Category: AI Infrastructure

  • Marvell’s 32TB Memory Tier Bets AI Inference Gains Hinge on Where KV Cache Lives

    Marvell’s 32TB Memory Tier Bets AI Inference Gains Hinge on Where KV Cache Lives

    TL;DR · 30-second read

    The Short Version

    Chipmaker Marvell unveiled new parts aimed at a quiet problem in artificial intelligence: the expensive chips that run chatbots often sit idle, waiting for data.

    When an assistant works through a long conversation or a multi-step task, it keeps a running memory of everything it has already read. That memory outgrows the chip. Marvell’s answer is a shared memory pool, reached by light over cables up to 50 meters long, that holds about as much as 32 laptops.

    Marvell says this lets the same building produce two to three times more answers on the same electricity. That figure has not yet been tested by anyone else.

    Marvell Technology announced on August 4, 2026, at the FMS 2026 memory and storage conference in Santa Clara, California, a set of products aimed at AI inference, the stage where trained models answer requests. The lineup spans three levels of the data center: the Bravera SC6 PCIe 6.0 SSD controller for individual servers, Structera X CXL memory expansion for racks, and Photonic Fabric memory modules, network cards and chiplets that form an optical shared-memory tier across multiple racks.

    Marvell says Photonic Fabric reaches up to 50 meters, supports up to 32TB of warm KV cache offload and delivers up to 2-3x higher token throughput within existing footprints and power budgets. It says Bravera SC6 doubles the performance of its Bravera SC5 predecessor and is expected to begin sampling in Q4 2026. Marvell gave no availability dates for the Structera X or Photonic Fabric products.

    Executive Summary

    Marvell is presenting memory, not processors, as the next place where AI inference performance is won or lost. All three products in the announcement target the same object: the KV cache. This is the working memory a language model builds as it reads a prompt, and it grows with longer conversations and with more simultaneous users. The company argues that expanding, pooling and sharing memory separately from GPUs reduces the time those GPUs spend idle, which it calls stalls, and raises output per watt.

    This matters to anyone who pays for AI infrastructure by the megawatt. If the claims hold, operators could get more usable work from GPU fleets they have already bought and powered, without adding floor space or grid capacity. The strongest number, the 2-3x throughput gain, is framed as an “up to” figure with no stated baseline or workload. Only one of the products, the SC6 controller, has a date, and that date is for sampling rather than volume shipment.

    Three Tiers Aimed at One Problem

    The announcement is best read as a memory hierarchy rather than three separate products. At the top sits high-bandwidth memory (HBM), the fast but scarce memory stacked beside each GPU. Below it, Structera X uses CXL (Compute Express Link, an open standard for attaching extra memory to processors over the same physical lanes as PCIe) to add and pool DRAM at the rack level. Photonic Fabric then extends a shared memory tier across racks over optical links of up to 50 meters. At the bottom, the Bravera SC6 controller makes SSDs fast enough that, in Marvell’s words, more KV cache can “move from high-bandwidth memory to SSD.”

    Every tier has the same stated purpose: holding the KV cache somewhere cheaper and larger than HBM, while keeping it close enough that the GPU does not wait. Marvell’s framing, that the portfolio helps hyperscalers “scale memory more independently from compute,” is a statement about data center design. It describes memory as a resource provisioned at the rack and pod level, not a fixed amount soldered into each server.

    Why KV Cache Placement Drives Throughput

    The case for the headline rests on how inference works. When a model processes a prompt, it computes intermediate values for every token and stores them in the KV cache so it does not have to recompute them for each new word it generates. That cache grows with context length and with the number of concurrent sessions. Agentic AI, where a model runs multi-step tasks and carries long histories between steps, pushes both factors up at the same time. Once HBM fills, the operator has three options. It can evict the cache and recompute it later, which spends GPU cycles on repeated work. It can fetch the cache from storage, which is slow. Or it can fetch it from a faster intermediate tier.

    Marvell’s throughput claim sits on that third option. The company says Photonic Fabric enables KV cache “to be loaded from a shared memory tier rather than storage,” offering up to 32TB of warm cache (recently used but not currently active) across multiple XPUs (Marvell’s umbrella term for GPUs and custom accelerators). From that it claims up to 2-3x higher token throughput within the same footprint and power envelope. The GPUs are unchanged. What changes is how often they sit idle waiting for context to be rebuilt or retrieved. Operators most affected are those running long-context, many-user inference on power-capped sites, where adding GPUs is not an option but raising their utilization is.

    This supports a narrower claim than “memory has replaced GPUs as the bottleneck,” and the release itself does not go that far. It says memory is becoming “as critical as compute.” 650 Group analyst Alan Weckel, quoted in the release, calls memory capacity, bandwidth, latency and data movement “primary constraints.” The defensible reading is that for inference workloads with large KV caches, gains now come from memory placement as much as from raw compute.

    What the Numbers Do and Do Not Establish

    The concrete figures are the 32TB capacity, the 50-meter reach, the 2x performance step from SC5 to SC6, and the Q4 2026 sampling date for SC6. The 50-meter reach is significant for facility planners. It means the shared tier spans multiple racks, enough to cover a GPU pod, and it adds optical cabling inside the pod.

    The 2-3x figure is the least specified. It carries an “up to” qualifier and no stated model size, context length, batch size or comparison system. Throughput gains from cache offload depend heavily on how often the cache is reused, so the figure is best read as an upper bound for favorable workloads, not a fleet-wide expectation. Likewise, “doubling the performance” of SC5 fits the generational jump from PCIe 5.0 to 6.0 but does not specify which metric doubles. “Extremely low latency” for the optical tier is not quantified at all.

    Sourcing Flexibility and the Buyer’s Calculus

    One of the more practical details concerns procurement. Bravera SC6 is NAND-agnostic, meaning it works with flash memory from multiple suppliers. For cloud providers, that lowers dependence on a single flash vendor, a real consideration in a market prone to supply swings. Structera X is pitched in similar terms, as a way to extend “the value of existing memory investments” rather than forcing wholesale replacement.

    The trade-off is architectural complexity. Each tier adds devices, interconnects and software orchestration that must decide, in real time, where each session’s cache should live. The added power draw of CXL expanders, optical network cards and memory modules also counts against the power envelope the throughput gains are meant to protect. Buyers will need measured results on their own workloads before the net efficiency claim can be verified.

    Background

    Marvell Technology (NASDAQ: MRVL) is a Santa Clara-based semiconductor company focused on data infrastructure: chips that move, store, process and secure data for cloud, enterprise and carrier networks. It says it has served leading technology companies for over 30 years. Its Bravera line covers storage controllers, its Structera line covers CXL memory, and Photonic Fabric is its optical interconnect and memory technology.

    As AI shifts from training models to serving them at scale, the memory attached to each GPU has become a practical limit. Models that handle long documents, long conversations or multi-step agent tasks build large KV caches that can exceed on-package high-bandwidth memory. Chipmakers and cloud providers are responding with tiered designs that spill that cache into pooled DRAM, shared memory and fast flash storage. FMS, the annual memory and storage conference, is a common venue for these announcements.

    Sources

    Source: Marvell Advances AI Memory Infrastructure Portfolio to Accelerate Agentic AI Inference: Marvell Technology’s August 4, 2026 announcement of Bravera SC6, Structera X and Photonic Fabric memory products at FMS 2026.

  • NVIDIA’s Reported $50B Lease Is Really a $19.6B Bet on Powered AI Capacity

    NVIDIA’s Reported $50B Lease Is Really a $19.6B Bet on Powered AI Capacity

    TL;DR · 30-second read

    The Short Version

    NVIDIA, the company whose chips run most artificial intelligence, has reportedly agreed to rent a huge computer campus in Texas from a firm called Hut 8. NVIDIA has not confirmed it.

    The deal is described as worth $50 billion, but only $19.6 billion is locked in for the first 15 years. The rest depends on renewals after that.

    What is really being rented is electricity: a guaranteed grid connection roughly equal to what one large nuclear reactor produces. For artificial intelligence, that has become as important to secure as the chips themselves.

    Data Center Frontier reported on August 3, 2026 that NVIDIA has emerged as the tenant behind Hut 8’s lease commitments at its one-gigawatt Beacon Point AI data center campus in Nueces County, Texas. Hut 8 has disclosed two 15-year leases covering 704 megawatts of IT capacity with a combined base-term value of $19.6 billion. That figure rises to as much as $50.2 billion if all renewal options are exercised. The Financial Times identified NVIDIA as the tenant. Reuters said it could not independently verify that, and NVIDIA neither confirmed nor denied its role, saying only that it works with ecosystem partners to deploy AI infrastructure through its DSX AI factory architecture.

    In a separate announcement, Aalo Atomics and Crusoe said they will partner on a nuclear-powered AI data center. They plan to demonstrate it at Idaho National Laboratory in 2027 and then deploy 50-megawatt Aalo Pod reactors at Crusoe data centers by the end of 2029.

    Executive Summary

    Hut 8 has fully commercialized Beacon Point with two 352-megawatt phases, each carrying a base-term contract value of about $9.8 billion. The campus runs on a one-gigawatt interconnection agreement with AEP Texas and is being designed around NVIDIA’s DSX reference architecture. If the reported tenant is confirmed, NVIDIA would move beyond supplying GPUs and become an anchor tenant whose credit helps make a multibillion-dollar campus financeable.

    The headline number overstates what is firm. Only $19.6 billion of the $50.2 billion is committed. The remaining $30.6 billion depends on renewals beyond the initial 15 years. What the firm commitment secures is 704 megawatts of capacity that is already power-connected, and in the current AI buildout that is the scarcest input.

    The Aalo–Crusoe partnership takes the opposite approach to the same constraint. Instead of locking up a gigawatt of grid power, it aims to manufacture both a small reactor and a modular data center and to add them 50 megawatts at a time. Beacon Point itself is grid-connected, not nuclear-powered.

    Most of the $50 Billion Sits in Renewal Options

    The $50.2 billion figure is a ceiling, not a check. Hut 8 has disclosed two 15-year leases with a combined base-term contract value of $19.6 billion, about $9.8 billion for each 352-megawatt phase. The remaining $30.6 billion, roughly 61 percent of the headline, materializes only if every renewal option beyond the initial terms is exercised.

    The firm commitment is still very large. Spread evenly, $19.6 billion over 15 years comes to about $1.3 billion a year, or roughly $1.9 million per megawatt of IT capacity annually. That is a simple average and not a rent schedule, because payment start dates and escalators have not been published.

    Whether the renewals happen depends on something no one can promise today: that a campus designed in 2026 will still be competitive decades from now. Computing systems turn over rapidly, while the land, the shell and the grid connection last far longer. Power density and cooling will need repeated upgrades to keep the site useful. The base term prices the next 15 years of AI demand, and the option value prices how long the power connection itself stays valuable.

    What the Lease Actually Buys Is Powered Capacity

    Set aside the headline value, and the scarce asset at Beacon Point is electricity that has already been promised. The campus holds one gigawatt of utility capacity under an interconnection agreement with AEP Texas. An interconnection agreement is the utility contract that lets a large load plug into the grid, and obtaining one has become as important as obtaining GPUs. The two leases commit 704 megawatts of IT capacity, meaning the power delivered to servers, networking and storage. The gap between that figure and the 1 GW connection typically covers cooling and electrical overhead.

    Hut 8’s own disclosure shows which variable is fixed. The company said it redesigned the first Beacon Point data hall around NVIDIA’s architecture and increased capacity by 57 percent within the same land and utility footprint. The grid connection did not grow. What changed was how much compute the design could extract from it. NVIDIA pitches DSX on the same logic, emphasizing token performance per megawatt (AI output per unit of power) rather than floor space.

    The parties most affected are those who cannot secure this kind of power allocation on their own. Some reporting has speculated that NVIDIA could sublease capacity. If it does, smaller GPU cloud providers, often called neoclouds, would gain access to power and space they could not finance independently. For Hut 8, the leases bring its contracted AI portfolio to 949 megawatts on 1,330 megawatts of utility capacity, with $26.6 billion of aggregate base-term value. Beacon Point accounts for roughly three-quarters of that value.

    A Chip Supplier’s Credit Becomes Construction Finance

    If the Financial Times identification holds, the bigger shift is NVIDIA’s role in the market. Campuses at this scale are typically financed against the creditworthiness of the tenant, and Hut 8 has described its counterparty as investment-grade. An NVIDIA lease would make the company an anchor tenant whose balance sheet helps get a multibillion-dollar campus built. It would no longer be only the supplier of the GPUs, networking, systems and software inside.

    The strategic logic is clear. A supplier that underwrites capacity can pull forward projects that might otherwise stall on financing. It can also standardize those projects around its own reference design and widen the pool of customers able to rent its hardware. The open question, which only NVIDIA can answer, is how much of the demand behind such a campus comes from end customers and how much is the supplier’s own commitment. A lease anchored by the chip vendor and sublet to the vendor’s customers concentrates risk around a single forecast of compute demand. That is the tradeoff the model accepts in exchange for speed.

    The Other Model: Putting the Reactor Next to the Racks

    The nuclear model in this news cycle comes from Aalo Atomics and Crusoe, not from Beacon Point. The two companies plan to run a Crusoe Spark modular data center on power from an Aalo reactor at Idaho National Laboratory in 2027. They then intend to deploy Aalo Pods, which are 50-megawatt-electric nuclear plants, at Crusoe data centers by the end of 2029. Crusoe puts its Spark factory lease, buildout and initial module fleet at more than $200 million. Aalo has secured more than $136 million in funding.

    The milestone so far should be read precisely. On July 4, 2026, Aalo’s Critical Test Reactor reached criticality, meaning it sustained a nuclear chain reaction. It did so at zero power, before sodium coolant or any electricity-generating systems were added. The 10-megawatt-electric Aalo-X power reactor is being built next door. Commercial sites will still require regulatory approval, fuel, waste-management plans, insurance, financing and community support. For that reason, 2027 and 2029 are better read as ambitious targets than as delivery dates.

    Side by side, the two projects answer the same constraint in opposite ways. Beacon Point locks up a gigawatt of grid power and fills it in 352-megawatt blocks. Aalo and Crusoe aim to manufacture both the power plant and the data center and to add capacity 50 megawatts at a time. In both cases, securing power comes first in the contract.

    Background

    Hut 8 built its business in bitcoin mining and has been repositioning toward AI data center development, leasing power-connected capacity to large tenants under long-term contracts. Its Beacon Point campus in Nueces County, Texas, is now fully contracted and sits within a 949-megawatt contracted AI portfolio supported by 1,330 megawatts of utility capacity. NVIDIA supplies the GPUs, networking, systems and software that many AI data centers run on. It promotes DSX as a reference architecture for designing entire facilities around its platforms.

    Crusoe is an AI cloud and data center company that sells computing through Crusoe Cloud and builds prefabricated Spark data center modules. Aalo Atomics is developing small advanced reactors, including the 10-megawatt-electric Aalo-X, a sodium-cooled design, and the 50-megawatt-electric Aalo Pod. The company is expanding toward a one-million-square-foot manufacturing operation, and its planned demonstration with Crusoe is sited at Idaho National Laboratory.

    Sources

    Source: NVIDIA’s Reported $50B Lease and the Nuclear-Powered AI Factory: Data Center Frontier’s August 3, 2026 analysis of Hut 8’s Beacon Point leases and the Aalo–Crusoe nuclear data center partnership.

  • At $130-145bn Capex, Meta’s AI Data Centers Start to Look Like a Cloud Business

    At $130-145bn Capex, Meta’s AI Data Centers Start to Look Like a Cloud Business

    TL;DR · 30-second read

    The Short Version

    Meta, the company behind Facebook and Instagram, now expects to spend between $130 billion and $145 billion this year. The cost of the huge computer buildings that power artificial intelligence keeps climbing. This is the second time this year Meta has raised that figure.

    The bill is biting. In the latest three months, the cash Meta had left after all its spending shrank to $784 million, from $8.5 billion a year earlier. Its shares fell about 10 percent.

    The twist: other companies are offering to pay Meta more than it paid for its computing power. Meta may start renting that power out.

    Meta has raised its 2026 capital expenditure guidance to between $130bn and $145bn, DatacenterDynamics reported on August 2, 2026. The range had been $115bn–$135bn at the start of the year and $125bn–$145bn in April. The increase comes as AI data center costs rise. In the same week, Amazon lifted its own capex guidance by $20bn and cited the cost of memory for AI servers.

    Meta shares fell around 10 percent. Quarterly free cash flow dropped to $784m, its lowest level in five years, compared with $8.5bn a year earlier. Revenue rose 28 percent to $61bn and profit fell 14 percent to $6bn. Meta again confirmed its interest in becoming a cloud provider. It recently hired AWS cloud lead David Brown, and it is reportedly in early talks to rent compute to Anthropic in a deal that could be worth as much as $10bn.

    Executive Summary

    Meta now expects to spend $130bn to $145bn this year on capital projects, the long-lived assets such as AI servers, data center buildings and the electrical and cooling systems that support them. This is the second upward revision of 2026. The midpoint of the range has risen from $125bn at the start of the year to $137.5bn. The latest step lifts the bottom of the range and leaves the $145bn ceiling set in April unchanged.

    The spending is now visibly squeezing Meta’s cash generation. Free cash flow of $784m for the quarter was a fraction of the $8.5bn Meta generated a year earlier, and the stock fell around 10 percent. Amazon’s shares rose after its own capex increase. The difference shows that investors distinguish between AI spending that feeds an existing cloud business and AI spending whose return is still prospective.

    The bigger strategic signal is Mark Zuckerberg’s remark that Meta is receiving offers for its compute “at a significant premium over what we paid for it.” Three things point the same way: that remark, a senior cloud hire from AWS, and reported talks with Anthropic. Together they suggest Meta’s AI fleet may become a product it sells as well as a cost it carries. That would matter for AI developers hunting for capacity, for incumbent clouds, and for the suppliers who build and power data centers.

    The Raise Lands at the Floor, Not the Ceiling

    Meta’s new range of $130bn to $145bn is its second upward revision of 2026. The year opened at $115bn–$135bn, and April’s update moved it to $125bn–$145bn. This time the top of the range held at $145bn and the bottom rose by $5bn. Since the start of the year, the midpoint has climbed from $125bn to $137.5bn, an increase of $12.5bn or 10 percent. The latest step, however, is a narrowing. Meta is signalling that it is less likely to land at the low end. It is not signalling that it has found room to go higher.

    The increase comes as AI data center costs rise. In the same week, Amazon raised its guidance by $20bn and pointed specifically to the price of memory for AI servers. That context matters for how the number is read. Capital expenditure can rise because a company is building more. It can also rise because each unit it builds costs more. Meta has not said how its increase splits between the two. Amazon’s explanation is a reminder that at least part of the industry’s recent guidance creep reflects price rather than volume.

    Why Meta’s Data Centers Start to Look Like a Cloud Business

    The most consequential line from the earnings call was not the capex figure. It was Zuckerberg’s observation that Meta is “getting a lot of offers for compute at a significant premium over what we paid for it.” Compute here means the processing capacity of Meta’s AI server fleet. If outside buyers will pay more than Meta’s cost for that capacity, every server Meta installs has two possible uses. It can train and run Meta’s own models, or it can be rented to someone else at a margin. That shifts the capex line from pure cost toward something closer to inventory.

    Meta has taken steps consistent with acting on that. It again confirmed its interest in becoming a cloud provider, it recently hired AWS cloud lead David Brown, and it is reportedly in early talks to rent compute to Anthropic. That deal could be worth as much as $10bn. For scale, $10bn is roughly 7 percent of the low end of this year’s capex range. That is meaningful, but a single deal of that size would not by itself change the investment case, and the talks are described as early.

    The effects would reach both sides of the market. AI developers that cannot secure enough capacity from established clouds would gain another large supplier. Incumbent cloud providers would face a competitor whose capacity was financed by an advertising business rather than by cloud customers. Data center developers and equipment vendors could find Meta’s demand extending beyond what its own apps require. None of this is settled. Running a cloud business requires customer contracts, service-level commitments, support and sales functions that Meta has not historically built. The direction, though, is visible in what the company has said and done.

    Why Investors Treated Meta and Amazon Differently

    Amazon raised its guidance and its shares jumped. Meta raised its guidance and its shares fell around 10 percent. The cash figures help explain the gap. Free cash flow is the cash left after operating costs and capital spending. Meta’s was $784m for the quarter, its lowest level in five years, against $8.5bn a year earlier. Revenue grew 28 percent to $61bn, but profit fell 14 percent to $6bn, so costs are growing faster than sales.

    One reasonable reading of the split reaction is about visibility. Amazon’s spending feeds a business that already sells computing capacity to customers, so investors can see where the return comes from. Meta’s spending serves its own products, and much of the payoff depends on AI products that Zuckerberg described as “coming.” Meta’s third-quarter revenue forecast of $61bn to $64bn has a midpoint of $62.5bn, below the $63.1bn Wall Street expected, so it offered little near-term offset. Renting out capacity is one way Meta could make the return on its data centers easier to see. That is another reason the cloud ambition and the capex number belong in the same story.

    What the Spending Means Down the Supply Chain

    For suppliers of power, cooling and construction, guidance at this level signals sustained demand from one of the largest private buyers of data center infrastructure. For a transformer maker or a cooling vendor, though, the useful question is not the dollar total. It is how many megawatts of capacity the money buys, on what timetable, and where. Meta has not broken that out. And if more of each dollar is going to pricier components such as memory, then $137.5bn buys less physical build-out than the same sum would have a year ago.

    A cloud push would add another layer. Capacity built for external customers generally has to meet those customers’ expectations for reliability and connectivity. It could also extend Meta’s need for sites, power and cooling beyond its internal requirements. Until Meta says how much capacity it intends to offer to others, suppliers can plan only against the capex range.

    Background

    Meta Platforms, founded by Mark Zuckerberg as Facebook, owns Facebook, Instagram and WhatsApp. It is one of a handful of “hyperscalers,” companies that design and operate their own data centers at enormous scale. Unlike Amazon Web Services, Microsoft Azure or Google Cloud, Meta has historically built that capacity only for its own apps and advertising systems. It has not rented it to outside customers.

    The race to build artificial intelligence has changed the scale of that spending. Training and running large AI models requires dense clusters of specialized servers, along with the electricity, cooling and networking to support them. That has made capex guidance from Meta, Amazon and their peers one of the most closely watched demand signals for chipmakers, memory suppliers, data center developers and utilities.

    Sources

    Source: Meta boosts AI data center capex, forecasts $130-145bn spend — DatacenterDynamics on Meta’s raised 2026 capital expenditure guidance, second-quarter results and cloud ambitions.

  • Hyperscale Data’s 20 MW AI Contract Still Needs Bitcoin to Get Built

    Hyperscale Data’s 20 MW AI Contract Still Needs Bitcoin to Get Built

    TL;DR · 30-second read

    The Short Version

    A small company building a data center for artificial intelligence in Michigan is paying for it partly by selling bitcoin.

    Hyperscale Data sold about 100 bitcoin and set up a loan that uses its remaining coins as a guarantee, at roughly 4.5% to 5% interest. The building already has a customer lined up, under a deal that could bring in more than $1.2 billion over 20 years.

    The surprise is that a signed customer does not pay for construction. The money has to be found first.

    Hyperscale Data, Inc. (NYSE American: GPUS) said on July 30, 2026 that it has sold approximately 100 bitcoin and put the proceeds directly into its Michigan AI data center campus. It has also set up a credit facility that lets it borrow at variable rates, expected at about 4.5% to 5.0%, with bitcoin pledged as collateral.

    The money will fund construction and long-lead equipment for a previously announced master services agreement with an unnamed neocloud customer. That agreement covers about 20 megawatts of critical AI compute capacity for an initial 10-year term. The company says it could exceed $1.2 billion in revenue if the customer exercises two five-year extensions, and $3.0 billion if the customer also takes an optional 32 additional megawatts.

    Executive Summary

    Hyperscale Data is turning part of its bitcoin treasury into data center capacity. It is doing this in two ways: selling some coins outright and borrowing against the rest. Management calls the Michigan campus one of its highest-priority long-term investments. It presents the plan as a way to avoid issuing new shares at what Executive Chairman Milton “Todd” Ault III calls the company’s current market valuations.

    The announcement matters beyond one small-cap company because of what it shows about how AI infrastructure gets funded. Hyperscale Data already has a signed long-term customer, yet it still reached for its most liquid asset to pay for construction. The company says the alternatives were dilutive equity or “onerous” debt. A contract with a customer and the capital to build for that customer are separate problems, and the second one sets the pace.

    The trade-off is real. Borrowing against bitcoin is cheap only while bitcoin holds its value. The company has not disclosed the facility’s size, terms or collateral requirements, the campus’s total budget, or a delivery date.

    A Signed AI Customer Is Not a Construction Budget

    The master services agreement at the center of this plan is substantial on paper. It covers about 20 megawatts of critical AI compute capacity, meaning the power delivered to the computing equipment itself, excluding cooling and other overhead. The initial term is 10 years, and the customer may choose to add two five-year extensions. The contract reaches the “in excess of $1.2 billion” the company cites only if both extensions are exercised. Spread across 20 years and 20 megawatts, that is roughly $3 million per megawatt per year, or about $250 per kilowatt per month.

    That revenue arrives only after the capacity is built and handed over. Before then, the landlord has to pay for the building, the electrical and cooling plant, and what the release calls “long-lead equipment.” These are items such as large transformers and switchgear that typically must be ordered well in advance. Hyperscale Data says proceeds from the roughly 100 bitcoin it sold are being invested “directly” into this work, and that the credit facility will help fund it further.

    The telling detail is how the company describes its alternatives. It says the bitcoin route reduces reliance on “third party short-term equity financing or onerous debt financing.” Ault says that at current valuations, bitcoin is a more attractive source of capital than issuing stock. A neocloud is a cloud provider that specializes in renting out GPU computing for AI. Having signed one as a customer had not, by itself, made construction money cheap enough for management. The company has not said whether it sought project debt secured by the contract itself. For smaller developers the operational lesson is that the customer queue and the capital queue are separate. The company’s own word for what the new funding does is “accelerate,” which means the capital queue is what sets the build schedule.

    Why Bitcoin Became the Cheaper Money

    The credit facility carries variable rates expected at about 4.5% to 5.0%, secured by pledged bitcoin. For comparison, the company’s listed 13.00% Series D Cumulative Redeemable Perpetual Preferred Stock pays holders 13% a year. Against that benchmark, and against selling common shares at a valuation management considers unattractive, borrowing against an asset already on the balance sheet looks inexpensive.

    Lenders accept bitcoin as collateral because it trades around the clock in deep markets and can be sold quickly if a loan goes bad. That lets them lend on terms closer to secured lending than to speculative construction debt. The company keeps most of its coins and their potential upside. Ault calls the treasury, which the company recently said had reached approximately $71 million in bitcoin, “a significant source of financial flexibility.” Chief Executive Officer William Horne describes the move as “shifting one balance sheet asset for another.”

    That description is accurate as far as it goes. The company is converting a liquid, volatile asset into an illiquid, long-lived one that is meant to produce contracted cash flow. Whether the exchange creates value depends on the build cost, the delivery date and whether the customer exercises its extensions. The release quantifies none of these.

    Volatile Collateral Behind a Long-Lived Asset

    Bitcoin-backed loans typically advance only a fraction of the collateral’s market value. They usually require the borrower to post more collateral or repay if the price falls far enough. Hyperscale Data has not stated the facility’s size, lender, loan-to-value ratio or margin terms, so the scale of that exposure cannot be measured. The structural mismatch is clear, though. A data center takes years to pay back, while the collateral behind its funding is repriced every minute.

    The variable rate adds a second moving part. The 4.5% to 5.0% range is an expectation, not a fixed cost. Rising rates would increase the cost of carrying debt that funds an asset whose revenue is locked into a long-term contract. Neither risk is unusual alone. Together, they are what investors should weigh against the stated benefit of avoiding dilution.

    Why Speed Carries a Premium

    The contract gives the customer a right to 32 additional megawatts. If that right is exercised within the first two years and the contract runs through both extensions, the company expects total revenue above $3.0 billion across about 52 megawatts. That is roughly the same per-megawatt economics as the base deal. Because the expansion window is limited in time, any delay on the first 20 megawatts shortens the period in which the larger deal can materialize. That is a concrete reason to pay for speed with the most readily available capital.

    The company is also narrowing its focus. It expects to divest Ault Capital Group, its private equity and operating-company arm, in the second quarter of 2027. What remains would own and operate data centers and hold digital assets. The Michigan campus is the asset that narrower company will be judged on, which helps explain why management ranks it at the top of its capital priorities.

    Background

    Hyperscale Data, Inc. trades on NYSE American under the ticker GPUS and describes itself as an AI data center company anchored by bitcoin. Its subsidiary Sentinum owns and operates a data center where it mines digital assets and offers colocation and hosting. Its other subsidiary, Ault Capital Group, is a hybrid private equity firm and operating company with holdings in financial services, industrial, defense, hospitality and other sectors. The company expects to divest Ault Capital Group in the second quarter of 2027, leaving a business centered on data centers and digital assets.

    A number of public companies now hold bitcoin as a treasury asset, and some bitcoin miners have been repurposing power and sites for AI computing. AI workloads need dense electrical and cooling infrastructure, and specialized GPU cloud providers, often called neoclouds, increasingly lease that capacity from third-party landlords under long-term contracts.

    Sources

    Source: Hyperscale Data Repurposes Bitcoin Treasury Strategy to Accelerate Development of Michigan AI Data Center (PR Newswire, July 30, 2026): Hyperscale Data announces a bitcoin sale and a bitcoin-backed credit facility to fund its Michigan AI campus.

    Primary sources: Hyperscale Data Form 8-K (filed September 18, 2026); Exhibit 99.1: Hyperscale Data Declares Monthly Cash Dividend on 13.00% Series D and 10.00% Series E Preferred Stock (September 18, 2026).

  • Meta Hands 80% of a $14B AI Campus to BlackRock, Turning Capex Into Rent

    Meta Hands 80% of a $14B AI Campus to BlackRock, Turning Capex Into Rent

    TL;DR · 30-second read

    The Short Version

    Meta, the company behind Facebook and Instagram, is building a huge computer complex in El Paso, Texas, to run its artificial intelligence products. It will draw about as much electricity as a large nuclear reactor produces.

    The twist is that Meta is not paying for most of it. Funds run by BlackRock, the world’s biggest investment manager, will own 80% of the roughly $14 billion project, and Meta will rent the space back.

    It is like building a house, selling most of it, and signing a long lease to keep living there. That shows how expensive the artificial intelligence race has become.

    Meta Platforms and BlackRock announced on July 28, 2026 a joint venture to develop and operate a data center campus in El Paso, Texas, with development costs of about $14 billion, Reuters reported in a story carried by The Journal Record. BlackRock-managed funds will own 80% of the venture and Meta 20%. Meta is contributing land and in-progress construction valued at about $2.3 billion, BlackRock about $4.9 billion in cash, and part of BlackRock’s investment will be financed with $12.5 billion of debt. Meta will also receive a $1 billion distribution to align ownership.

    The campus, already under construction near the Texas-New Mexico border, is designed for 1 gigawatt of compute capacity, with operations expected to begin in 2028. Meta will sign lease agreements with the venture, securing the capacity without funding or owning the campus outright.

    Executive Summary

    Meta has taken one of its largest AI construction projects and moved most of the ownership, and most of the funding burden, to an outside investor. Rather than paying for a 1-gigawatt campus upfront, Meta will hold a minority stake and pay rent for the capacity it uses. The deal lands as AI-related bond issuance has reached $270 billion by early July, according to BofA Global Research, nearly double all of 2025, and as Meta’s shares had fallen about 10% this year on concerns about AI spending.

    The significance is structural. Hyperscalers, the handful of companies that run cloud and internet services at global scale, have mostly funded their own data centers. This venture shows a different model: an asset manager owns the building, the power and the land, lenders finance much of it, and the tech company becomes an anchor tenant. For investors watching Meta’s $600 billion U.S. infrastructure plan, it is a signal that not all of that money has to come from Meta’s own cash flow.

    From Owner to Tenant: The Math Behind the 80/20 Split

    The disclosed figures show how the ownership split was engineered. Meta contributes land and construction already underway, valued at about $2.3 billion. BlackRock-managed funds put in about $4.9 billion in cash. Meta then receives a $1 billion distribution to align ownership, which leaves its net contribution at roughly $1.3 billion. Against BlackRock’s $4.9 billion, that is about 21% of the combined equity, consistent with the stated 20% stake. In effect, Meta recovers a meaningful slice of money it had already spent on the site and hands the remaining funding burden to its partner.

    The second half of the structure is the lease. Capital expenditure, or capex, is spending on long-lived assets such as buildings and equipment, paid upfront. Rent is paid over time. By signing lease agreements with the venture, Meta converts much of the cost of a gigawatt-scale campus from the first kind of obligation into the second. The $12.5 billion of debt that finances part of BlackRock’s investment means lenders to the venture, rather than Meta’s own bond investors, carry a large share of construction financing.

    This is not money that disappears. Long-term leases are commitments, and under U.S. accounting rules they generally appear on a company’s balance sheet as liabilities. What changes is the shape of the obligation: spread over the lease term, paired with an outside owner that bears the long-term value risk of the asset, and freeing Meta’s cash for chips, other campuses or shareholders. The people most affected are Meta shareholders, who see lower near-term spending on this site; BlackRock’s fund investors, who take on a long-dated, single-tenant asset; and the venture’s lenders, who take construction and tenant risk.

    Why Meta Wants Rent Instead of Receipts

    The timing explains the appeal. Meta has said it plans to invest $600 billion in U.S. AI infrastructure, data centers and jobs by 2028, and is building several gigawatt-scale campuses, including a Louisiana site it expects to expand to 5 gigawatts with investment above $50 billion. At that scale, even a company with Meta’s cash flow runs into limits on how much it can spend or borrow directly without unsettling investors.

    Matt Britzman, senior equity analyst at Hargreaves Lansdown, put the concern plainly: the scale of spending raises questions about cash flow, operating costs and returns, particularly because Meta does not yet have a large cloud business selling spare capacity to outside customers. Amazon, Microsoft and Google can rent surplus computing to others; Meta has mostly built for its own use. Reported talks to lease computing power to Anthropic, in a potential deal worth up to $10 billion over two years, suggest Meta is exploring that route, but no such deal had been announced.

    Against that backdrop, a structure that trims upfront spending while keeping full access to the capacity is a direct answer to the question investors are asking. Whether it improves Meta’s economics depends on the lease rate, which has not been disclosed.

    What BlackRock Is Actually Buying

    For BlackRock’s funds, the attraction is a long-term stream of lease payments from one of the world’s largest technology companies, secured by physical assets and, crucially, by access to power. A campus designed for 1 gigawatt, roughly the output of a large nuclear reactor, is hard to replicate quickly, which gives the site value beyond its current tenant.

    The risks are concentration and specialization. The venture has one disclosed tenant, and a building designed for dense AI computing is not easily repurposed. The asset’s long-term value depends heavily on Meta renewing its leases or on the venture finding other tenants willing to pay similar rates. Operations are not expected until 2028, so construction and schedule risk sit with the venture for the next two years.

    One Deal, Not Yet a Template

    Reuters framed the transaction within a broader pattern of tech companies turning to debt markets and fund managers for external capital. That pattern is visible in the bond figures, but one joint venture does not establish that Meta, or its peers, will finance every campus this way. The more telling test will be whether Meta applies the structure to Louisiana or other gigawatt-scale projects.

    What the ownership change does not alter is the local footprint. The campus still needs the same power, land and community acceptance whether Meta or BlackRock owns it, and AI data center expansion faced 142 protests across 42 states in July. For utilities, local governments and residents, the relevant questions about electricity, water and tax treatment remain the same regardless of whose balance sheet carries the building.

    Background

    Meta Platforms, the parent of Facebook, Instagram and WhatsApp, has become one of the largest builders of AI data centers in the United States, with 28 sites in operation or under construction. It had earlier described El Paso as a more than $10 billion project and is also building a Louisiana campus it expects to grow to 5 gigawatts. Earlier in 2026 Meta appointed Dina Powell McCormick as president and vice chairman to drive partnerships with governments and investors for its AI projects.

    BlackRock is the world’s largest asset manager. Its role here reflects a wider shift in which pension funds, insurers and other long-term investors, reached through managers like BlackRock, finance the physical infrastructure behind AI, much as they have long financed toll roads, pipelines and power plants.

    Sources

    Source: Meta, BlackRock announce $14 billion AI data center project in Texas — Reuters report, via The Journal Record, on the El Paso joint venture’s ownership, financing and capacity.

  • Pentagon Offers AI Data Centers Land at 12 Bases; the Real Price Is Off-Grid Power

    Pentagon Offers AI Data Centers Land at 12 Bases; the Real Price Is Off-Grid Power

    TL;DR · 30-second read

    The Short Version

    The US military is inviting private companies to build giant computer warehouses for artificial intelligence on at least 12 of its bases, from Texas to Alaska. Two projects, each estimated at about $2 billion, have been tentatively awarded.

    The military pays nothing up front. In exchange for the land, it gets to use the computers. The catch is that each project must make its own electricity instead of plugging into local power lines, and must not drain local water. In effect, the builders are being asked to build their own power plant too.

    Some proposed sites sit half a mile from family homes, and Congress is still arguing over the rules.

    Military.com reported on July 27, 2026 that the Army and the Department of the Air Force have opened at least a dozen military installations to commercial data center developers over the past year and a half. Only two deals have been awarded, and both are conditional. In late March the Army selected Carlyle for roughly 1,384 acres at Fort Bliss, Texas. It selected CyrusOne, a data center operator jointly held by funds managed by KKR and BlackRock, for a roughly 1,201-acre parcel at Dugway Proving Ground, Utah. The Financial Times estimated each project at about $2 billion.

    The other sites are still solicitations. The Army is considering Fort Hood, Texas, and Fort Bragg, North Carolina. The Air Force has sought bids at Arnold, Edwards, Joint Base McGuire-Dix-Lakehurst, Davis-Monthan and Robins, and more recently at Joint Base Elmendorf-Richardson, Eielson Air Force Base and Clear Space Force Station in Alaska. Proposals must include net-zero water usage and a power plan that does not draw on the local electrical grid. Initial operating capability is projected for fiscal 2027 at Fort Bliss and fiscal 2029 at Dugway.

    Executive Summary

    The Pentagon is turning spare base land into a currency for AI computing. Under the Army’s Enhanced Use Lease program, private developers finance, build, operate and eventually decommission hyperscale data centers on military land. Hyperscale refers to the warehouse-sized campuses built for the largest cloud and AI workloads. In return the service gets access to computing power, at what the Army calls “no upfront cost to taxpayers.” Army Secretary Dan Driscoll has described Fort Bliss as “the first hyper-scale data center that the Pentagon has ever done.”

    The terms matter more than the land. Unlike commercial projects off base, these bids must bring a power plan that does not rely on the local grid and must achieve net-zero water use. For developers, that changes the job from building a data center to building a data center plus its own utility. For base commanders and the surrounding communities, siting and power have become part of the procurement itself.

    The program is also politically unsettled. A House defense bill provision would bar leases unless developers exclude certain components from China, Russia, Iran and North Korea, and the Army has objected that the rule would penalize building on federal land.

    The Land Is Free. The Power Plant Is Not.

    In the commercial market, the scarcest input for a large AI campus is usually a grid connection, not land. A grid connection is the utility’s commitment to deliver large, steady amounts of electricity, and it often comes only after a long wait in an interconnection queue, the line of projects waiting for approval to plug in. The Army and Air Force solicitations rule that path out. Proposals must include a power plan that does not draw on the local electrical grid, a requirement that does not apply to identical projects built off base.

    That changes what is actually being procured. The Army’s side of the bargain, land in exchange for computing access, is accurate in cash terms, but it moves the heaviest capital item onto the bidder. Whoever builds at Fort Bliss or Dugway is committing to finance and run generation as well as the data halls. The Army’s long-term vision, as described by Col. John Oliver, is a campus with a commercial side, a classified military data side and onsite power generation. The roughly $2 billion estimate for each project has not been broken down publicly, so nobody outside the negotiations can yet say how much of it is power plant and how much is computing.

    The rule is likely to favor sponsors with deep capital and power-development experience, such as a global investment firm like Carlyle or a KKR- and BlackRock-backed operator like CyrusOne, over developers whose model depends on buying utility power. It also changes the utility relationship. El Paso Electric attended the Army’s spring listening session at Fort Bliss, and the Army says excess power generated on base could potentially be sold back to civilian grids. A would-be large customer could become a supplier instead. For residents, the rule speaks directly to the concern behind a Virginia state-funded study, which found monthly household energy bills there could rise by $14 to $37 by 2040 as data centers grow.

    Water Is the Second Utility the Bidder Must Bring

    Data centers use water to cool their servers, and the net-zero water requirement is hardest to meet at the Army’s flagship site. Defense Department procurement documents reported by Data Center Dynamics rate the water risk for the El Paso area as “Extremely High.” One option under discussion is having Carlyle drill a new well to feed El Paso’s desalination plant and offset what the data center uses. Oliver was candid that this is unproven: “We don’t know if that’s an engineering solution that we can actually get to yet.”

    Water and power are linked. Cooling designs that use less water generally use more electricity. In a desert location, a strict water target therefore adds to the size of the on-site generation the developer must build and pay for. The two requirements compound rather than sit side by side.

    A Dozen Sites, Two Awards, Very Different Neighbors

    The sites vary widely. Dugway, where the military conducts its primary chemical and biological weapons testing, offers three parcels totaling about 3,466 acres, roughly an hour’s drive from the nearest community. Fort Hood has 207 acres on offer, and Fort Bliss has nearly seven times that. Contracting documents show a Fort Hood parcel within a half-mile of residential and commercial property, and several Fort Bragg locations within a mile of civilian areas and a half-mile of civilian housing. Smaller parcels close to homes leave less room for generation, and less buffer for its noise, than a remote range does.

    Opposition near homes is not a hypothetical risk. Data Center Watch counts nearly $156 billion in projects nationwide delayed or canceled after local opposition. The Army has built community engagement into the bid: developers must describe outreach plans and assess “any risks or opposition.” Brookings senior fellow Darrell M. West said communities that accept these facilities tend to be the ones told the details up front, including energy source, water use, cost and noise.

    Congress Is Still Writing the Rules

    Rep. Cory Mills secured a provision in the House fiscal 2027 defense authorization bill barring data center leases unless developers avoid certain printed circuit boards, advanced semiconductors and chipsets made in China, Russia, Iran or North Korea. The Army objected that it would create a “federal land penalty,” imposing rules on installations that would not apply to identical projects anywhere else. Rep. John Garamendi’s proposal to require evaluation of energy, water, noise, light and security effects before any deal was rejected by House Armed Services Republicans. A Senate amendment to pause leasing until the Pentagon finalizes a data center strategy was also defeated.

    Each condition adds to the same balance sheet. Free land has to pay for off-grid generation, net-zero water and possibly a restricted supply chain, and the Army’s own argument shows it knows the margin is not unlimited. Whether the economics still work at scale is the question the two conditional awards will answer first.

    Background

    AI workloads have pushed demand for hyperscale data centers faster than utilities can connect them, and local opposition has become a real constraint, with Data Center Watch counting nearly $156 billion in projects delayed or canceled. The Pentagon holds large tracts of underused land at remote and secure installations, and a Trump executive order directed agencies to open non-excess federal land to data center development. The Army’s Enhanced Use Lease program, with the U.S. Army Corps of Engineers handling lease negotiations and environmental review, is the vehicle.

    Carlyle is a global investment firm. CyrusOne is a data center operator jointly held by funds managed by KKR and BlackRock. Dugway Proving Ground in Utah is where the U.S. military conducts its primary chemical and biological weapons testing, and Fort Bliss in El Paso is home to the Army’s 1st Armored Division.

    Sources

    Source: Pentagon Plans AI Data Centers at Military Bases Across Multiple Branches (Military.com), on Army and Air Force plans to lease base land to commercial AI data center developers.

  • AMD-Cerebras 5x Per-Watt Claim Rests on Splitting AI Inference Across Vendors

    AMD-Cerebras 5x Per-Watt Claim Rests on Splitting AI Inference Across Vendors

    TL;DR · 30-second read

    The Short Version

    Two chipmakers, AMD and Cerebras, are teaming up so that a single chatbot answer is produced by both companies’ hardware working together.

    When you ask an artificial intelligence assistant something, the computer does two jobs: it reads your question, then writes the reply word by word. AMD’s machines will do the reading. Cerebras’ dinner-plate-sized chip will do the writing.

    The companies say the pairing could produce up to five times more output for the same electricity. That matters because power is one of the biggest limits on building the giant computer buildings behind these services. It is due later this year.

    AMD and Cerebras Systems announced on July 23, 2026, at AMD’s Advancing AI 2026 event, a technical partnership that combines AMD’s Helios rack-scale systems, built on AMD Instinct GPUs, with the Cerebras Wafer-Scale Engine in a single disaggregated inference workflow. In the companies’ joint announcement, Helios handles prompt processing and large context windows, while the Wafer-Scale Engine handles token generation, the step that streams the answer back to the user.

    The companies say the combined system is expected to deliver up to 5x higher tokens per second per watt. Cerebras plans to deploy Helios systems in its own data centers, with the joint solution expected to be available first through Cerebras Cloud in the second half of 2026.

    Executive Summary

    The announcement pairs two very different kinds of AI hardware inside one request. AMD’s GPU racks take the first, compute-heavy pass over a prompt. Cerebras’ wafer-scale processor then generates the output, a phase limited mainly by how fast data can be pulled from memory. The companies aim the product at the ultra-low-latency segment of inference: coding tools, real-time copilots, live agents and agentic workflows, where response time directly shapes how useful the system is.

    It matters for three reasons. First, the headline efficiency figure is measured per watt, and electricity is increasingly the scarcest input for new AI capacity. Second, it makes disaggregated inference, the practice of running the two phases of a request on separate hardware, a cross-vendor design rather than something done inside one company’s GPU fleet. Third, it turns Cerebras into an AMD customer and operator of AMD systems, with the integration burden sitting in Cerebras’ data centers rather than with end buyers.

    What is not yet established is the performance itself. The 5x figure is an ‘up to’ projection. The product is forward-looking, with availability targeted for the second half of 2026, and neither company has disclosed deployment scale or commercial terms.

    The 5x Claim Rests on the Split, Not on Either Chip

    Every response from a large language model is produced in two phases. Prefill reads the whole prompt in one parallel pass, whether that prompt is a question, a pasted contract or an entire codebase. That work suits hardware built for massive arithmetic throughput. Decode then produces the answer one token, roughly a word fragment, at a time. Each step has to re-read the model’s weights and the stored context of the conversation from memory, so decode speed is capped by memory bandwidth more than by raw compute. The AMD–Cerebras announcement describes exactly this division of labour. Helios processes ‘prompts and large context windows’, while the Wafer-Scale Engine ‘accelerates the memory-bandwidth-intensive token generation.’

    That division is the mechanism behind the headline number. The companies say the two engines together are expected to deliver up to 5x higher tokens per second per watt. They present the gain as a property of the pairing, not of either system alone. A GPU rack running both phases spends much of decode waiting on memory. A wafer-scale system running both phases spends capacity on prefill work that GPUs batch efficiently. Assigning each phase to the hardware whose strength matches that phase’s bottleneck is how a combined system could outperform either one, provided the handoff between the two machines is fast and cheap enough.

    The choice of metric matters for the people who build and power these facilities. Tokens per watt means output per unit of electricity, and grid power, more often than floor space, now caps how much AI capacity an operator can bring online. The words ‘up to’ mark the figure as a best case. Until AMD and Cerebras publish the configuration and baseline behind it, it is a design target rather than a measured operating result.

    One Answer, Two Vendors’ Hardware: What Operators Inherit

    Splitting prefill from decode is not new in itself; large inference providers have separated the phases across different pools of GPUs. The new element here is that every request crosses a vendor boundary. When Helios finishes prefill, the prompt’s working state, known as the key-value cache (the model’s stored record of everything it has just read), has to move to the Cerebras system before the first output token appears. That transfer sits directly on the latency path the product is being sold on, so the network link between the two systems effectively becomes part of the product.

    It also puts two hardware platforms, two software stacks and two support relationships inside one serving pipeline. It means two distinct system designs, likely with different power densities and cooling requirements, sharing the same data hall. Scheduling is harder too, because prefill and decode capacity must be sized against each other: too few Helios racks starve the wafer-scale engines, and too few wafer-scale engines leave GPU output queued.

    The launch model shields buyers from most of that. Cerebras plans to deploy Helios in its own data centers and offer the joint solution first through Cerebras Cloud in the second half of 2026, so customers would consume it as a service rather than integrate it themselves. Cerebras, not the end customer, absorbs the integration risk, along with the capital cost of buying, powering and hosting AMD racks.

    What Each Company Gets Out of It

    For AMD, the partnership extends Helios into the most latency-sensitive end of the inference market without AMD building a dedicated decode chip. Helios stays positioned, in the announcement’s words, as ‘the foundation for high-throughput and balanced inference workloads across the data center.’ It also makes Cerebras a buyer and operator of AMD systems, adding another deployment of Helios in a production cloud.

    For Cerebras, the arrangement concentrates its wafer-scale hardware on the phase where it says it is strongest. Prompt processing, which grows heavier as context windows lengthen and agentic workflows chain many calls together, goes to GPU racks built for throughput. Andrew Feldman framed the benefit as reach: partnering with AMD lets Cerebras ‘bring that performance to even more customers.’ That fits Cerebras’ own risk disclosures in the announcement. It lists dependence on a small number of significant customers, including OpenAI, Group 42 Holding, Mohamed bin Zayed University of Artificial Intelligence and AWS, and the need to secure data center capacity and capital for its cloud offerings. The partnership broadens the product, but it also adds a capital-intensive deployment commitment.

    For the wider market, this is one partnership, not proof of an industry shift. It does put a concrete test in front of buyers: whether heterogeneous inference, which matches different chips to different phases of a single request, delivers enough latency and energy gains to justify the added operational complexity. If Cerebras Cloud demonstrates those gains in production, similar cross-vendor pairings become more likely. If it does not, single-platform stacks keep their simplicity advantage.

    Background

    Cerebras Systems, listed on Nasdaq under the ticker CBRS, builds AI systems around its Wafer-Scale Engine, a processor made from a whole silicon wafer rather than many chips cut from it. The company sells hardware on premises and, increasingly, inference as a cloud service, and it markets itself on speed of token generation. Its disclosures name OpenAI, Group 42 Holding, Mohamed bin Zayed University of Artificial Intelligence and AWS as significant customers.

    AMD is the main GPU alternative to Nvidia in AI data centers through its Instinct accelerator line, and Helios is its rack-scale system, which packages GPUs, processors and networking into a single deployable unit. The partnership reflects a wider industry move toward splitting inference into prefill and decode stages and optimising each separately, an approach previously used mainly within single-vendor GPU clusters.

    Sources

    Source: AMD and Cerebras Announce Industry-Leading Ultra-Low-Latency and High Throughput AI Inference Solution, the joint AMD and Cerebras Systems announcement of July 23, 2026, unveiling a disaggregated inference solution pairing AMD Helios with the Cerebras Wafer-Scale Engine.

  • Arm’s $53B Passing x86 Shows AI Servers Are Now Bought by the Rack

    Arm’s $53B Passing x86 Shows AI Servers Are Now Bought by the Rack

    TL;DR · 30-second read

    The Short Version

    Companies spent $89.7 billion on the computers behind artificial intelligence in just the first three months of 2026, according to the research firm IDC. That is about a third more than a year earlier.

    The bigger surprise is which chips won. Servers built on Arm’s chip designs, best known from smartphones, now take more of that money than servers built on the Intel and AMD design that has run data centers for decades.

    The reason is how big buyers shop. They now order whole ready-built cabinets of equipment, and the chip simply comes with the cabinet.

    IDC reported on July 21, 2026 that worldwide AI infrastructure spending reached $89.7 billion in the first quarter of 2026. That was up 33.1% from a year earlier but essentially flat against the record fourth quarter of 2025. Servers accounted for $87.6 billion of the total. Within accelerated servers, meaning machines built around GPUs or other AI chips, Arm-based systems took $53.0 billion and x86 systems took $34.6 billion.

    The research firm also raised its full-year 2026 forecast to $497 billion. That is roughly 56% growth over 2025’s $318 billion. IDC now projects the market will pass $1 trillion in 2029, at $1.08 trillion, and reach $1.21 trillion in 2030.

    Executive Summary

    The headline number is large, but the more consequential finding is structural. In two quarters, x86 accelerated server value fell from $51.9 billion to $34.6 billion, while Arm-based value rose from $29.8 billion to $53.0 billion. IDC says the crossover began in Q4 2025. It attributes the shift to large buyers consolidating around NVL72/GB200-class rack-scale platforms and moving volume away from custom x86 rack designs.

    That matters well beyond chip vendors. In these deals the buyer is purchasing a pre-integrated rack rather than an individual server. The processor architecture therefore arrives bundled with the accelerators, the interconnect, and the rack’s power and cooling requirements. Server makers, data center operators and the utilities that feed them all end up planning around whichever rack design the buyer picked.

    IDC’s other signals point the same way. Spending remains concentrated in the United States, at 75.7% of the global total. Enterprises are now catching up on storage refreshes they had deferred. And IDC names power and grid capacity as the primary operational constraint on commissioning new data centers.

    The Processor Now Rides In With the Rack

    IDC’s platform figures show a clean swap. Arm-based accelerated server value went from $29.8 billion in Q3 2025 to $47.5 billion in Q4 and $53.0 billion in Q1 2026. Over the same three quarters, x86 went from $51.9 billion to $42.7 billion to $34.6 billion. Added together, the two barely moved between the last two quarters: $90.2 billion in Q4 2025 and $87.6 billion in Q1 2026. The market’s size held roughly steady while its mix flipped. IDC research director Juan Seminara described it this way: “that’s not demand destruction, that’s an architecture shift.”

    The mechanism lies in how these systems are bought. IDC ties the crossover to large buyers standardizing on NVL72/GB200-class rack-scale platforms. These are Nvidia designs in which a full cabinet of GPUs is wired together to behave like one large machine, with Nvidia’s Arm-based Grace processors serving as the host CPUs. GPUs dominate the cost of a system like that. As a result, these dollar figures largely measure which rack design buyers chose. The Arm processor comes with that choice; it did not win a separate CPU-by-CPU evaluation against Intel or AMD. The customer buys the rack, and the CPU architecture follows.

    The effects reach several groups. IDC flags execution risk for x86-focused OEMs and ODMs, the branded server makers and contract manufacturers, that have not diversified their rack-scale roadmaps. For data center operators, the unit of planning changes too. A pre-integrated rack arrives with fixed power and cooling requirements, so a facility must be ready for it on delivery rather than adapting one server at a time. IDC is explicit that the contest is not settled, because new x86 platforms are coming. But the competition has moved from the individual processor to the whole rack.

    A Flat Quarter Behind a Raised Forecast

    Q1 2026 spending of $89.7 billion was up 33.1% year over year but essentially flat against Q4 2025. IDC reads this as growth off a much larger base, not a softening of demand. The base has expanded quickly: spending went from $153 billion in 2024 to $318 billion in 2025, more than doubling.

    The raised full-year forecast is where the numbers need scrutiny. IDC now expects $497 billion in 2026, about 56% growth, up from a pace of roughly 53% it estimated a quarter earlier. Subtracting Q1 leaves $407.3 billion for the remaining three quarters. That is an average of about $136 billion a quarter, roughly 50% above the Q1 run rate. The forecast therefore assumes the flat first quarter was a pause before sharp sequential acceleration, not a new plateau. IDC’s own suggested test is the right one: Q2 2026 capital expenditure guidance from the leading hyperscalers, the largest cloud operators, and from AI platform providers.

    Further out, IDC projects $1.08 trillion in 2029 and $1.21 trillion in 2030, a compound annual growth rate of about 30% from 2025. It lists four possible accelerants: broader inference deployment, sovereign AI programs in the Middle East, Southeast Asia and Europe, new model architectures and agent frameworks, and AI demand that does not run on GPUs at all.

    Not All AI Demand Lands on GPUs

    Servers made up 97.6% of Q1 AI infrastructure value, and storage made up 2.4%, or $2.2 billion. Inside the server figure, IDC points to a growing share of AI-related demand landing on hardware that is not GPU-accelerated. That includes orchestration tooling, data-pipeline workloads, and CPU-only inference clusters that hyperscalers run to contain costs alongside their GPU buildouts. This is a second, quieter market for general-purpose processors. The Arm-versus-x86 decision there is made on different terms from the rack-scale GPU buildout.

    Storage is catching up after a deliberate delay. IDC says enterprises spent the past one to two years redirecting budget to GPU servers and treating storage refresh as something they could postpone. They no longer can, so pent-up refresh demand is now arriving on top of genuine AI-driven demand. IDC also warns that memory and storage component scarcity can raise server bills of materials, meaning the component cost of each system, and slow procurement.

    Regionally, the United States spent $67.9 billion, or 75.7% of the global total, and grew 30.3%. China returned to growth at $7.8 billion, up 9.3%. The fastest growth came from smaller bases. The Middle East and Africa rose 233% to $1.1 billion, and Asia/Pacific excluding Japan and China rose 62% to $5.8 billion. Western Europe reached $5.1 billion. IDC cautions that Middle East growth is concentrated in a small number of large, government-backed Gulf deals. It adds that regional tensions, including the conflict involving Iran, could delay procurement.

    Power, Not Hardware, Sets the Schedule

    IDC names power generation and grid capacity as the primary operational bottleneck for commissioning new data centers in major markets. It says commissioning timelines are increasingly driven by utility capacity rather than hardware lead times. That observation connects directly to the platform shift. When buyers standardize on dense, pre-integrated racks, the question becomes whether a site can deliver the power and cooling those racks require, not whether the servers can be sourced.

    For operators and their financiers, a $497 billion forecast is only achievable to the extent that energized capacity exists to host it. IDC adds two more external constraints: export controls and data-sovereignty rules. Either could reshape where AI workloads are deployed and which vendors win enterprise deals.

    Background

    IDC is a technology market research firm and part of IDG. Its Worldwide Quarterly AI Infrastructure Tracker measures spending on the servers and storage used for AI workloads, broken out by region, by component and by platform. The tracker showed spending more than doubling from $153 billion in 2024 to $318 billion in 2025 as the largest cloud operators expanded training infrastructure.

    x86 is the processor architecture developed by Intel and also used by AMD, and it has dominated data center servers for decades. Arm is a rival architecture that Arm Holdings licenses to chip designers. Nvidia’s Grace CPU, which pairs with its Blackwell GPUs in GB200 systems, is built on Arm. In an accelerated server, GPUs or other AI chips do most of the computation while the CPU acts as host, coordinating work and moving data.

    Sources

    Source: AI Infrastructure Spending Holds Near $90 Billion in Q1 2026 as ARM Overtakes x86 in Accelerated Servers; 2026 Forecast Raised to $497 Billion. IDC’s Q1 2026 results from its Worldwide Quarterly AI Infrastructure Tracker, with its raised 2026–2030 forecast.

  • Savannah River Pick Shows DOE’s 16 AI Sites Are a Power Bet, Not Just a Land Deal

    Savannah River Pick Shows DOE’s 16 AI Sites Are a Power Bet, Not Just a Land Deal

    TL;DR · 30-second read

    The Short Version

    The United States government is opening some of its own land to companies that want to build the huge computer warehouses that run artificial intelligence. One of the first chosen sites is in South Carolina, at a lab that supports the nation’s nuclear weapons work.

    The twist: these projects are meant to come with their own power plants. The biggest artificial intelligence facilities can need as much electricity as a large power plant produces, and building power on-site is meant to keep that cost off ordinary households’ bills.

    Engineering firm Amentum was picked to build it. No price, size or start date has been announced.

    Microgrid Knowledge reported on July 20, 2026 that the National Nuclear Security Administration (NNSA), part of the U.S. Department of Energy (DOE), has selected Virginia-based engineering firm Amentum to develop a co-located artificial intelligence data center and on-site power generation project at the Savannah River National Laboratory site in South Carolina. The selection followed a competitive request for proposals that DOE began last year, after identifying 16 federal sites it judged could host AI data centers paired with their own energy infrastructure.

    The effort, formally the Savannah River Site Artificial Intelligence Infrastructure and Energy Generation Project, has no announced timeline, generating technology, capacity or cost. DOE says the project could meet its own electricity needs without shifting large costs to utility customers and could also increase the power available to the grid.

    Executive Summary

    NNSA has named Amentum to lead construction of both an AI data center and the power plant that would feed it, on federal land at one of the country’s principal nuclear security sites. It is a concrete step in DOE’s program to put AI computing on government property, and the first selection at Savannah River under that program.

    The significance lies less in the land than in the pairing. DOE did not simply offer acreage; it screened its sites for their ability to host data centers together with co-located generation, and it justified the Savannah River project on the grounds that it would not push large grid-upgrade costs onto utility customers. That makes federal sites a potential route around the slowest and most politically sensitive part of AI buildout: getting enough electricity to the building.

    What has been announced is a selection and a scope, not a financed, sized project. The generating technology, megawatt capacity, tenant, financing and schedule are all still undisclosed, so the model’s claims about cost and grid benefits remain untested.

    Sixteen Sites, One Condition: Bring the Power

    DOE’s list of 16 federal sites was not a catalogue of empty lots. According to the department, it identified sites that could support AI data centers paired with co-located energy infrastructure. The Savannah River award makes that pairing explicit: generation is in the project’s official name, and Amentum is contracted to build the power supply alongside the data center rather than to plug into whatever the local utility can provide.

    The reason is the grid connection itself. A large new electricity user normally joins the grid through an interconnection process, in which the utility studies whether its lines and power plants can serve the new load and what must be built if they cannot. At AI scale, those upgrades are substantial: industry projections cited in the coverage put future AI server racks at up to 1 megawatt each and the largest AI campuses at around 1 gigawatt, roughly the output of a large power plant. When a utility builds for a customer that size, the question of who pays tends to fall on its other customers. DOE’s stated case for Savannah River answers that directly, saying the project could generate its own needs “without shifting huge costs to utility customers.”

    That is where federal land earns its place. A large, secured, government-controlled site offers space, a single landowner and an existing security perimeter for putting generation right next to the load, which many sites closer to population centers struggle to offer. The trade-off is engineering: AI training creates wide, fast swings in power demand, and a self-supplied site has less grid backing to absorb them. Whatever generation Amentum ultimately builds will need to be designed around that volatility, which makes the power system as central to the project as the data center it serves.

    Why a Nuclear Security Site, and Why Amentum

    NNSA Administrator Brandon Williams framed the award as a way to “move faster, apply innovative technologies, and make productive use of federal land while maintaining our commitment to mission delivery.” That last clause matters. Savannah River hosts sensitive defense work, and any private-sector partner operating there has to coexist with it. Amentum already does; it was selected earlier this year to provide design and engineering services for the site’s Tritium Finishing Facility, which supports the nuclear weapons program.

    That existing footprint is a competitive advantage that is hard to replicate. Contractors that already understand federal security, procurement and site operations are positioned to win the first wave of these awards, while conventional data center developers may find their route onto federal land runs through partnerships with such firms. Amentum and NNSA say they will collaborate on expanding private-sector partnerships in the project, which leaves room for tenants, financiers or generation owners still to be named.

    What Is Settled, and What Still Rests on “Could”

    The announcement establishes a selection made through a competitive process, a defined scope covering both compute and generation, and a named site. Those are real commitments of intent from a federal agency. The headline benefits are conditional, however: the project could meet its own electricity needs, and it could increase power available to the grid. Williams himself called it a “proposed partnership.”

    Until the parties disclose the generation type, capacity and commercial structure, the ratepayer-protection claim cannot be measured. Its value depends on details not yet public: whether the site stays connected to the grid for backup, who pays for that connection, and on what terms any surplus power is sold. The first projects under DOE’s program will set the template for the remaining sites, so those terms will matter well beyond South Carolina.

    Background

    The Savannah River Site in South Carolina is a Department of Energy complex overseen by the National Nuclear Security Administration, the DOE agency responsible for nuclear weapons and nuclear security. Its national laboratory supports defense programs including tritium production. Amentum, a Virginia-based firm spun out of AECOM about six years ago, ranks as the second-largest U.S. government services contractor behind Leidos and was already working at the site on the Tritium Finishing Facility.

    The award sits within a wider federal push to use government land for AI infrastructure. DOE identified 16 of its sites as candidates for data centers paired with their own power, and Energy Secretary Chris Wright has described the AI race as “the next Manhattan project.” The pressure behind that push is electricity: an analysis by the Penn State Institute of Energy and the Environment projects AI’s share of global electricity use could rise from 4% to 20% by 2035.

    Sources

    Source: Amentum to Develop Co-Located AI Data Center and Power Infrastructure at Savannah River National Laboratory (Microgrid Knowledge, July 20, 2026): NNSA selects Amentum to build an AI data center with on-site power generation at the Savannah River Site.

  • Fireworks’ $1.5B Round Shows AI Inference Is Now a Layer Above the GPU Clouds

    Fireworks’ $1.5B Round Shows AI Inference Is Now a Layer Above the GPU Clouds

    TL;DR · 30-second read

    The Short Version

    A young company called Fireworks just raised $1.5 billion from investors. It doesn’t invent its own artificial intelligence. It runs freely available artificial intelligence programs for other businesses, using computer chips spread across more than a dozen cloud companies.

    Its yearly sales pace reportedly jumped from $100 million to over $1 billion in 16 months. Shopify and Twilio took four to five years to make the same climb.

    Why it matters: the money in artificial intelligence isn’t only going to chipmakers and tech giants. The middleman who keeps these systems fast and cheap is now big enough to fund on its own.

    Fireworks, the training and inference platform for open-weight AI models, has raised a $1.5 billion Series D, according to Bessemer Venture Partners, which said on July 17, 2026 that it had joined the round as an existing investor. Bessemer said Fireworks has passed $1 billion in annual recurring revenue (ARR), up from $100 million 16 months earlier. It also said the company processes about 43 trillion tokens a day. Tokens are the word-fragments AI models read and write.

    The company runs on more than a dozen cloud providers across over 20 regions. It is a first-party partner shipping natively inside Microsoft Foundry on Azure. It is led by CEO Lin Qiao, who led the Meta engineering organization that built the PyTorch framework, and by President George Hu, who joined in April 2026.

    Executive Summary

    Fireworks does not sell a chatbot or a proprietary model. It runs other people’s open-weight models, and the fine-tuned versions customers build on them, as fast and as cheaply as possible. It charges by the token. A $1.5 billion round for that business, at more than $1 billion in ARR, is a large bet that inference is a market of its own. Inference means running a trained model to answer requests, and on this bet it is neither a feature of the chip clouds nor of the model labs.

    For infrastructure operators, the important detail is the architecture. Fireworks describes its platform as a virtual GPU cloud. It pools graphics processors from more than a dozen providers and sits between those providers and the enterprises consuming AI. The more demand flows through layers like this, the more GPU owners compete as suppliers to software platforms rather than selling directly to end customers.

    The growth figures are striking. So far, though, they are headline numbers without the detail needed to judge durability. That detail includes margins, customer concentration, how much capacity is contracted, and how ARR is defined for a usage-based business.

    A Platform That Sits Above the Chips

    What Fireworks sells is not compute capacity in the usual sense. Its platform is described as a virtual GPU cloud. That is software pooling graphics processors (GPUs, the chips that do most AI work) from more than a dozen cloud providers in over 20 regions. On top sits a proprietary inference engine that decides how to run each model at maximum speed and efficiency. Customers buy output, measured in tokens, not servers. At roughly 43 trillion tokens a day and more than $1 billion in ARR, that software layer is now a business of real scale. The $1.5 billion Series D finances it as one.

    That is the operational shift. In the familiar cloud model, whoever owned the machines owned the customer. In this model, GPU owners become suppliers. Those owners include hyperscalers (the largest cloud providers, such as Azure) and neoclouds (newer GPU-specialist clouds). The platform holds the customer relationship, decides where each workload runs, and keeps the margin from running it efficiently. Bessemer states the ambition directly: Fireworks wants to be the software layer that abstracts away inference running on “heterogeneous silicon, hyperscalers, neoclouds, and air-gapped environments” — the inference platform “running on every GPU in the world.”

    For GPU cloud and data center operators, this cuts both ways. An aggregator pools demand from thousands of customers, which can make it a large and steady buyer of capacity. But demand routed through a multi-cloud layer is also portable: it can move toward whichever provider offers the best price, availability or location. The Microsoft relationship shows the layer can coexist with hyperscalers rather than simply competing with them. Azure distributes Fireworks natively through Foundry. Still, this is one company’s round. It shows the independent inference layer can be financed at scale, not yet that it will hold its margin once hyperscalers price their own inference services against it.

    Open Models Move the Spend From the Model to the Serving

    The investment thesis rests on a claim about models, not chips. Open-weight models, whose trained parameters are published for anyone to run, now match closed frontier models on the workloads most enterprises care about, at a fraction of the cost. On that premise, enterprises will increasingly post-train an open model on their own proprietary data instead of paying per call for a closed API. They would, in Bessemer’s phrase, stop wanting to “rent” their core intelligence.

    If that premise holds, value shifts toward whoever serves and tunes those open models well. Fireworks’ pitch is that it does both in one place. Its platform combines fine-tuning and reinforcement learning (methods for adapting a model with examples and feedback) with production serving. Every interaction can feed the next version of a customer’s model. That combination also creates switching costs. A customer whose tuned model, data pipelines and serving setup all live on one platform has reasons to stay. That stickiness is what justifies valuing the layer as more than a GPU reseller.

    The load-bearing premise, however, is asserted rather than demonstrated. No benchmarks or customer workloads accompany the claim that open models match frontier quality. The answer will vary by task. Buyers weighing an open-model strategy should test it on their own workloads rather than assume parity.

    What the Growth Numbers Establish, and What They Don’t

    Growth from $100 million to more than $1 billion in ARR in 16 months is, by Bessemer’s comparison, a climb that took Twilio and Shopify four to five years. It establishes that demand for hosted open-model inference is very large right now. For a usage-based business, however, ARR usually annualizes current consumption. It can fall if customers optimize their prompts, switch to smaller models, or move volume to another provider. That makes it a different quantity from contracted subscription revenue.

    Revenue is also not margin. A platform built on capacity drawn from other clouds has to pay for that capacity. Its economics depend on how many tokens its engine can produce per GPU-hour relative to what it pays for that hour. The inference engine is the core asset for that reason: efficiency is the margin. The broader market figures cited alongside the round are context, not evidence about Fireworks. They include hyperscaler capital spending of more than $800 billion in 2026 and token consumption expected to rise more than 30-fold by the end of the decade, and no source is given for either projection.

    Background

    Fireworks was founded by engineers who built and scaled PyTorch at Meta, the open-source framework on which much of today’s AI software is built. CEO Lin Qiao led that engineering organization. The company positions itself as a training and inference platform for open-weight models. It combines a proprietary inference engine, a multi-cloud GPU pool and tools for fine-tuning and reinforcement learning. Bessemer Venture Partners was already an investor before the Series D.

    The round lands as AI spending shifts from building models toward running them. Enterprises that first used closed frontier-model APIs are weighing open models they can tune on private data and serve more cheaply. A growing set of providers now competes to run those models: hyperscalers, GPU-specialist neoclouds and independent inference platforms.

    Sources

    Source: Fireworks raises $1.5B Series D for open model inference – Bessemer Venture Partners: Bessemer’s announcement of its participation in Fireworks’ Series D and its investment thesis.