market reliably open. Traditional project finance lenders want long-term contracted revenue; a miner mid-pivot often cannot yet show it. The willingness of convertible buyers to absorb $3 billion of IREN paper says the market is pricing meaningful upside into the equity, but it also means the company is, in effect, pre-selling a slice of that upside to fund the buildout.

    Winners, Losers, and the Sorting of the Mining Sector

    The miner-to-AI transition is sorting the sector into tiers. Companies with large, well-located power portfolios and access to capital markets can finance real conversions; smaller miners without either are left competing in a bitcoin mining business whose economics tighten with every halving — the programmed event that cuts mining rewards roughly every four years. A raise like this one widens that gap: capital compounds, because funded buildouts attract customers, and customer contracts attract cheaper follow-on capital.

    For the broader data center industry, well-capitalized former miners are becoming genuine competitors for AI workloads, particularly in the cost-sensitive middle of the market. Incumbent operators retain advantages in reliability track record and enterprise relationships, but the energized-power advantage is real, and $3 billion buys a lot of construction.

    What a Closed Raise Does and Does Not Prove

    It is worth being precise about what this announcement substantiates. It proves investor appetite: sophisticated buyers committed $3 billion. It does not, by itself, prove customer demand for IREN’s AI capacity, the economics of its contracts, or the timeline on which the capital becomes revenue-generating infrastructure. The AI infrastructure boom has featured both genuinely contracted buildouts and speculative capacity built ahead of demand, and a financing headline cannot distinguish between them. The next meaningful data points will be customer agreements, deployment milestones, and disclosed note terms — not the raise itself.

    Background

    IREN began as Iris Energy, an Australian-founded bitcoin miner that listed publicly and built a portfolio of power-intensive data center sites, emphasizing access to low-cost and renewable energy. Like much of the mining sector, it faced the structural squeeze of bitcoin’s halving cycle, which periodically cuts mining revenue, just as the generative AI boom created enormous demand for exactly the kind of powered data center capacity miners control.

    Over the past two years, the miner-to-AI pivot has become the defining strategic story of the sector, with a handful of large operators securing AI and high-performance computing deals while smaller players remained pure miners. Capital markets have increasingly rewarded the pivot, and large convertible note offerings have become the sector’s signature financing tool for funding GPU purchases and data center conversion at scale.

    Source: IREN closes $3 billion convertible notes offering as Bitcoin miner’s AI infrastructure push accelerates — The Block’s May 16, 2026 report on IREN’s completed $3 billion capital raise.