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

Category: AI Infrastructure

  • Applied Digital’s $36B AI Lease Book Is Now a 1.41 GW Delivery Test

    Applied Digital’s $36B AI Lease Book Is Now a 1.41 GW Delivery Test

    TL;DR · 30-second read

    The Short Version

    Applied Digital, a company that builds giant computer warehouses for artificial intelligence, says customers have signed leases worth about $36 billion. Before five new deals this past year, that figure was $7 billion.

    The buildings covered by those leases are designed to draw about as much electricity as roughly a million American homes.

    The catch: in this business, rent generally starts only once a building is finished, powered and handed over. The company now pays its president partly on how much capacity actually gets delivered, not just how much gets signed.

    Applied Digital (Nasdaq: APLD) reported in its definitive proxy statement, filed with the SEC on September 25, 2026, that five leases signed during fiscal 2026 (the year ended May 31, 2026) raised its initial-term contracted revenue from $7 billion to approximately $36 billion, before any customer renewal or expansion options. Stock Titan, summarizing the filing, reported that total contracted critical IT load stands at 1.41 gigawatts (GW).

    The same proxy sets out five-year equity awards for CEO Wes Cummins, President Jason Zhang and CFO Saidal Mohmand, and calls stockholders to a virtual annual meeting on November 4, 2026, to elect six directors, ratify CBIZ CPAs P.C. as auditor, and cast an advisory vote on executive pay.

    Executive Summary

    The headline number is a roughly fivefold jump in contracted revenue in a single fiscal year: from $7 billion to about $36 billion, driven by five leases. That places Applied Digital among the AI data center developers with the largest disclosed lease backlogs, and it is tied to 1.41 GW of critical IT load, meaning the electricity available to the servers themselves, before cooling and building overhead.

    What makes the filing more than a backlog update is how the company has chosen to pay its leadership against it. President Jason Zhang’s performance shares are measured half on contracted revenue and half on ready-for-service delivery, CFO Saidal Mohmand’s on net operating income, and CEO Wes Cummins’s on share-price hurdles the proxy describes as interdependent with the other two. In effect, the board has written into its pay plan that signing leases is only half the job; turning them into energized, revenue-producing capacity is the other half.

    That framing matters because it arrives after only 52.6% of votes supported the company’s 2025 say-on-pay proposal, and because the conversion from contract to delivered megawatt is where AI infrastructure schedules most often slip.

    From $7 Billion to $36 Billion in One Fiscal Year

    The proxy’s figure is initial-term contracted revenue: the total rent customers have committed to pay over the base length of their leases, excluding renewal and expansion options. On that measure, five leases signed between June 2025 and May 2026 added roughly $29 billion. Dividing $36 billion by 1.41 GW gives a blended figure of about $25.5 million of initial-term revenue per contracted megawatt. That is simple arithmetic on two disclosed totals, not an annual rate, because the lease terms are not broken out; a 15-year lease and a 10-year lease at the same headline value imply very different yearly income.

    The excluded options cut both ways. They represent potential upside if tenants renew or expand, but they are not commitments, and the $36 billion should be read as the contracted floor rather than a projection of lifetime value. For industry readers, the useful comparison is not against other developers’ press-release totals but against Applied Digital’s own delivery rate: how quickly 1.41 GW of contracted load becomes billing capacity.

    Contracted Is Not Delivered: Why 1.41 GW Is the Real Test

    In wholesale data center leasing, rent typically begins when a hall or building is declared ready for service (RFS): powered, cooled, commissioned and handed to the tenant. Until then, a signed lease is a promise on both sides, and the developer carries the construction cost, the equipment procurement and the utility interconnection work. A 1.41 GW critical IT load also understates the site-level power requirement, since total facility draw is higher once cooling and electrical losses are added; the proxy lists power usage effectiveness (PUE), the ratio of total facility power to IT power, among its key terms for exactly this reason.

    Applied Digital’s compensation design makes this conversion explicit. Jason Zhang’s performance stock units (PSUs, shares that vest only if targets are hit) are split evenly between contracted revenue and ready-for-service delivery. Saidal Mohmand’s rest on two net operating income hurdles, which can only be cleared once leased capacity is operating and paying. Wes Cummins’s three share-price hurdles are described as interdependent with the success of the other two awards. The board has, in other words, tied the top three executives’ largest equity grants to the same chain: sign, deliver, earn.

    The people most exposed to that chain sit outside the company as well. Tenants are planning AI deployments around RFS dates. Lenders and equity holders are underwriting construction against rent that has not yet started. Utilities, transformer makers and cooling suppliers are the ones who have to deliver the physical inputs on schedule. A backlog this large is a strong demand signal; whether it is a strong earnings signal depends on those delivery dates, which is the measure the company itself has chosen to be judged on.

    A Five-Year Pay Plan After a 52.6% Vote

    The awards to Cummins, Zhang and Mohmand use a five-year performance and vesting horizon, a 25/75 split between time-based restricted stock units (RSUs) and performance-based PSUs, and a two-year holding requirement after PSUs vest. They are intended to be the three executives’ only company equity awards for five years from grant. The Compensation Committee says the structure reflects longer project cycles and that stockholder feedback, following 52.6% support for say-on-pay in 2025, informed its fiscal 2026 decisions. Support at that level is well below the threshold at which proxy advisers commonly expect a board to show it has responded.

    Both sides of the debate have a fair point to test. Supporters of front-loaded, multi-year grants argue they match compensation to multi-year construction timelines and reduce annual grant churn. Skeptics of the design generally ask how demanding the hurdles are, how share-price targets interact with dilution from future capital raises, and whether a single large grant concentrates risk. The November 4 advisory vote will indicate which reading stockholders favor.

    The proxy also notes that on April 29, 2026, the board formed separate committees of directors it identifies as independent to review and approve or ratify related-party transactions with ChronoScale and Base Electron expected to exceed, or potentially exceed, $120,000 in a fiscal year. Establishing dedicated committees is a governance safeguard; the scale and nature of those relationships are what stockholders will want to follow.

    Background

    Applied Digital is a Dallas-based developer and operator of data centers that began by hosting high-power computing for cryptocurrency mining and has since repositioned toward purpose-built campuses for AI and high-performance computing, leased to large compute customers on long-term contracts. Its fiscal year ends May 31.

    The wider AI infrastructure market has shifted toward multi-year, gigawatt-scale leases in which developers secure land and power, build to a tenant’s specification and begin collecting rent once capacity is ready for service. Backlogs have grown quickly across the sector; the pace at which utilities, equipment suppliers and construction crews can turn those contracts into energized capacity has become the practical constraint on how fast they convert into revenue.

    Sources

    Source: Five leases lifted Applied Digital (APLD)’s initial-term contracted revenue beyond $7 billion. – Stock Titan, a summary of Applied Digital’s definitive proxy statement reporting initial-term contracted revenue rising from $7 billion to about $36 billion across 1.41 GW of contracted critical IT load.

    Primary sources: Applied Digital Corp. Definitive Proxy Statement (Form DEF 14A), filed September 25, 2026 — SEC EDGAR.

  • Applied Digital’s $270M Brookwood Pledge Makes Local Terms Part of AI Campus Cost

    Applied Digital’s $270M Brookwood Pledge Makes Local Terms Part of AI Campus Cost

    TL;DR · 30-second read

    The Short Version

    A company called Applied Digital plans to spend $3.2 billion on a huge building full of computers for artificial intelligence near Brookwood, a town in Alabama. At full size it would have about as much floor space as 20 American football fields. It should start running in 2028.

    The notable part is the deal with the town. Applied Digital promised to pay the local area $270 million over 20 years. That is on top of an estimated $131 million in taxes, even though only about 100 people will work there permanently.

    Applied Digital has confirmed that Delta Forge 2, a $3.2 billion, 210MW AI data center, will be built on a 1,300-acre site near Brookwood in Alabama’s Tuscaloosa County, Data Center Dynamics reported on September 25, 2026. The campus is planned at 1.2 million square feet at full build-out, with operations expected to begin in 2028. The tenant is an unnamed customer that Applied Digital describes as an “investment grade hyperscaler,” meaning one of the very large cloud and AI platform operators.

    The company is committing $270 million in community benefits payments to the Tuscaloosa County Economic Development Authority over 20 years, alongside an estimated $131 million in tax revenue for the area. The project is expected to support up to 1,000 construction jobs and 100 permanent roles. The Tuscaloosa Thread reported that the county board approved incentives for the project unanimously.

    Executive Summary

    Applied Digital said in June that it had signed a customer for a new 210MW data center, but it did not say where. Naming Brookwood completes a basic picture of the project: a customer, a site, a budget, a target date and local approval. It also puts a campus on the company’s map outside North Dakota, where its Polaris Forge 1 and 2 campuses sit.

    The most informative number is not the $3.2 billion headline. It is the $270 million community payment. That is roughly 8% of the stated investment and larger than the project’s estimated tax contribution. For a facility with 100 permanent jobs, the negotiated fiscal package, more than employment, is what the host community receives in return. That makes the local agreement a real cost of bringing AI capacity online, one that developers have to budget for alongside land, power and construction.

    The $270 Million Pledge Is a Line Item, Not a Gesture

    Applied Digital’s commitment of $270 million over 20 years to the Tuscaloosa County Economic Development Authority is large enough to register in the project’s economics. Against the $3.2 billion stated investment, it is roughly 8%. Spread across 210MW of capacity, it comes to about $1.3 million per megawatt. Averaged over two decades, it is about $13.5 million a year, although the company has not said whether payments are level, front-loaded or tied to construction milestones.

    The pledge is also larger than the other fiscal figure attached to the project, an estimated $131 million in tax revenue. The period behind that tax estimate was not specified, so the two numbers are not a strict like-for-like comparison. Still, of the two local benefits disclosed, the negotiated payment is the bigger one, not the ordinary tax base. For site selection, that changes the calculation. When a county’s approval comes with a multi-decade cash commitment on this scale, the community agreement becomes part of the cost of delivering capacity. It has to be priced into the lease the same way power and construction are.

    Three groups are affected. Other developers pitching large AI campuses to rural and suburban counties may now find this package used as a reference point in negotiations. Local economic development authorities gain a contracted revenue stream instead of relying only on assessed taxes, which can be reduced by abatements. The tenant’s lease pricing ultimately has to cover the developer’s full cost of delivering the site.

    Why Capital-Heavy, Job-Light Projects Pay Their Way in Cash

    A $3.2 billion project that supports 100 permanent roles represents about $32 million of investment per permanent job. That ratio is typical of large data centers. Most of the spending goes to buildings, electrical equipment and cooling systems, not payroll. The up-to-1,000 construction jobs are substantial, but they are temporary by definition.

    This is the mechanism behind the size of the payment. Traditional economic development deals trade tax incentives for jobs. A data center cannot offer many jobs, so it offers direct fiscal commitments instead. Brookwood mayor Joe Barger’s comments point in that direction. The mayor framed the project around “strengthening our schools” and benefits “for generations to come,” which reflects a long-term revenue case more than a hiring case. The unanimous board vote suggests the county found the exchange acceptable. The value of the incentives the county granted in return has not been published, so the net fiscal benefit cannot yet be calculated.

    A Portfolio Built on a Few Very Large Tenants

    Applied Digital, founded in 2021 as Applied Blockchain, now reports a contracted portfolio of five AI Factory campuses totaling 1.4GW of critical IT load. Critical IT load is the power available to servers and networking gear, excluding cooling and other facility needs. The company reports approximately 2.15GW of grid-connected utility power behind that portfolio. The gap of roughly 750MW is the extra electricity a site must secure to run cooling and other facility systems and to keep some headroom. Securing grid power, not just signing tenants, is the harder half of this business.

    Chairman and CEO Wes Cummins called Delta Forge 2 part of “repeatable AI infrastructure delivered at scale.” Part of that claim is supported by disclosed facts: multiple campuses under contract, one tenant named (CoreWeave at Polaris Forge 1), and repeat business from one customer. The same high-investment-grade hyperscaler that signed at Delta Forge 1 also took 300MW at Polaris Forge 3. The company’s statement that it has a “proven track record of delivering on our commitments to our host communities” is not accompanied by delivered-capacity or payment-history figures. Much of the portfolio is concentrated in a small number of unnamed hyperscale tenants. That concentration brings strong credit, but it also means the business depends heavily on a few counterparties.

    What Has to Go Right by 2028

    Delta Forge 2’s 2028 start follows Polaris Forge 3, where initial operations are targeted for August 2027. That puts overlapping large builds on the company’s schedule. Hitting both dates depends on things not addressed in the announcement: the utility and grid connection for a 210MW load in Tuscaloosa County, procurement of long-lead electrical equipment, and financing for a project of this size. The move to Alabama widens the company’s geographic reach, but it also means building a new set of utility, contractor and permitting relationships far from its North Dakota base.

    Background

    Applied Digital was founded in 2021 as Applied Blockchain and has repositioned itself as a developer of data centers for artificial intelligence and high-performance computing. It builds large campuses and leases them to a small number of major tenants. Its Polaris Forge 1 campus in Ellendale, North Dakota, is set to be leased to CoreWeave. Polaris Forge 2 in Harwood, North Dakota, is largely committed to an unnamed investment-grade hyperscaler. Polaris Forge 3, on 600 acres in an unnamed northern state, will provide 300MW to the same hyperscaler that signed at Delta Forge 1, with initial operations targeted for August 2027.

    AI data centers differ from conventional ones mainly in power density. Training and running large AI models requires far more electricity per rack of servers, so developers compete on how much grid power they can secure and how quickly. That has made site selection a negotiation over power, land and local fiscal terms, as much as over proximity to users.

    Sources

    Source: Applied Digital reveals $3.2bn Delta Forge 2 AI data center will be built in Alabama (Data Center Dynamics). Location, scale, timeline and community commitments for Applied Digital’s 210MW Brookwood campus.

    Additional coverage: County Board Unanimously OKs Incentives for $3.2 Billion Applied Digital Data Center in Brookwood (Tuscaloosa Thread); $3.2 Billion AI Data Center Campus Announced in Brookwood (WBMA); Applied Digital confirms $3.2 billion data center campus in Brookwood (WBRC); Applied Digital confirms Brookwood data center (WBRC); Massive $3.2B AI data center coming to Brookwood (WVTM Channel 13).

  • Cadence’s 2.5x A14 Gain Is in Analog Layout; AI Chip Timing Now Rides on IP

    Cadence’s 2.5x A14 Gain Is in Analog Layout; AI Chip Timing Now Rides on IP

    TL;DR · 30-second read

    The Short Version

    Before a new computer chip can be made, engineers draw and redraw its design in specialized software. Cadence, which makes that software, says its tools are now approved for the next, even smaller manufacturing process at Taiwan Semiconductor Manufacturing Company, the world’s biggest contract chipmaker.

    Cadence says one tricky part of the job now needs up to two and a half times fewer redraws. That could speed up the custom artificial intelligence chips that giant cloud companies build. The catch: many ready-made parts Cadence sells for such chips are still unfinished, and some are not due until the end of this year.

    Cadence Design Systems announced on September 23 an expanded partnership with TSMC, according to the company’s release as published by Wccftech. Cadence’s digital full flow, signoff flow and custom/analog flows are now certified for TSMC’s A16 and A14 process technologies, and its analog, digital and signoff flows are certified for N2P. Cadence says Virtuoso Studio’s automated layout generation and design migration to A16 and A14 reduces design iterations by up to 2.5x.

    On intellectual property (IP) — pre-designed circuit blocks that chipmakers license rather than build themselves — Cadence reported a UCIe-64G chiplet interface taped out on A14, LPDDR6 memory-interface IP in development, and A14 foundation and mixed-signal IP expected in Q4 2026. It also cited support for TSMC’s SoW-X wafer-scale flow with hundreds of chiplets, a UALink demonstration on N3P, and design-services firm Global Unichip Corp. (GUC) as a customer.

    Executive Summary

    Tool certification is the gate that lets chip designers commit real products to a new manufacturing process. It means TSMC has validated Cadence’s software against its A14 and A16 design rules and models, so the layouts, timing and power numbers the tools produce can be trusted to match what the fab will build. Cadence has now cleared that gate for its digital, signoff and custom/analog flows on both nodes.

    It matters because the next generation of custom AI accelerators — the chips hyperscalers design in-house to run alongside or instead of merchant GPUs — will migrate to these nodes, and the time between a node becoming available and silicon shipping in volume is set by tools, IP and packaging together. Cadence’s headline 2.5x figure is narrower than it first reads: it applies to automated custom and analog layout and migration, and it is stated as an upper bound.

    The more consequential detail is in Cadence’s own status table. With tools certified, most of its A14 IP is still in development, with foundation and mixed-signal blocks due in Q4 2026. For design teams relying on that IP, the A14 schedule question has moved from software qualification to IP readiness. On the manufacturing side, TSMC’s latest 6-K shows its board approved US$16,035 million for advanced-technology capacity machinery in August — the fab is spending; design-side readiness determines how quickly customers can use it.

    The 2.5x Applies to Analog Layout, Not the Whole Chip

    Cadence ties its “up to 2.5X” reduction in iterations specifically to Virtuoso Studio’s design-rule-correct layout generation and automated design migration to A16 and A14 — the custom and analog side of the flow. Analog circuits handle continuous electrical signals rather than ones and zeros: the transceivers that move data off a chip, the physical interfaces to memory, clocking and power-management blocks. Digital logic can largely be re-mapped to a new node by synthesis and place-and-route software; analog layouts have traditionally been redrawn by specialists at each node, because exact transistor placement affects how the circuit behaves.

    That makes analog a sensible target for automation. “Design-rule-correct” means the generated layout already obeys the foundry’s geometric rules on spacing and widths, cutting the draw-check-fix loop that consumes much of a migration. Shortening the workstream that scales worst with headcount is a real gain. But Cadence does not state the baseline — fewer iterations than which previous method, on which designs — and “up to” marks a best case. No comparable multiplier is claimed for the digital full flow, where most of an AI accelerator’s transistors sit.

    For anyone estimating time to market, the 2.5x is a figure about iterations in one workstream, not a claim that chips ship 2.5 times sooner. Cadence’s own summary of the benefit — “faster time to market and improved PPA” (power, performance and area) — is not quantified.

    Tools Are Certified; the A14 IP Ladder Is Still Being Climbed

    Semiconductor IP climbs a ladder before customers trust it: designed, taped out (sent to the fab to make test chips), tested in the lab when silicon returns, then characterized — measured across voltage and temperature so its behavior is documented. Cadence’s UCIe-64G chiplet interface shows that ladder across nodes in a single announcement: characterized on N3P, in the lab on N2P, taped out on A14.

    On A14 itself, the list is mostly future tense. UCIe-64G is taped out. LPDDR6 IP is in development. Artisan foundation IP — standard cells, the basic logic building blocks, plus GPIO and memory compilers — and mixed-signal IP are in development and expected in Q4 2026. Other HPC and AI memory IP and chiplet protocols have been “prioritized for development.” HBM4/4E and PCIe 7.0 appear in Cadence’s broader portfolio without an A14 status.

    That is the mechanism behind the headline. A custom AI accelerator is mostly compute logic, but it cannot ship without the interfaces that connect it to high-bandwidth memory, sibling chiplets and the network — precisely the analog and mixed-signal blocks still in progress on A14. With certified tools in hand, a design team relying on Cadence’s A14 IP now faces a different schedule question: not whether the software is qualified, but when each block reaches proven silicon. Teams can source some blocks from TSMC or other IP vendors — Synopsys has announced its own TSMC partnership on agentic AI and advanced design — so this is a statement about Cadence-based flows and about where the gating item has moved, not a ceiling on the whole market.

    Packaging Is Where the AI System Gets Assembled

    The announcement gives as much space to packaging as to transistors. Cadence says its tools support TSMC’s SoW-X system-on-wafer flow with hundreds of chiplets and millions of bumps (microscopic solder connections), and CoWoS-L-style packages five times the reticle size. The reticle is the largest area a lithography scanner prints in one exposure, which effectively caps the size of a single die; a 5x-reticle package combines the silicon of several maximum-size dies in one module. A certified die-to-die auto-routing flow targets UCIe requirements across CoWoS architectures.

    Cadence also cites TSMC-COUPE, a photonic engine that converts electrical signals to light close to the chip, enabled with its 224G SerDes IP and aligned with UALink, UltraEthernet, ESUN and co-packaged optics (CPO). These are the links that turn individual accelerators into clusters, and their readiness shapes what data center operators will be powering, cooling and cabling. Cadence discloses no power or thermal figures for these designs.

    TSMC’s capital decisions point the same way. In a 6-K filed September 24, TSMC reported that its board approved US$16,035 million for advanced-technology capacity machinery, US$4,791 million for advanced packaging, mature and/or specialty capacity, and US$8,616 million for real estate and capitalized leased assets in August. The filing does not isolate packaging from the combined category or allocate spending by node, so it supports the direction of travel rather than any A14-specific conclusion.

    Who Gains From Shorter Node Ramps

    The most direct beneficiaries are the firms that build custom accelerators for hyperscalers. GUC, named as a customer, delivers custom silicon for AI accelerators and hyperscale systems on TSMC’s leading-edge nodes. Its senior vice president, Louis Lin, said success on N3P and N2P depends on tightly optimized execution, “especially when working with the early‑adopters where schedule risk is highest.” His remarks address N3P and N2P; the announcement names no A14 customer.

    For hyperscalers, being early on a new node matters because each generation aims to deliver more computing per watt, which compounds across fleets of data center racks. But the constraints stack — certified tools, proven IP, packaging capacity, wafer capacity — and a design moves at the pace of the slowest. On A14, Cadence has cleared the first and is partway through the second.

    Some claims need more support before buyers can weigh them: the UALink demonstration is described as “industry-leading” without published bandwidth, latency or power data, and “agentic AI chip design flows” are named but not defined. Both are marketing language until measured results follow.

    Background

    Chips are designed with electronic design automation (EDA) software — the programs engineers use to describe, lay out and verify circuits before a factory builds them. Cadence Design Systems and Synopsys are the two largest EDA suppliers, and both also license IP blocks such as memory and chiplet interfaces. TSMC, the largest contract chipmaker, manufactures chips designed by other companies, including many AI accelerators; its leading-edge roadmap runs from N3 and N2 to A16 and A14.

    Every new process requires design tools and IP to be requalified before customers can commit products to it, which is why certification announcements cluster around each node. Hyperscalers increasingly design custom AI accelerators, often with design-services partners such as GUC, and those chips are now assembled from multiple chiplets in advanced packages — TSMC’s 3DFabric family, including CoWoS and SoIC — making packaging and die-to-die links as important to the schedule as transistor density.

    Sources

    Source: Cadence Locks In TSMC A14 Tool Certification, Cutting Design Iterations By 2.5X As It Races Toward Agentic AI Chips — Cadence’s announcement of A14/A16/N2P tool certification, new IP and 3DFabric support. Also referenced: Synopsys and TSMC Partner to Accelerate AI Systems Innovation with Agentic AI and Advanced Design.

    Primary sources: Taiwan Semiconductor Manufacturing Company, Form 6-K filed 2026-09-24 (report for the month of August 2026, including board-approved capital appropriations).

  • CleanSpark’s AI Pivot Costs 7.875% to Fund, Even With Meta Guaranteeing Rent

    CleanSpark’s AI Pivot Costs 7.875% to Fund, Even With Meta Guaranteeing Rent

    TL;DR · 30-second read

    The Short Version

    CleanSpark, a company best known for mining bitcoin, is borrowing about $2.3 billion to finish a huge building full of computers for artificial intelligence. The tenant is a company owned by Meta, the parent of Facebook and Instagram, and Meta itself guarantees 20 years of rent.

    Even with that promise, lenders are charging nearly 8% interest, or about $179 million a year.

    Why so expensive? The building is not scheduled to start paying rent until late 2027, and lenders are betting it gets finished. That interest rate shows what it costs a bitcoin company to become a landlord for artificial intelligence.

    On September 18, 2026, CleanSpark (Nasdaq: CLSK) said in an 8-K that its wholly owned subsidiary CSDC Finance I, LLC had priced $2.276 billion of 7.875% senior secured notes due 2031 at 98.5% of face value. The deal is expected to close on September 25. It is larger than the $2.227 billion CleanSpark proposed a day earlier. The notes were sold privately to large institutional investors under Rule 144A and to non-U.S. buyers under Regulation S.

    The proceeds will finish building CleanSpark’s Sandersville Facility, reimburse the company for equity it has already put into the project, and fund debt service reserves. The investor presentation CleanSpark furnished on September 17 describes Sandersville as a 175 MW AI data center campus. It is leased for 20 years to Anviran, LLC, a wholly owned Meta subsidiary, and Meta guarantees the rent and operating expenses. First-phase rent is targeted for Q4 2027.

    Executive Summary

    CleanSpark has raised project-level debt for its first large AI hosting campus, and the price is specific: a 7.875% coupon, sold at a 1.5-point discount. The notes are secured by first-priority liens on the Sandersville project and the issuer’s equity, a completion guarantee from CleanSpark backs them, and part of the proceeds funds cash reserves for interest payments. Behind all of that sits a 20-year triple-net lease guaranteed by Meta, a company the presentation values at about $1.7 trillion.

    The notable feature is the gap between the tenant and the terms. Meta’s credit is not in question. The coupon and the protective package reflect what is still unfinished: a building that is not scheduled to earn rent until late 2027, owned by a sponsor that is moving from bitcoin mining into data center development. For operators trying to turn mining power into AI capacity, this deal is a concrete, disclosed reference point for what that conversion costs to finance.

    Meta Signed the Lease. Lenders Still Priced the Build.

    On paper, Sandersville is about as strong a credit story as a data center gets. The investor presentation CleanSpark furnished on September 17 describes these terms:

    • a 20-year base lease with Anviran, LLC, a Meta subsidiary, plus two five-year extension options;
    • Meta itself as guarantor of rent and operating expenses;
    • a triple-net structure, meaning the tenant covers taxes, insurance and upkeep;
    • a 3% annual rent escalator;
    • about $6.6 billion of base-term contract value.

    To underline the tenant’s strength, the presentation cites Meta’s $117 billion of first-half 2026 revenue.

    Even so, the notes carry a 7.875% coupon, the fixed annual interest rate. They were also sold at 98.5 cents on the dollar, which by rough estimate lifts investors’ effective yield to a little over 8%. Bondholders also received three protections. They hold first-priority liens on substantially all assets of the issuer and its guarantor subsidiary, plus the issuer’s equity. Reserves funded from the proceeds cover debt service. And CleanSpark gives a completion guarantee, a promise to supply more money if the proceeds cannot finish the facility on time. Each protection guards against the same thing. The question is not whether Meta can pay. It is whether the building gets delivered.

    The timeline explains why. First-phase rent is targeted for Q4 2027, so interest accrues for more than a year before rent arrives, and the lease pays only once the capacity exists. For a bitcoin miner moving into AI hosting, a lease with a hyperscaler (one of the giant cloud and internet platforms) turns megawatts into something lenders will finance. It does not, on its own, earn the borrowing costs of the investment-grade tenant. Construction and sponsor risk sit in between. In this deal, CleanSpark carries the construction risk directly through its guarantee.

    The Per-Megawatt Math Behind $2.276 Billion

    The presentation puts development cost at about $11.9 million per megawatt of critical IT load, which is the power delivered to the servers themselves, not counting cooling and building overhead. For 175 MW, that implies roughly $2.1 billion. At 98.5% of face value, $2.276 billion of notes raises about $2.24 billion before fees, which is more than that implied total cost. That fits the stated uses: finishing the build, reimbursing CleanSpark’s prior equity contributions, and funding reserves. In practice, the debt is sized to fund close to the whole project, and CleanSpark recovers part of what it has already spent.

    Annual interest on the notes comes to about $179 million, or roughly $1.0 million per megawatt per year. Net operating income (NOI) is rent minus operating expenses, and under a triple-net lease it is nearly all of the rent. CleanSpark projects average NOI of about $330 million a year, or roughly $1.9 million per megawatt. That covers interest about 1.8 times on average. The early years are thinner. The presentation’s schedule shows NOI of about $220 million in 2028, the first full rent year, which covers interest about 1.2 times, before the escalator widens the margin.

    The resizing is worth noticing. CleanSpark proposed $2.227 billion on September 17 and priced $2.276 billion the next day. At the 98.5% price, the larger deal raises slightly more than the original principal target. The extra $49 million of face value more than offsets the discount. CleanSpark has not said why it changed the size. Either way, the issuer now pays 7.875% on the larger amount.

    What 7.875% Means for Miners Sitting on Megawatts

    CleanSpark says it controls more than 1.8 GW of power, land and data centers. Sandersville’s 175 MW of IT load is less than a tenth of that figure. The comparison is only rough, because IT megawatts and grid megawatts are measured differently. The company’s own forward-looking disclaimers list the risks of converting mining sites into data centers: permitting and utility constraints, construction delays, cost overruns, financing, supply chains and tenant performance. Sandersville is where that list first meets billions of dollars of committed capital.

    The 2031 maturity matters as much as the coupon. The notes come due about four years after targeted rent start, while the lease runs 20 years. If the campus is delivered and paying rent by then, the construction risk that shaped this coupon will be gone, which could support cheaper refinancing. That outcome still depends on rates and credit markets in 2031. The refinancing risk moves to a later date; it does not disappear.

    For other miners pursuing the same path, this is one data point, not a trend. It shows that an investment-grade lease plus secured, project-level debt can fund a full AI build, at this price and with the parent still responsible for completion. Three groups are affected. CleanSpark shareholders get part of their equity back but take on a contingent obligation. Noteholders hold a claim on a single asset. Other operators now have a public benchmark that their own conversions will be compared against.

    Background

    CleanSpark, based in Las Vegas and listed on Nasdaq as CLSK, built its business on bitcoin mining. It now describes itself as a data center developer that controls more than 1.8 GW of power, land and data centers across the United States. Its investor materials describe an evolving strategy: expanding into data center development, high-performance computing and AI, including by converting some of its bitcoin mining sites into data centers for other uses.

    The move follows a wider pattern. Bitcoin miners hold what AI developers most need: land with large grid connections already in place. Turning a mining site into an AI data center, however, takes a different class of building, with heavier cooling, redundant power and stricter uptime requirements. It also takes financing on a different scale. That is why Sandersville’s financing terms draw attention well beyond CleanSpark itself.

    Sources

    Source: Is CleanSpark (CLSK) Undervalued On Its New 7.875% Notes And Expansion Plans? (Yahoo Finance), coverage of CleanSpark’s new senior secured notes and data center expansion.

    Primary sources: CleanSpark Form 8-K filed September 18, 2026; CleanSpark, Inc. Announces Pricing of $2.276 Billion of Senior Secured Notes (Exhibit 99.1); CleanSpark Form 8-K filed September 17, 2026; Sandersville Investor Presentation, September 2026 (Exhibit 99.1); CleanSpark, Inc. Announces Proposed Offering of $2.227 Billion of Senior Secured Notes (Exhibit 99.2).

  • Hut 8’s $140M Poolin Win Tests AI Power’s $3M-to-$27M-per-Megawatt Gap

    Hut 8’s $140M Poolin Win Tests AI Power’s $3M-to-$27M-per-Megawatt Gap

    TL;DR · 30-second read

    The Short Version

    Hut 8, a company that mines bitcoin and builds computer centers, won a bankruptcy auction for two Texas computer sites with a $140 million offer. A judge still has to approve the sale on September 29.

    The prize is the electricity. Power-ready space at crypto miners has been valued at less than one-ninth of what similar space fetches once an artificial intelligence company rents it. Buying sites that already exist can mean skipping a crowded line to connect to the Texas grid.

    The catch: Hut 8 has not said how much power the sites have, and the Texas governor just ordered a pause on permits for these facilities.

    Hut 8 has won the bankruptcy auction for Poolin’s two Texas data centers, the Pyote and Tarbush sites, with a $140 million bid, Bitbo reported on September 24, 2026, citing court records. The sale still requires approval from the U.S. Bankruptcy Court for the District of New Jersey, which has scheduled a hearing for September 29.

    The sites would add to the roughly 1.5 gigawatts of power capacity Hut 8 already controls in Texas across operating and development-stage sites. The result arrives days after Texas Governor Greg Abbott directed the Texas Commission on Environmental Quality to halt data-center permits until ERCOT, the state’s main grid operator, completes an audit.

    Executive Summary

    Hut 8 placed the winning $140 million bid for two Texas data centers being sold out of Poolin’s bankruptcy. If the court approves the sale on September 29, the Pyote and Tarbush sites join a Texas portfolio of about 1.5 gigawatts that Hut 8 already has operating or in development.

    The deal matters because grid access has become one of the scarcest inputs for both AI data centers and bitcoin mining in Texas. Data centers now account for 87% of ERCOT’s large-load queue, which totals 410 gigawatts, according to CoinShares. An existing site whose power arrangements come with it is a way to acquire capacity without starting at the back of that line.

    Whether $140 million is cheap or expensive cannot yet be judged. Data-center capacity is priced per megawatt, and the megawatt capacity of Pyote and Tarbush has not been disclosed. The purchase also lands just as Texas has paused data-center permitting, which could either raise the value of sites that already hold permits or slow any plan to convert them.

    The $3M-to-$27M Megawatt Gap Behind the Bid

    Data-center capacity is priced per megawatt, meaning the amount of electrical power a site can deliver to the computers inside it. Two recent market marks bracket what sites like Pyote and Tarbush could be worth. At the low end, the energized but unleased capacity of some listed bitcoin miners is valued at less than $3 million per megawatt. At the high end, three fully leased AI facilities in Northern Virginia recently sold for about $27 million per megawatt. Between those two marks sits the cost of moving from one to the other: converting a mining site to AI-grade infrastructure runs roughly $8 million to $15 million per megawatt.

    Stacking those figures shows where the value is. Power at the miner mark plus conversion comes to about $11 million to $18 million per megawatt, against a $27 million mark for leased AI capacity. The residual, roughly $9 million to $16 million per megawatt on these numbers, is what a buyer captures only if it converts the site and signs a tenant. The comparison is directional rather than precise. Northern Virginia is a more mature data-center market, the Virginia sales were of finished and leased buildings, and Texas sites differ in power, land and permitting.

    Where the $140 million lands on that ladder depends on a figure that has not been published: the sites’ capacity. The arithmetic is simple. If Pyote and Tarbush together carry more than about 47 megawatts, Hut 8 bid less per megawatt than the sub-$3 million mark for miner capacity. If they carry less, the bid priced in a premium to that mark, effectively paying in advance for part of the conversion upside. In either case, the auction tests how much buyers will pay for a position on the gap before any tenant exists.

    Buying Around a 410-Gigawatt Queue

    ERCOT’s large-load queue is the line of big electricity customers asking to connect to the Texas grid. CoinShares puts its total at 410 gigawatts, with data centers making up 87% of it. A gigawatt is 1,000 megawatts. A new project joins that line and waits for grid studies and upgrades before it can draw power. An existing data center lets a buyer start further up the line, provided its grid connection and power agreements survive the change of ownership. That is the core case for buying distressed sites: the asset is the interconnection as much as the buildings.

    Governor Abbott’s September 21 directive to halt data-center permits until ERCOT completes an audit, alongside reviews of grid and water impacts, cuts both ways for this deal. If the pause applies only to new applications, sites that already hold their permits become scarcer relative to demand, which strengthens the case for buying over building. If expanding or converting an existing facility requires new or amended permits, the pause could reach Pyote and Tarbush too and delay any AI conversion until the audit concludes.

    On August 10, Hut 8 said it welcomed state scrutiny focused on grid reliability, local resources and community impacts. That statement predates the permit pause. How the company paces its Texas expansion against the audit is now an operational question as well as a policy one.

    A Small Check Against a Large Book

    For Hut 8, $140 million is modest. In the second quarter the company reported 949 megawatts of contracted IT capacity, worth an estimated $26.6 billion in base-term contracts, and said it had secured $7.5 billion in investment-grade financing so far. The bid equals less than 2% of that financing. For a developer at this scale, the binding constraint looks less like capital and more like power that can be connected and permitted, which is what a bankruptcy auction can offer.

    CoinShares estimates AI work can earn about $1.5 million in annual profit per megawatt, compared with $0.5 million for bitcoin mining. That tripling explains the pull toward conversion. On the same figures, though, the extra $1 million per megawatt a year takes roughly 8 to 15 years to repay an $8 million to $15 million conversion, before financing costs. The economics therefore rest on long, contracted leases, which is why a leased Virginia megawatt trades at many times an unleased one and why Hut 8 reports contracted capacity and contract value rather than raw megawatts alone.

    Mining Distress Is Supplying the Sites

    The second quarter was hard on miners. Bitcoin ended June at $58,400, while listed miners faced a weighted-average cash cost of about $75,500 per coin, which left the sector below breakeven overall. Conditions improved by mid-August as bitcoin recovered and a stronger hash price, the expected revenue per unit of computing power, pushed most operators back above cash breakeven. Stretches like that test weaker balance sheets, and assets that change hands through court processes are one result.

    The likely beneficiaries are creditors and estates selling energized sites, because the pool of bidders now includes developers who value power for AI, not only for mining. Listed miners whose capacity carries sub-$3 million marks could see those marks revisited if more auctions clear at higher per-megawatt prices. One auction is a single data point, though. The September 29 hearing, and any disclosure of the sites’ capacity, will say more about the market than the headline price does.

    Background

    Hut 8 is a publicly traded company (Nasdaq: HUT) that began in bitcoin mining and has built a large data-center business alongside it, reporting 949 megawatts of contracted IT capacity in the second quarter and about 1.5 gigawatts of operating or development-stage power capacity in Texas. Poolin is best known in bitcoin mining as a mining-pool operator. Its two Texas data centers, Pyote and Tarbush, are being sold through a bankruptcy proceeding in the U.S. Bankruptcy Court for the District of New Jersey.

    Texas has become a center of both bitcoin mining and AI data-center development because of its large, largely self-contained grid run by ERCOT. Demand to connect has surged, with data centers making up most of ERCOT’s large-load queue. That has made sites with existing power connections valuable and drawn closer scrutiny from state officials over grid reliability and water use.

    Sources

    Source: Hut 8 Wins Poolin’s Texas Data Centers With $140M Bid (Bitbo), on Hut 8’s winning bankruptcy-auction bid for Poolin’s Pyote and Tarbush sites and the Texas market context around it.

  • Everyone Counts AI Packaging Capacity. TSMC’s 2029 Hub Targets Qualification Time

    Everyone Counts AI Packaging Capacity. TSMC’s 2029 Hub Targets Qualification Time

    TL;DR · 30-second read

    The Short Version

    The chips behind AI tools like chatbots are not just manufactured. They are also assembled: several chips are placed side by side or stacked, then wired together into one package. How well that assembly works now affects how fast new AI chips can ship.

    TSMC, the Taiwanese company that makes many of the world’s AI chips, is building a center in southern Taiwan to test new assembly machines and materials before they reach the factory floor. The goal is to catch problems early, instead of fixing them while production is already running.

    Do not expect quick results. The center is due to open in late 2029.

    Kaohsiung’s Baipu Industrial Park broke ground on September 21, 2026. SemiVision reported that TSMC will anchor the park with an advanced-packaging training and validation center. The center is not a production line. It is a facility where TSMC and its suppliers will jointly develop, test and qualify packaging equipment and materials. Earlier plans put TSMC’s initial site at roughly three hectares with two buildings, and operations are expected to begin around the fourth quarter of 2029.

    The park covers about 88.7 hectares, including roughly 53 hectares of industrial land. Around 25 hectares of that are expected to focus on advanced back-end semiconductor equipment and related supply chains. Official estimates suggest the park could eventually generate around NT$300 billion in annual output and more than 4,000 jobs.

    Executive Summary

    TSMC’s Baipu center is built around a reconfigurable “Mini-Loop” validation platform. It will not house a complete CoWoS production line. Instead, selected process modules will be assembled, tested and swapped out as technologies change. CoWoS (Chip-on-Wafer-on-Substrate) is TSMC’s technique for placing an AI processor next to stacks of high-bandwidth memory (HBM) inside a single package, and it sits at the core of today’s AI accelerators.

    The significance lies in what TSMC is separating out. It is moving equipment and materials validation out of its production fabs and into a dedicated site shared with suppliers. Validation is the work of proving that a new tool or material performs reliably enough for mass production. The move is a bet that AI chip supply increasingly depends on how quickly new packaging tools and materials can be qualified, and not only on how many packaging lines exist.

    For suppliers, the promise is earlier entry into TSMC’s development cycle. For AI chip designers, the payoff, if it comes, is fewer surprises when new package designs reach volume production. Neither benefit is expected before late 2029.

    Why the Gating Step Is Qualification Time, Not Floor Space

    Front-end chipmaking, the step where transistors are patterned onto wafers, relies on equipment giants such as ASML, Applied Materials, Lam Research and Tokyo Electron. These companies run large R&D organizations and global field-service teams that help get new tools qualified. Advanced packaging has no equivalent bench strength. Many back-end equipment suppliers are much smaller, yet a long list of technologies is changing at the same time:

    • CoWoS itself
    • Hybrid bonding, which fuses chips copper-to-copper without solder bumps
    • Thermocompression bonding (TCB)
    • Temporary bonding and debonding
    • Wafer and panel handling
    • Automated optical inspection and metrology
    • Underfill, dielectric and thermal materials

    The product cadence on top of that is accelerating. NVIDIA, AMD, Broadcom and custom AI chip (ASIC) developers keep introducing new architectures. HBM is moving from HBM3E toward HBM4 and beyond. Package size, die count, interconnect density and heat load are all rising together. The practical result is that one generation of equipment may not finish qualification before the next product generation arrives. Historically, some packaging tools were still being modified after installation, with engineers tuning them while production was already underway. A line that is being debugged is not yielding at full rate.

    Baipu is TSMC’s structural answer: move troubleshooting upstream, so that equipment, material and integration problems surface before mass production. This is why capacity counts alone understate the constraint. A packaging plant full of tools that are not yet qualified for the next package design cannot ship the next accelerator. Two groups are most exposed. The first is accelerator designers, whose launch timing depends on new package formats. The second is smaller back-end toolmakers, who have lacked a place to prove their equipment at the pace TSMC’s customers now demand.

    The Mini-Loop Is a Capital-Efficiency Choice

    A fixed, full-scale experimental packaging line would require heavy capital expenditure. Because 3D IC, CoWoS, HBM and chiplet architectures are evolving quickly, that line’s configuration could be outdated within a few years. The Mini-Loop concept avoids this. TSMC brings in only the module that needs validating, tests it, modifies it, and then reconfigures for the next job. The validation infrastructure evolves with each technology generation instead of being written off along with the previous one.

    The trade-off is fidelity. A modular loop tests modules. Some integration problems only appear when every step runs together at volume, and those may still surface in production. How much time Baipu actually saves will depend on two things: how TSMC bridges that gap, and whether validation results at Baipu carry directly into production qualification or must be partly repeated in the fabs.

    From Equipment Vendor to Co-Development Partner

    The conventional sequence runs like this: the customer sets a specification, the supplier builds the tool, the tool is installed in the fab, qualification begins, and then the equipment is modified in response to feedback. Baipu aims to put customers, equipment makers, materials companies and research organizations into the same development loop before mass production begins. In that loop, equipment fixes can happen immediately, and materials can move straight into another design of experiments (DOE, a structured test matrix). Software, automation, process recipes and metrology can all be tuned in parallel.

    This is a different advantage from simple proximity. Being near TSMC saves shipping and service time. Entering its development cycle early lets a supplier help shape the specification instead of chasing it. The groundbreaking drew companies from across the stack: TSMC, ASE, Foxconn, ASPEED, WinWay, King Slide, Gudeng, C SUN, Grand Process Technology, Scientech, AblePrint, HIWIN and Gallant Precision Machining. Attendance does not confirm investment. The first phase is expected to include validation work linked to TSMC and its suppliers, along with Foxconn AI-related R&D and manufacturing. A later phase could support further advanced-packaging expansion, including requirements from companies such as ASE.

    Kaohsiung presents Baipu as one link in a southern-Taiwan corridor:

    • Nanzih: advanced semiconductor manufacturing
    • Renwu: packaging and test
    • Ciaotou: materials and critical components
    • Gangshan and Lujhu: existing precision-machinery, metal-processing and fastener industries that the city hopes to pull into the equipment supply chain

    If that works, the regional advantage would not rest on any single fab. It would rest on how tightly manufacturing, packaging, equipment and qualification can iterate over short physical distances.

    What a 2029 Opening Means for Today’s AI Supply

    With operations expected around the fourth quarter of 2029, Baipu sits outside the window of current packaging constraints. Tools and materials qualified there would serve product generations well beyond today’s HBM3E-based parts. Buyers of AI accelerators should not read Baipu as near-term relief. It is an investment in the rate at which packaging technology can change over the following years.

    The headline economic figures are NT$300 billion in annual output and more than 4,000 jobs. These are park-wide official estimates, not TSMC commitments, and they say little about the metric that matters most to the industry: how much shorter qualification cycles actually become. No such target has been published. The argument that time-to-qualification is becoming the gating factor is coherent and fits the direction of packaging technology. Its payoff, however, will only be measurable in how many Baipu-validated tools reach production lines at TSMC, ASE or other manufacturers.

    Background

    TSMC (Taiwan Semiconductor Manufacturing Company) manufactures chips designed by other companies. In the AI era, its role has extended from wafer fabrication into advanced packaging. Packaging is the step where multiple chips, such as a processor and stacks of high-bandwidth memory, are combined into a single package. Technologies such as CoWoS, hybrid bonding and 3D IC stacking have made packaging a performance-critical part of AI accelerators, and HBM, substrates, thermal management, metrology and automation have become increasingly interdependent.

    Southern Taiwan is a focal point of that build-out. Kaohsiung’s semiconductor base already spans Nanzih, Renwu and Ciaotou. The city is positioning Baipu Industrial Park, which broke ground in September 2026, as both an advanced semiconductor equipment and materials validation hub and an AI advanced-manufacturing base.

    Sources

    Source: Beyond Capacity Expansion: TSMC Builds an Advanced Packaging Validation Hub in Kaohsiung (SemiVision), an analysis of TSMC’s planned advanced-packaging validation center at Kaohsiung’s Baipu Industrial Park.

  • Marvell’s 102.4Tbps 2nm Optics Make AI Scale-Out a Watts-Per-Bit Problem

    Marvell’s 102.4Tbps 2nm Optics Make AI Scale-Out a Watts-Per-Bit Problem

    TL;DR · 30-second read

    The Short Version

    Computers that train artificial intelligence have to talk to each other constantly, and today most of that traffic still travels over copper wiring that burns electricity and fades over distance.

    Marvell has shown chips that move the traffic as light instead, built on the most advanced chipmaking process now available. One demonstration moves data fast enough to copy well over a thousand high-definition movies every second.

    The point is not speed. Light uses less power per unit of data, and electricity is the hard limit on how big these computing sites can get.

    Marvell Technology has demonstrated optical interconnect silicon built on a 2nm process for AI data centers, including a co-packaged optics engine running at 102.4 terabits per second and supporting 200G and 400G per electrical lane. 24/7 Wall St. reported on September 21 that the company described the work as an industry first and said it is “pushing the envelope of what is possible in optical tech,” with shares trading around $248 in Monday’s first hour, up 1.5% on the session.

    The demonstration sits on top of a quarter Marvell has already reported: record fiscal second-quarter revenue of $2.739 billion, up 37% year over year, with data center revenue of $2.17 billion, up 46%, now 79% of the total. Guidance filed August 27 put the current quarter at $3.150 billion plus or minus 5%, and raised the fiscal 2027 and fiscal 2028 outlooks. Management has scheduled an Investor Day for October 6 in New York.

    Executive Summary

    The headline number is 102.4 Tbps of co-packaged optics, but the operational story is the denominator. Co-packaged optics — CPO — moves the optical engine off the front panel of a switch and onto the same substrate as the switch or accelerator silicon. That shortens the electrical path a signal has to travel before it becomes light, and shortening that path is primarily a way to spend fewer watts per bit, not a way to go faster.

    That distinction matters because a rack has a fixed power envelope set by the building, not by the silicon roadmap. Every watt the network consumes inside that envelope is a watt unavailable to the accelerators the network exists to connect. As Marvell steps its lane rates from the 800G generation deployed today, past the 1.6T generation now ramping, toward 3.2T, copper’s usable reach shrinks and more of the interconnect budget migrates to optics — which is exactly where the energy cost concentrates.

    The demo also gives silicon substance to forecasts management has already made. On the earnings call, CEO Matt Murphy said “the magnitude of our scale-up optics opportunity next year is much larger than we thought just a quarter ago,” and that Marvell expects scale-across bandwidth requirements to exceed 10 times current front-end data center interconnect networks. What a demonstration does not establish is yield, production timing, field serviceability or a named customer.

    Optics Stopped Being a Bandwidth Problem and Became a Power Budget

    Here is the mechanism the 102.4 Tbps figure is really about. In a conventional switch, optical modules plug into the faceplate, and the switch chip’s electrical lanes have to drive signals across the package, the circuit board and a connector before they reach the laser. Each of those hops costs energy and frequently requires retiming circuitry to clean the signal up. Co-packaged optics collapses that journey by placing the optical engine on the same substrate as the switch or accelerator die — in Marvell’s framing, shortening data pathways and cutting energy per bit.

    The lane rates are what force the issue. By demonstrating 200G and 400G per lane, Marvell is aiming past the 800G optics deployed across hyperscale fleets today and the 1.6T generation now ramping, at the 3.2T generation beyond it. As per-lane signalling rates climb, the distance a passive copper cable can carry a clean signal falls, which is why the release notes copper runs out of reach at the speeds these clusters require. Once copper is out, the interconnect is optical by default — and its power draw becomes a line item in the rack budget rather than a rounding error.

    Who this lands on is specific. A facility commits a fixed number of megawatts to a hall; a rack design commits a fixed number of kilowatts to a cabinet. If the network layer connecting tens of thousands of accelerators across multiple racks consumes a smaller share of that envelope, the operator either fits more compute per rack or fewer racks per hall. That arithmetic is the reason a demo with no customer attached still moves the roadmap conversation for switch vendors, optical module suppliers, rack integrators and the people underwriting the power contracts underneath them.

    A 79% Data Center Mix Makes the Optical Roadmap the Whole Company

    Marvell’s reported quarter explains why an engineering milestone carries this much weight. Revenue of $2.739 billion, up 37% year over year, came with data center revenue of $2.17 billion, up 46% and equal to 79% of the total. Non-GAAP earnings were $0.94 per share, and the company guided the following quarter to $3.150 billion plus or minus 5% with earnings of $1.10 plus or minus $0.05, while raising its fiscal 2027 and fiscal 2028 outlooks in the August 27 filing.

    A business that derives roughly four dollars in five from data centers has effectively converted itself into a single-cycle instrument. That is an advantage while hyperscale capital expenditure expands and a concentrated exposure when it does not. It also means the optical and custom silicon roadmaps are not adjacent product lines — they are the earnings model, which is why a 2nm demonstration is treated as corroborating evidence for guidance rather than as a lab result.

    The market has priced that trajectory rather than waiting on it. The stock has advanced 192% year to date, carries a market capitalisation near $217.4 billion and a forward price-to-earnings multiple around 58, against an average analyst target of $289 with 39 buy or strong-buy ratings and five holds. A multiple set that far ahead of current earnings is a statement about execution risk: the roadmap has to arrive roughly on the schedule management has sketched, and the October 6 Investor Day is where that schedule gets specified.

    What Being First Actually Buys — and What It Costs

    The optical push runs alongside a custom silicon business that has been given an unusual structure. Management disclosed an expanded partnership with Google that includes a warrant allowing Google to acquire up to 7% of Marvell’s shares, tied to revenue milestones, and Murphy said custom revenue should more than double in fiscal 2028. Linking a customer’s equity upside to the volume it buys aligns incentives efficiently, but it also converts a commercial relationship into potential dilution for existing holders, on terms whose triggers Marvell has not published. A separate expanded NVLink Fusion arrangement with NVIDIA positions Marvell in scale-up switching, and the earlier Celestial AI photonic fabric acquisition, closed February 2, 2026, and the XConn Technologies chiplet connectivity deal, closed February 10, 2026, supply the building blocks.

    Against Broadcom, the incumbent in high-speed networking silicon, first-to-2nm is a claim on the next design cycle rather than on current sockets. Hyperscalers qualify interconnect years ahead of deployment, so the practical value of a demonstration is optionality in roadmap reviews happening now for clusters that power up later. It converts into revenue only if the process yields at volume, the thermal and reliability behaviour survives qualification, and a named platform adopts it.

    The demand side of that equation is funding itself in size, if not always for named projects: Amazon’s Form 8-K filed September 14 records the close of £4.242 billion of notes maturing between 2029 and 2045, with net proceeds of roughly £4.235 billion and no specific use of proceeds identified in the filing. Capital of that scale is available to the companies buying interconnect; whether it reaches co-packaged optics specifically is the question Marvell has not yet answered with a customer name.

    Background

    Marvell Technology designs custom silicon and networking chips for hyperscale data centers. Its portfolio spans custom AI accelerators built to a single customer’s specification, electro-optics, the digital signal processors that drive high-speed optical links, and data center interconnect products that carry traffic between facilities. Over the past two years the company has reweighted itself decisively toward that market, with the data center segment now accounting for 79% of revenue.

    Interconnect has become one of the more consequential bottlenecks in AI infrastructure. Training and inference clusters link tens of thousands of accelerators, and the network that connects them has to scale in step with the compute — first inside a rack, then across racks, then between buildings. Marvell has assembled capability for that through acquisition as well as internal development, closing the Celestial AI photonic fabric deal on February 2, 2026 and the XConn Technologies chiplet connectivity deal on February 10, 2026, while expanding partnerships with NVIDIA on scale-up switching and with Google on custom silicon. Its principal competitor in high-speed networking silicon is Broadcom.

    Sources

    Source: Marvell Smashes the AI ‘Power Wall’ with Industry-First 2nm Optical Chips — 24/7 Wall St. report on Marvell’s 2nm optical interconnect demonstrations, its fiscal second-quarter results and its custom silicon partnerships.

    Primary sources: Amazon.com, Inc., Form 8-K filed September 14, 2026 (close of £4.242 billion of notes due 2029 through 2045; net proceeds of approximately £4.235 billion).

  • HIVE’s $350M AI Cloud Deal Is Twice Its GPU Run-Rate — Delivery Is the New Risk

    HIVE’s $350M AI Cloud Deal Is Twice Its GPU Run-Rate — Delivery Is the New Risk

    TL;DR · 30-second read

    The Short Version

    A company best known for mining Bitcoin has signed a deal worth 350 million US dollars over five years to rent out computing power for artificial intelligence.

    The company, HIVE Digital Technologies, runs warehouses full of computers, largely on renewable electricity. It is steadily pointing those machines at artificial intelligence work instead of Bitcoin.

    Why it matters: the deal promises roughly 70 million dollars a year, about double what HIVE’s artificial intelligence business brings in today. The hard part now is getting the machines installed and running on schedule.

    scanx.trade reported that BUZZ High Performance Computing Inc. (“BUZZ HPC”), the wholly owned subsidiary of HIVE Digital Technologies Ltd. (TSX: HIVE; Nasdaq: HIVE), has signed a $350 million, five-year AI cloud services agreement — a contract that would supply GPU compute capacity, the specialised processors used to train and run artificial-intelligence models, on a multi-year committed basis rather than by the hour.

    The deal arrives in an unusually dense month of disclosure for HIVE. In an 8-K filed September 16, the company told investors that on August 21, 2026 it exceeded $1 million in combined average daily revenue from Bitcoin hashrate and GPU cloud services, with the GPU cloud business contributing roughly $100,000 of that per day. A second 8-K, filed September 17, disclosed a partnership between BUZZ HPC and Vancouver-based applied-AI firm ProCogia announced on September 16.

    Executive Summary

    A $350 million contract with a five-year term is a different financial object from a GPU rental business. Spot and short-term GPU capacity is priced like a commodity and repriced constantly; a dated multi-year agreement converts that capacity into something closer to a lease — a revenue line with a start date, an end date and a counterparty obligation behind it. For HIVE, which spent years as a Bitcoin miner monetising renewable power through a volatile, unhedged commodity, that distinction is the whole point of the BUZZ HPC strategy.

    The scale is worth measuring carefully in both directions. Against HIVE’s disclosed GPU cloud run-rate of about $100,000 per day — roughly $36.5 million annualised, per the company’s September 16 filing — the contract’s straight-line average of $70 million a year is roughly double the entire AI segment as it stands. Against the whole company, which HIVE says has been running above $1 million in combined daily revenue, the same contract is smaller than a single year of current total revenue. It is transformative for one engine and incremental for the business overall.

    What changes with a signed multi-year contract is the location of the risk. Demand risk falls; delivery risk rises. HIVE’s own forward-looking language in the September 17 filing names the exposure precisely: the ability to convert “signed total contract value into recognized revenue on the timing anticipated,” alongside “delays in equipment delivery or commissioning” and “capital availability.” Those are procurement, construction and financing problems, not sales problems.

    The Arithmetic: What $350 Million Means in a $100,000-a-Day Business

    Start with the two numbers side by side. HIVE disclosed in an 8-K filed September 16 that since August 21, 2026 its GPU cloud operations generated approximately $100,000 in average daily revenue. Annualised on a straight line, that is roughly $36.5 million. A $350 million agreement spread evenly across five years implies about $70 million a year. The contract, on its face, is worth close to twice the current annual run-rate of the business it is meant to feed.

    That ratio is the operationally interesting part, because a contract that large relative to an existing book cannot be served out of spare capacity. It has to be built. Real contracts also rarely run straight-line: multi-year compute agreements typically ramp as racks energise, which means the revenue curve is back-weighted and the capital spend is front-weighted. Between signature and recognition sit accelerator delivery slots, power energisation, cooling and network commissioning, and the working capital to pay for hardware before the customer pays for capacity. HIVE says as much in its own risk disclosure, listing equipment delivery delays, commissioning delays and capital availability as the factors that could break the timing.

    Who is affected downstream is concrete. Accelerator suppliers and system integrators gain a scheduled order rather than a speculative one. Electrical contractors, cooling vendors and interconnection counterparties — the utilities and operators who connect a site to the grid — get a date to work against. And lenders get what they have been asking for from this asset class: a named term and an amount, which is what makes GPU fleets financeable at something other than venture pricing. The flip side is that a dated contract removes the operator’s option to slip. A miner can idle machines when economics turn; a contracted AI host that misses a commissioning window is in breach, not merely idle.

    Dual Engines, Different Risk: 12 Bitcoin a Day Versus Contracted Compute

    HIVE’s September 10 release, furnished to the SEC on September 16, laid out the current mix plainly: approximately 12 Bitcoin per day, which the company put at roughly 2% of global Bitcoin network production, plus about $100,000 a day from GPU cloud, together exceeding $1 million in combined average daily revenue. On those figures the AI business is still well under a tenth of the top line. The $350 million contract does not reverse that overnight — but it changes the character of the revenue it adds.

    Mining revenue is spot revenue. It moves with Bitcoin’s price, with network difficulty, and with the scheduled supply reductions built into the protocol, and no amount of operational skill hedges those. Contracted compute revenue is dated and, depending on structure, may be substantially insensitive to what the spot GPU market does over the same five years. That is the trade the entire miner-to-AI cohort is attempting: exchange high-velocity, high-variance cash generation for lower-variance, longer-duration cash flows that a lender will underwrite. HIVE’s own framing in the September 10 release — Bitcoin as a “high-velocity cash-generating foundation” funding a “recurring, contracted AI infrastructure business” — is an accurate description of the mechanism, and it is also an admission that the second engine is not yet self-financing.

    The portfolio reallocation is already visible. HIVE said its Swedish Bitcoin mining operations represented less than 5% of global daily revenue during August and that it intends to wind down mining there while evaluating repurposing that infrastructure for HPC and AI. The candour of that number matters: it tells you the wind-down is a low-cost decision, not a painful one. It also warrants a caution the release does not dwell on. Repurposing a mining hall for AI is not a software change. Mining sites are built for cheap power, high temperature tolerance and thin redundancy; accelerated compute for enterprise customers needs denser power per rack, liquid or high-capacity cooling, carrier-grade networking and the uptime engineering implied by the Tier-III designation HIVE cites for BUZZ HPC. Some megawatts convert cheaply. Others are effectively a new build on an old plot.

    Sovereign Compute Is the Sales Motion, Not Just the Slogan

    The ProCogia partnership disclosed on September 17 shows how BUZZ HPC intends to fill capacity beyond a single anchor contract. ProCogia will procure dedicated GPU capacity in BUZZ HPC’s Canadian data centres to run its ZeroBoxx AI framework and vertical products — CallYeah, a voice-AI tool for healthcare providers, and PolyKode, a code-migration model for regulated environments — and will in turn act as a preferred applied-AI services partner to BUZZ HPC’s own customers, with the two intending to co-sell into the United States and Europe.

    Read commercially, that is a channel strategy aimed at a specific buyer problem. “Sovereign” compute means infrastructure where the data and the processing stay inside a defined legal jurisdiction, which matters to healthcare, financial services and public-sector buyers governed by data-residency rules. Those buyers are also the ones least able to assemble an AI stack themselves, so bundling infrastructure with an applied-AI delivery partner addresses the real obstacle — nobody to build the thing — rather than just the rack. The hire of Mark Volk as BUZZ HPC’s Senior Vice President, Revenue, announced in the same September 10 release, fits the same thesis: an enterprise and government sales motion is a different discipline from selling hashrate.

    The honest caveat is that the partnership discloses no dollar value, no committed capacity and no term, which places it in a different category from the $350 million agreement. Executive Chairman Frank Holmes referenced “a 400 MW coast-to-coast Canadian pipeline” — 400 megawatts being, very roughly, the electricity draw of a mid-sized city — but a pipeline is not energised capacity, and the release does not break it into secured power, sites under construction and sites under evaluation. Competition is real, too: the major cloud providers have been standing up in-country regions with sovereignty commitments of their own. BUZZ HPC’s argument is price and availability for Canadian companies that cannot get either, which is a defensible niche — and one that a five-year contract, if it converts on schedule, would do more to prove than any amount of positioning.

    Background

    HIVE Digital Technologies Ltd. (TSX: HIVE; Nasdaq: HIVE; BVC: HIVECO), headquartered in San Antonio, Texas, began as a Bitcoin miner built around low-cost renewable power and data centre operations. Its stated model is to convert electricity into compute and then sell that compute — historically as hashrate for the Bitcoin network, and increasingly as GPU capacity for artificial-intelligence workloads through its wholly owned subsidiary BUZZ High Performance Computing Inc.

    That pivot is an industry-wide pattern. Miners already control three scarce inputs for AI infrastructure — secured power, energised sites and operational experience running large fleets of hot, dense hardware — while facing a commodity revenue stream that is volatile and periodically halved by protocol design. Converting megawatts to AI hosting trades that volatility for contracted, longer-duration cash flows, but requires heavier capital per megawatt and a customer base that demands uptime guarantees, data-residency assurances and enterprise service levels that mining never did.

    Sources

    Source: HIVE BUZZ HPC signs $350 million five-year AI cloud services agreement — report that HIVE Digital Technologies’ BUZZ HPC subsidiary entered a $350 million, five-year AI cloud services agreement.

    Primary sources: HIVE Digital Technologies Ltd., Form 8-K filed September 17, 2026; Exhibit 99.1 — “HIVE’s BUZZ HPC Partners with ProCogia to Deliver End-to-End Sovereign AI to Enterprise and SMB Clients Across Canada, the U.S., and Europe”; HIVE Digital Technologies Ltd., Form 8-K filed September 16, 2026; Exhibit 99.1 — “HIVE’s BUZZ HPC Appoints AI and HPC Veteran Mark Volk as Senior Vice President, Revenue; HIVE Surpasses $1 Million in Average Daily Revenue”.

  • IREN’s $2.8B Order Book Shows AI Capacity Now Gets Built on Contracts, Not Bets

    IREN’s $2.8B Order Book Shows AI Capacity Now Gets Built on Contracts, Not Bets

    TL;DR · 30-second read

    The Short Version

    IREN used to mine bitcoin. Now it rents out computing power for artificial intelligence — and Microsoft has signed on as a customer.

    The company has $2.8 billion in new orders. That is money customers have agreed to pay over time, not cash sitting in the bank today.

    Why an ordinary person should care: the buildings that run artificial intelligence cost enormous sums to build before they earn anything. A signed commitment from a buyer as large as Microsoft makes it far cheaper and easier to borrow the money to build them.

    Moomoo reported that IREN Ltd has secured a contract with Microsoft and booked $2.8 billion in new orders, a combination the company presents as accelerating its transformation into an artificial-intelligence infrastructure provider.

    IREN is a Nasdaq-listed operator that began as a bitcoin miner and has spent recent years redirecting its data center estate — land, substations and grid connections built for mining — toward hosting AI computing workloads. A named hyperscale customer and a multi-billion-dollar order figure are the two most material markers of that shift to date.

    Executive Summary

    Two facts carry this story: a dollar figure and a name. The $2.8 billion order book is the number that will travel; the Microsoft contract is the part that changes how the business is valued and financed.

    The distinction matters because AI data center capacity gets built in one of two ways. It is built on spec, in the hope that rental rates for AI chips hold up long enough to repay the cost — or it is built against a signed, multi-year commitment from a buyer. The two look identical from the outside and cost radically different amounts of money to finance.

    For the broader market, IREN is a test case for a pattern now visible across the sector: large cloud providers contracting directly with converted crypto-mining operators rather than waiting on their own construction pipelines. That arrangement transfers the demand risk off the operator’s balance sheet — and transfers a different set of risks, concentration chief among them, onto it.

    The Counterparty Is the Number That Matters

    The headline figure is $2.8 billion in new orders. The more consequential disclosure is the name attached to it. In AI infrastructure, capacity is either merchant — built speculatively and rented at whatever the spot market pays — or contracted, built against a signed offtake agreement. The physical racks are the same. The cost of capital behind them is not.

    Here is the mechanism. A merchant fleet of AI accelerators is a commodity asset: its revenue depends on rental prices that move quickly, on hardware that depreciates fast, and on an upgrade cycle measured in quarters rather than decades. Lenders discount that cash flow heavily, so the build gets funded mostly with equity, which is the most expensive money there is. A multi-year contract with a large, creditworthy technology buyer converts the same racks into a stream of receivables, and lenders then price the risk substantially against the buyer’s credit rather than the operator’s. Nothing about the building changes; the interest rate does.

    That is what makes a named hyperscaler contract structurally different from a strong quarter of rental demand, and it is why the signature is worth more analytically than the order total. The effect radiates outward. Operators with an anchor tenant can commit to chip allocations and power contracts earlier and more cheaply. Operators without one bid for the same accelerators and the same megawatts on worse terms. Chip vendors gain firmer order visibility; utilities and power developers gain load commitments they can actually plan generation and transmission around. The party squeezed is the marginal builder with capacity but no counterparty — the one whose economics depend on renting to whoever shows up.

    What Transfers From Mining, and What Doesn’t

    The reason converted miners are credible counterparties at all is that they already hold the genuinely scarce inputs. Bitcoin mining required cheap, abundant electricity at industrial scale, which meant acquiring land, building substations, negotiating power purchase agreements and — critically — securing a place in the interconnection queue, the multi-year waiting list to connect a large load to the electricity grid. Those assets cannot be bought quickly at any price, and they are precisely what an AI operator needs.

    What does not transfer is the interior. Mining rigs tolerate heat, dust, single-feed power and interrupted service; a mining site that goes dark for an afternoon simply loses a day’s revenue. AI training runs do not tolerate any of that. They require redundant power paths, far higher rack power densities and usually liquid cooling, high-bandwidth low-latency networking between thousands of chips acting as one machine, and contractual uptime guarantees with financial penalties attached. Retrofitting a mining shed into that is substantial capital expenditure the mining business never had to carry.

    The practical read, then, is that a converted site is a head start rather than a finished product — and that a hyperscaler signing a contract is meaningful third-party evidence that specific sites cleared that engineering bar. That is a stronger signal than any self-described transformation, because the buyer had every incentive to inspect before committing.

    An Order Book Is a Promise, Not a Profit

    An order book is contracted future revenue, and its meaning depends almost entirely on the term over which it is delivered. The same $2.8 billion is a very different business spread across three years than across ten, and the annualised figure implied by each is not something a reader can derive from the total alone. Orders also convert to revenue only as capacity is delivered, which means the recognition schedule follows a construction timeline, not a sales announcement.

    The corresponding obligation is capital. Serving AI contracts means buying accelerators, energising power and installing cooling ahead of the revenue those contracts will eventually produce. In this business, a growing order book and rising cash consumption arrive together; the relevant question is the size and funding of the gap between them, not the size of the backlog.

    Stated even-handedly: what is substantiated here is a named customer, a headline order figure and a clear strategic direction — all three are material and none are trivial. What is not yet substantiated is the margin profile, the contract term, the contracted capacity or the delivery schedule, and those four variables determine whether $2.8 billion is a transformative book or a demanding one.

    Concentration Is the Price of Credibility

    The anchor tenant that makes capacity financeable also concentrates the business. A contract large enough to move a company’s valuation is, by definition, a customer whose departure at renewal would move it just as far in the other direction — and large buyers know that when renewal terms are negotiated. Cheap capital today and pricing leverage against you later are two faces of the same agreement.

    There is a second-order consideration worth stating plainly rather than dramatically. Hyperscalers lease third-party capacity alongside building their own, and the leased portion is a flexible variable in their planning: it can expand when their construction pipeline lags demand and contract when it catches up. That is a legitimate procurement strategy, not a criticism of any buyer. But it means operators in this position are underwriting capacity against a demand source that has an alternative and they do not. The mitigation is diversification of the order book, which is exactly why the split between the Microsoft contract and the remainder of the $2.8 billion is the most useful single disclosure still outstanding.

    Background

    IREN Ltd was founded in Australia as a bitcoin miner and listed on Nasdaq, building its business around securing large volumes of low-cost, largely renewable electricity and constructing its own data center sites to consume it. That model required the company to become, in practice, a power and land developer — acquiring sites, building substations and working through grid interconnection processes — with mining as the initial use of the capacity.

    Since 2023, a broad reordering has pulled that asset base toward artificial intelligence. Demand for AI computing outran the supply of energised, grid-connected data center space faster than new construction could respond, while bitcoin mining economics tightened. Operators holding power capacity found that the same sites could host far higher-value workloads, and a number of listed miners repositioned as AI hosting providers. The competitive question across that group has been which of them can pair their power advantage with the engineering and the customer relationships that AI workloads actually require — which is why a contract with a named hyperscale buyer reads as a milestone rather than a routine sale.

    Sources

    Source: IREN Secures Microsoft Contract; $2.8 Billion in New Orders Accelerates AI Transformation — report that IREN has signed a contract with Microsoft alongside $2.8 billion in new orders as it converts data center capacity to AI hosting.

  • AOI’s 1.8M Sq Ft in Texas and 10 Years in Ningbo Make AI Optics a Floor-Space Bet

    AOI’s 1.8M Sq Ft in Texas and 10 Years in Ningbo Make AI Optics a Floor-Space Bet

    TL;DR · 30-second read

    The Short Version

    Applied Optoelectronics makes the laser-based parts that let computers inside a data center talk to each other at high speed. Artificial intelligence has made those parts scarce.

    The company has now rented an entire business park outside Houston: 1.8 million square feet, or roughly 31 American football fields of floor. Days earlier it signed a ten-year lease on a factory in Ningbo, China, for less than a million dollars a year in rent.

    It has not said who will buy what comes out of either building.

    Applied Optoelectronics, the Sugar Land, Texas optical components maker listed on Nasdaq as AAOI, has taken the entire 1.8 million-square-foot Hightower Business Park, a Crow Holdings industrial development in the Houston area, for data center-related manufacturing. The Houston Business Journal reported on September 21 that the campus began life as a speculative project — built without a tenant in hand — and was absorbed in full by the single Sugar Land company, with the 737,621-square-foot Building 5 shown in the developer’s rendering as AOI space.

    The Texas commitment lands alongside a second, separately documented expansion. In an 8-K filed September 15, AOI disclosed that its wholly owned subsidiary Global Technology, Inc. signed a Factory Premises Lease Agreement on September 10 for a roughly 38,311.8-square-meter building — about 412,000 square feet — at No. 227 Kesheng Rd. in Ningbo, China. That lease runs ten years, from September 16, 2026 to September 15, 2036, at an annual rent of RMB 6,896,124.

    Executive Summary

    Two leases in eleven days put more than 2.2 million square feet of manufacturing and support space under Applied Optoelectronics’ name on two continents. Neither is a fab expansion in the semiconductor sense — no wafer line, no lithography tool order. Both are real-estate commitments, and that is what makes them interesting: the company is buying room and time before it has publicly named a single customer for the output.

    The Ningbo lease, because it sits in an SEC filing, comes with terms the Texas deal does not. AOI’s subsidiary negotiated a three-month rent-free renovation period, the right to alter the building as production requires, protection against the landlord terminating early absent force majeure or material breach, a right of first refusal to buy the property if the landlord sells, a right to bind any future owner to the lease, and a right of first refusal on renewal. Those are the terms a tenant asks for when it intends to bolt heavy, immovable equipment to the floor and stay.

    Read together, the two deals describe a supply-side posture: add assembly and test capacity in the United States and in China simultaneously, on long horizons, and absorb the fixed cost now in exchange for the ability to ship when AI network buildouts call for volume. The upside is lead-time relief for customers. The risk is that rent, fit-out and headcount all start before the orders do.

    Why Optics Capacity Is Measured in Square Feet

    An optical transceiver — the plug-in module that converts a switch’s electrical signal into light, pushes it down a fiber, and converts it back at the far end — is not made the way a processor is made. The laser chip inside it comes off a wafer, but everything after that is alignment, packaging, burn-in and test: benches, automated assembly cells, environmental chambers, racks of test gear running modules for hours before they ship. That work scales with floor area and with the people standing on it, not with a single expensive tool. Which is why a company that wants to ship a lot more optics goes looking for a building rather than a cleanroom bay.

    The numbers here are the argument. AOI did not take a suite in the Hightower park; it took all 1.8 million square feet, of which Building 5 alone is 737,621 square feet. Add the roughly 412,000 square feet in Ningbo and the company has committed to more than 2.2 million square feet in under two weeks. For anyone specifying an AI cluster, that reframes the transceiver lead-time question. The gating item is not only chip supply; it is whether the vendor has floor, fit-out, qualified lines and trained operators ready when the order lands. Space leased in 2026 is what determines whether modules ship in 2027 and 2028.

    The Houston-area deal carries a second signal worth naming. The park was developed speculatively — Crow Holdings built it without a committed occupant, betting general industrial demand would fill it. Full absorption by one manufacturer moves that risk from the landlord’s balance sheet to the tenant’s. Developers of large spec industrial product now have a live data point that AI-adjacent manufacturing can take a whole campus at once; AOI, in exchange, has a fixed obligation sized to demand it has not yet publicly contracted.

    Two Leases, Two Jurisdictions, One Product Line

    The Ningbo terms repay close reading. Annual rent of RMB 6,896,124 across 38,311.8 square meters prices out to exactly RMB 180 per square meter per year — roughly RMB 16.70 per square foot, and less than a million US dollars a year in total at recent exchange rates. Escalation is modest and slow: three percent every three years, beginning in the third lease year. Rent, in other words, is a rounding error next to what will go inside the building. The meaningful capital is the tooling, the test capacity and the payroll, and none of that is disclosed.

    The tenant protections say more than the price does. A right of first refusal to purchase, a clause binding any future buyer of the property, a renewal right on terms no worse than a competing tenant’s, and a landlord who cannot terminate early: that package is negotiated by an occupier planning a decade of production in one place, not by one hedging its options. Running that alongside a simultaneous Texas campus gives AOI capacity on both sides of the US–China trade boundary — useful if customers have domestic-content preferences or tariff exposure, and useful if cost pressure pushes volume the other way. It also means policy changes in either direction touch one footprint but not both.

    The Cost Structure Moves First

    Leases convert a variable problem into a fixed one. The Ningbo clock is explicit: the term begins September 16, 2026, with three rent-free months from actual delivery for renovation — a short runway for fitting out a 412,000-square-foot factory. Rent obligations start on the calendar; revenue from the space starts when lines are built, qualified and buying customers have signed. The gap between those two dates is where the operating leverage lives, in both directions.

    For buyers of AI networking hardware, that leverage is broadly favorable. Capacity added ahead of demand is how lead times compress and how second sources become real rather than theoretical. For investors, the same commitment reads as a schedule to monitor: fit-out spending, headcount, and the point at which each site’s output is qualified into customers’ builds. AOI has disclosed the space and, for Ningbo, the rent. It has not disclosed the contracted demand those buildings are meant to serve, and that asymmetry — space certain, orders unstated — is the honest summary of where this stands today.

    Background

    Applied Optoelectronics designs and manufactures optical components and the transceiver modules built around them — the laser-based parts that carry traffic over fiber inside and between data centers, and in broadband access networks. It is headquartered in Sugar Land, Texas, just southwest of Houston, and trades on Nasdaq as AAOI. Manufacturing has long been split across sites in the United States and Asia, with Global Technology, Inc. operating as its Ningbo, China subsidiary.

    Demand for high-speed optics has been reshaped by AI cluster construction, where tens of thousands of accelerators must be wired together at very high bandwidth, consuming transceivers in volumes that conventional cloud workloads did not. That has pushed component suppliers to add assembly and test capacity, which is space- and labor-intensive work — and made industrial real estate, not just chip supply, part of how quickly the optics side of an AI buildout can scale.

    Sources

    Source: AOI leases entire 1.8M-SF industrial park for data center-related manufacturing — Houston Business Journal, September 21, 2026, on Applied Optoelectronics taking all of Crow Holdings’ Hightower Business Park.

    Primary sources: Applied Optoelectronics, Inc., Form 8-K filed September 15, 2026 (SEC EDGAR, CIK 0001158114) — disclosing the ten-year Ningbo factory lease signed September 10, 2026, its rent, escalation schedule and tenant rights.