Who pays for the AI build-out, and who is left holding the debt if compute returns stall?

AI infrastructure financing is the machinery of cash, loans, leases, customer contracts and supplier commitments that turns the demand for computing into chips, servers, data centres and electricity. It grew out of several older stories: the cloud's pay-as-you-go business model, the graphics chip that turned out to be good at training neural networks, the property and power markets that have always financed physical infrastructure, and a newer kind of lending that treats GPUs and compute contracts as collateral. The demand and the spending are real. They do not guarantee that every new facility will earn enough to pay for itself. Hyperscalers with large, cash-generating businesses can absorb the most risk. The greatest direct danger falls on borrowers whose debts and fixed commitments last longer than their customers' contracts, or longer than their hardware stays useful.

Chapter 1: The wires beneath the cloud

On 24 April 1996, the Federal Energy Regulatory Commission issued Order No. 888. It required US utilities that owned transmission lines to offer open-access, non-discriminatory service to other people's power.¹ The commissioners were thinking about generators and wholesale electricity markets. Chatbots were not on their minds. Even so, the order helped set the rules under which independent power producers, companies that generate electricity without owning the regulated monopoly wires, could sell into the grid, and under which very large customers would one day bargain for power.

That is the first surprise about AI infrastructure financing: it did not begin with AI. It began with the old, slow, heavily regulated business of moving electrons, and with the equally old business of financing buildings and cables. Before anyone could rent a GPU by the hour, someone had to own a substation, a fibre route and a warehouse with very good air conditioning.

A factory whose product is time

The simplest way to picture a data centre is as a factory whose product is computing time. Electricity goes in, heat comes out, and in between the machines sell their cycles to whoever is paying. Like any factory, it needs land, a building, a power connection, cooling and production equipment, and someone has to pay for all of it before the first customer is billed.

The analogy breaks at the server rack. In most factories the expensive machines last for decades and the building is the cheap part. In an AI data centre the order is reversed. The power connection and the concrete can last a generation. The accelerators inside, the chips that do the work, can be overtaken by a faster generation within a few years. A lender looking at a data centre is really looking at two assets bolted together: one that ages like a bridge and one that ages like a smartphone. That mismatch runs through everything that follows.

Two kinds of power company

Order 888 also helped separate two threads that this story keeps apart. Regulated utilities own the poles, wires and often the power plants in a given territory. They earn a return that regulators approve on the money they invest, their "rate base". Independent and merchant generators sell power into competitive markets or under bilateral contracts, and they live with commodity price swings. Today's AI power deals sit on both sides of that line. A utility such as Entergy can recover the cost of serving a data centre through regulated rates, while a merchant nuclear owner sells to a tech company at whatever price the two agree. One is a public-policy question, the other a negotiation.

The property thread was taking shape too. Digital Realty $DLR priced its initial public offering in October 2004, raising public equity for a real-estate investment trust focused on technology property.² A REIT owns buildings and passes most of its income to shareholders. Its model was straightforward: finance land and shells with equity and debt, lease space and power to tenants that need it, and collect rent for years. That model now has to stretch over AI campuses whose tenants need far more power per square metre than any office.

The last time the pipes outran the traffic

Every infrastructure boom has a ghost, and for this one it is the telecom bust. In the late 1990s, investors and operators bet that internet traffic would need vastly more bandwidth. In a 2003 speech, Federal Reserve governor Ben Bernanke described what followed: companies had laid too much long-haul fibre, and internet start-ups had bought too many servers and routers.³ The demand forecast was not entirely wrong. Internet traffic did keep growing. What failed was the timing, the pricing and the financing. Capacity arrived years before revenue that could pay for it.

Then came WorldCom. On 25 June 2002 the company said it would restate its accounts after improperly capitalising $3.852 billion of line costs, ordinary operating expenses that it had treated as long-lived investment. On 21 July it filed for Chapter 11.⁴ The investigation report the SEC filed describes an accounting fraud, and it also describes an industry already under pressure. Keep the two causes separate. The fraud was a crime. The overbuild was a capacity cycle that would have hurt honest companies too, and it did. What linked them was how quickly credit disappeared once lenders and suppliers lost faith in reported numbers.

None of this shows that AI is repeating 2001. Today's largest spenders are profitable companies, while many telecom borrowers were not. The useful lesson is narrower and harder to wave away: a demand forecast can point in the right direction while the timing, the price per unit and the financing still fail. Fibre laid in 2000 eventually carried the streaming era. Many of the companies that laid it did not survive to see it.

So the earlier boom left a question for this one. If the pipes can be right and the borrowers still go broke, what decides who survives? Part of the answer begins with a chip designed to draw video-game pixels.

↑ Back to Top

Chapter 2: A graphics chip learns to think

In 2012, three researchers at the University of Toronto, Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, trained a deep neural network on a large image dataset and described it in a paper that came to be known by Krizhevsky's name: AlexNet. They trained it on graphics processors.⁵ A chip built to shade millions of pixels at once turned out to be very good at a different job: doing huge numbers of similar mathematical operations side by side.

Hinton was the senior figure, a researcher who had kept faith with neural networks through decades when much of the field had moved on. Krizhevsky wrote the GPU implementation that made training practical. Sutskever would go on to help found OpenAI. Their result did not make money that year. It changed what researchers believed was possible, and that changed what they wanted to buy.

A crew instead of a specialist

A GPU is like a large crew of workers all doing the same simple task at once, such as painting identical fence panels. A CPU, the conventional processor, is more like a small team of skilled generalists working through varied tasks one after another. Training a neural network means multiplying enormous grids of numbers, which is fence-painting on a heroic scale, so the crew wins.

The analogy stops at coordination. A real AI workload is not only arithmetic. The chips must fetch data from memory fast enough, pass results to thousands of other chips across a network, and run software that keeps the whole crew in step. That is why the AI build-out ended up spending heavily on memory, networking, optical links and circuit boards as well as on processors, and why the supply chain in Chapter 6 is so long.

Nvidia's long bet

The crew only became useful outside graphics because Nvidia $NVDA made it programmable. In November 2006 the company released CUDA, a platform that let developers use GPU parallel processing for general-purpose computing.⁶ Under chief executive Jensen Huang, Nvidia spent years supporting software for a market that barely existed. When deep learning arrived, the tools were already there, and researchers had already learned them.

In April 2016 Nvidia went a step further and launched the DGX-1, a ready-made deep-learning system that combined GPUs, software and tools in one box.⁷ That was a quiet but important shift. Nvidia was now selling a working system that a lab could plug in, as well as chips. The more of the system it supplied, the more of the customer's spending it captured, and the harder it became for a customer to swap one part for a rival's.

The architecture that hungered for compute

In June 2017 a team of Google researchers published "Attention Is All You Need", introducing the Transformer architecture.⁸ Transformers made it practical to train language models on very large datasets, because they process many parts of a text in parallel, which suits the GPU crew. The paper's headline translation model trained on eight GPUs for three and a half days. That is evidence that computing power shaped what a model could do. It is not a forecast of modern costs: today's commercial clusters are many orders of magnitude larger, and the paper's set-up cannot be scaled up into a modern budget.

Then the research became a product. On 30 November 2022 OpenAI released ChatGPT as a public conversational service.⁹ OpenAI's chief executive, Sam Altman, had become the most visible advocate of the idea that more computing would yield more capable models. Microsoft's chief executive, Satya Nadella, had already tied his company's cloud to that idea. In 2019 Microsoft announced a $1 billion investment in OpenAI and a partnership to build Azure supercomputing capacity for it.¹⁰

Training and inference

Two words need separating here, because the economics differ. Training is building a model: a large, lumpy, one-off use of computing, like constructing a factory. Inference is using the model to answer a request: a recurring service, like running the factory every day. Training creates the capability. Only inference can create a lasting revenue stream, and only if the price each answer earns covers the electricity, chips and staff behind it.

ChatGPT's reach showed that people wanted the product. It did not show that every GPU bought to serve such products would earn a return. Interest is a demand signal. Revenue per answer, minus cost per answer, is a business. The gap between the two would soon be filled with borrowed money, and the model for renting out computing at all came from a bookseller.

↑ Back to Top

Chapter 3: From cloud rent to GPU-backed debt

On 14 March 2006, Amazon Web Services launched S3, a service that let developers store data online and pay for what they used instead of buying their own storage hardware.¹¹ Later that year AWS released EC2, which rented out computing capacity in the same metered way.¹² The customer paid by the hour or the gigabyte. Amazon $AMZN paid for the physical fleet in advance, before most of its future customers had even heard of it.

That inversion was the cloud's financial invention. Thousands of IT departments had each bought servers they did not fully use. Now a few platforms with deep pockets and many customers bought the hardware once and spread the cost. A diversified customer base made the bet safer: if one tenant left, a hundred others remained. Microsoft $MSFT with Azure, Alphabet $GOOGL with Google Cloud, and later Oracle $ORCL followed with their own clouds.

The lab that needed a new kind of company

Training frontier models soon became too expensive for a pure non-profit. In March 2019 OpenAI created a "capped-profit" structure so that it could raise outside capital while keeping its non-profit governance.¹³ Microsoft's investment that year, and the much larger extension in January 2023, tied the lab's computing to a hyperscaler's balance sheet. A model lab burns cash on computing long before it earns much, and the only parties able to fund that at scale were the clouds themselves and investors betting on the outcome.

From mining coins to renting GPUs

Meanwhile a different kind of company was getting ready. CoreWeave was founded in 2017 and began by using GPUs to mine cryptocurrency. It launched a cloud platform in 2020 and turned toward AI workloads.¹⁴ Its own annual report is frank about how new this business is. Before 2022 its revenue was limited and came mostly from crypto offerings it later discontinued, and it has a short operating history at its current scale.¹⁴ CoreWeave has traded on Nasdaq since 28 March 2025.¹⁴

The crypto origin cuts both ways. Mining taught operators to find cheap power, pack hot chips densely and run them hard. Several miners, including Core Scientific $CORZ, IREN $IREN, Hut 8 $HUT and TeraWulf, later tried the same conversion. But crypto also taught a lesson in volatility: revenue tied to a coin price can vanish overnight, and sites designed for mining rigs are not automatically fit for AI clusters that need more cooling, denser networking and stricter reliability.

Why specialised clouds borrowed

Why would anyone rent GPUs from a young specialist instead of from Amazon or Microsoft? Mostly speed and access. When accelerators were scarce, a focused operator that could get hardware, power and a building ready quickly had something to sell. Its weakness was the mirror image of the hyperscalers' strength. It had fewer customers, no advertising or software business to lean on, and no stockpile of retained cash. To buy GPUs before the receipts arrived, it had to borrow against something.

That something became the GPUs themselves, along with the customer contracts they served. This is asset-backed lending: credit secured by specific equipment or cash flows rather than by the borrower's general creditworthiness. Think of it as a mortgage on a machine. The analogy fails at resale. A house in a decent neighbourhood is fairly easy to value and sell. A used GPU's value depends on whether software still supports it, whether there is a powered data centre to plug it into, whether anyone still wants that generation, and how much faster the next one is. A lender repossessing a house gets a house. A lender repossessing a rack of GPUs gets a question.

The cloud model had shown that customers would rent computing instead of owning it. It had not shown that a single-purpose provider could finance a fleet that ages this fast. That question arrived suddenly in 2023, when AI demand turned from curiosity into a scramble.

↑ Back to Top

Chapter 4: The sale of the future

On 23 January 2023, Microsoft announced a multiyear, multibillion-dollar extension of its partnership with OpenAI.¹⁵ Look at what it bound together. Microsoft was OpenAI's investor. OpenAI was a huge consumer of computing. Microsoft sold the cloud infrastructure that did the computing. Money flowed from Microsoft to OpenAI and, in large part, back to Microsoft as cloud usage. The same company sat on three sides of the transaction.

That is not sinister in itself. When a market is too young for arm's-length contracts, strategic partners often coordinate investment that no outsider would finance. But it changes how outside observers should read the demand figures. When a cloud's biggest customer is also its investee, the customer's spending is partly the investor's own money coming back.

Everyone reserves capacity at once

After ChatGPT, the threads began to pull each other along. Hyperscalers raced to build or reserve capacity. Independent GPU clouds signed customer contracts and took them to lenders. Private-credit managers, which lend outside the banking system, and infrastructure funds offered loans and joint ventures. Model labs other than OpenAI, among them Anthropic and xAI, raised large private sums. Their financial results are not filed with regulators in a standard form, so the public sees them mostly through funding announcements and media reports.¹⁶

The word that began to dominate earnings calls was backlog. A backlog is like a factory's order book: work promised but not yet done. The analogy fails on the fine print. One company's backlog may hold firm, non-cancellable contracts. Another's may hold framework agreements, options, or orders that depend on the provider delivering powered capacity by a certain date. Accounting gives one stricter measure, remaining performance obligations or RPO, which covers contracted revenue not yet recognised. Even RPO is not cash, and for a diversified cloud it includes far more than AI.

Four stages that announcements blur

The headlines of this period blurred four different things: capital sought, capital committed, money actually drawn or spent, and capacity that is switched on and earning cash. Each is a long way from the next.

In September 2024, BlackRock $BLK, Global Infrastructure Partners, Microsoft and the Abu Dhabi investor MGX launched the Global AI Infrastructure Investment Partnership. It said it would seek $30 billion of private equity capital over time, which could mobilise up to $100 billion of total investment once debt was included.¹⁷ That is a fundraising goal with an assumed ratio of borrowing on top. It is not money deployed.

On 21 January 2025, OpenAI announced Stargate, a company with SoftBank, Oracle and others that intended to invest $500 billion over four years, starting with $100 billion.¹⁸ Later that year OpenAI and Oracle announced an additional 4.5 gigawatts of capacity under the partnership.¹⁹ By 2026, OpenAI said Stargate had passed its original 10-gigawatt milestone, adding more than 3 gigawatts in 90 days, and it acknowledged that financing and partnership structures could change.²⁰ These are company statements of intent and progress. "Secured" capacity in a press release is not the same as energised capacity under a binding payment contract, and none of the announcements disclosed the full project financing behind them.

Why does the distinction matter so much? Because each stage carries a different risk. A target can be abandoned at no cost. A commitment can create fixed obligations. Spent money becomes assets that must be paid off. Only energised, paid-for capacity produces the cash that pays the lenders. The ripple running through the chain, from chip orders to power contracts, starts at the earliest stage, long before anyone knows whether the last stage will arrive.

Strategic money can lower one risk, that the pieces never get built, while raising another, that demand is less independent than it looks. To see where that second risk ends up, look inside the balance sheet of a company that sits between the money and the machines.

↑ Back to Top

Chapter 5: The balance sheet in the middle

On 30 July 2023, CoreWeave entered a $2.3 billion delayed-draw term loan facility to finance the GPU servers and infrastructure needed to perform its customer contracts.²¹ In May 2024 it announced a much larger facility, $7.5 billion, led by Blackstone $BX and the hedge fund Magnetar, with other institutional lenders taking part.²² The structure was simple to describe and hard to run. Lenders advanced money against GPUs and pledged contract cash flows. CoreWeave had to install the hardware, keep it running and collect from customers fast enough to pay the lenders back.

The stack, from safest to sharpest

It helps to walk through the whole financing stack, because each layer carries risk differently.

At the top, the hyperscalers mostly pay from operating cash and corporate bonds. Their buildings, chips and power contracts are backed by everything the company earns: advertising, software, retail, existing cloud. If one AI cluster disappoints, bondholders still have the rest of Microsoft.

Below that, property and power projects often use joint ventures and project equity. An infrastructure fund owns a share of a campus, collects rent and hopes to sell at a profit. Construction loans and long leases fill the gaps.

Then come the specialised clouds, with delayed-draw loans and GPU-backed facilities. A delayed-draw term loan is a committed credit line that the borrower taps in stages as equipment arrives or construction milestones are met, like a builder drawing on a construction loan floor by floor. It matches borrowing to spending, but undrawn amounts often come with conditions, so "available" is not always "available tomorrow".

Running alongside all of these are supplier commitments, deposits and customer prepayments, which move cash forwards or backwards along the chain. At the bottom sit the model labs, funded by venture and strategic equity, which have to repay nothing if they fail and therefore expose their investors to the full loss.

The key concept for lenders is residual value: what the collateral is worth if the borrower defaults.

A five-year loan on three-year contracts

In August 2026 CoreWeave disclosed a new facility, DDTL 5.5, of $2.6 billion, secured by infrastructure. It has a maturity of about five years, while the customer contracts underneath average about three years.²³ Put plainly, the lenders will still be owed money after the average contract has expired. They are betting that customers renew, or that the capacity can be rented to someone else, at prices that keep paying down the debt.

The company presents this as a deliberate choice. Shorter contracts can command higher prices, and lenders accepting renewal risk opens credit to a wider range of customers.²³ That is a reasonable argument, and the lenders' willingness is evidence that sophisticated money accepts it. It is not proof that the structure is safe through a full hardware cycle, because no such cycle has yet been tested with defaults and recoveries at this scale. The record leaves the bet standing but unproven. What would settle it is renewals at stable prices when today's chips are two generations old.

What the cash-flow statement says

CoreWeave's quarterly report for the six months to 30 June 2026 gives the clearest picture of a borrower in the middle. It spent $14.9 billion on investing activities, mainly building and equipping capacity, and raised $14.0 billion from financing.²⁴ Its funding mix included debt, equity, delayed-draw facilities and financing arranged through equipment and software vendors. The company said its cash, short-term investments and available borrowing should cover obligations due within a year, and it also said it expects to need more debt and equity in future.²⁴ CoreWeave had $13.6 billion drawn under delayed-draw facilities, $16.6 billion of notes, $16.3 billion of operating lease liabilities and about $10 billion of undrawn availability across specified facilities.²⁴

Read those numbers as a liquidity picture, not a solvency verdict. Liquidity asks whether the bills can be paid this year. Solvency asks whether the assets will ultimately earn enough to repay everything. A company can pass the first test and fail the second, or the reverse. What the figures do show is dependence. Each new block of capacity has so far needed fresh outside capital, and that will stay true until operating cash catches up with investment.

Other borrowers, other lenders

CoreWeave is the most visible case but not the only one. Lambda, a privately held GPU cloud, completed several large secured financings in 2026: a $1 billion senior secured financing, a $926 million syndicated term loan and a separate $1 billion secured facility.²⁵ Its public materials do not fully disclose customer quality or collateral terms. Crusoe, which builds AI campuses, announced a private Series F round in September 2026 at a reported $30.9 billion post-money valuation, and said it had more than $140 billion of total contracted value across its platform.²⁶ Both figures come from the company. Contracted value may combine very different kinds of commitment, and a private round valuation, often carrying preferred rights, cannot be compared directly with a listed market value. In Europe, Rumble completed its acquisition of the GPU cloud Northern Data in June 2026, taking about 85.2% ownership and an estate it put at roughly 22,000 Nvidia GPUs.²⁷

The supplier that also supports demand

The most interesting party in the middle may be the one that sells the chips. Nvidia has commercial arrangements and commitments that expose it to counterparty and execution risk, including the possibility that it buys capacity from AI clouds that cannot sell what they committed to.²⁸ That does not make Nvidia a lender. It does blur the line between supplier and supporter of demand. A customer whose unsold capacity might be taken up by its own supplier is a stronger borrower for it, and the supplier carries some of the risk that its own sales figures seem to have passed on.

The debt, then, sits on top of a chain. To see what it is actually paying for, follow one rack of servers from a chip order to a power bill.

↑ Back to Top

Chapter 6: The physical chain, from a chip order to a power bill

Picture a single rack of AI servers on its way to work. At every stage someone has to pay before the end customer does.

From design to rack

It starts with an accelerator design, usually from Nvidia, sometimes from AMD $AMD, or a custom chip designed with help from Broadcom $AVGO or Marvell $MRVL for a hyperscaler that wants to rely less on Nvidia. The design goes to a foundry, almost always 台積電 TSMC $2330.TW, which manufactures it on the most advanced processes. The chip then needs high-bandwidth memory, stacks of memory placed right next to the processor so data can move fast enough, from SK hynix (SK하이닉스) $000660.KS or Micron $MU. Advanced packaging joins chip and memory, a step where TSMC and assembly-and-test specialists such as ASE Technology $3711.TW sit.

Next come boards and assembly. Multi-layer printed circuit boards and chip-package substrates come from Taiwanese specialists including Elite Material $2383.TW, Gold Circuit Electronics $2368.TW, Unimicron $3037.TW and Nan Ya PCB $8046.TW. Optical components that link racks come from contract makers such as Fabrinet $FN. Contract manufacturers including Quanta $2382.TW, Hon Hai (Foxconn) $2317.TW, Wiwynn $6669.TW, Wistron $3231.TW and Celestica $CLS assemble the servers, as do branded system makers Dell $DELL and Super Micro $SMCI.

At each step, a supplier may enter a purchase commitment: a reservation for scarce production, like booking a place in a queue. The reservation secures supply, but it can create a future cash obligation even if the buyer's product mix or demand changes. Nvidia reported $279 billion of supplier purchase commitments, Broadcom $127 billion and Super Micro $34.2 billion.²⁹ Those are promises to buy from suppliers, not costs already incurred. They pass through procurement, inventory and cash over the following quarters.

That is a separate risk from operating costs, and Nvidia has already shown how it can bite. In fiscal 2026 the company took a $4.5 billion charge for inventory and purchase obligations related to its H20 chip after US export restrictions changed demand.³⁰ Nvidia could absorb it. A thin-margin assembler holding the same kind of stranded order could not do so as easily.

From rack to building

The finished rack goes into a data centre. The building might belong to a landlord such as Digital Realty, Equinix $EQIX, Iron Mountain's data-centre arm $IRM, CoreSite (owned by American Tower $AMT), Goodman Group's data-centre platform, GDS $GDS in China or Keppel DC REIT in Singapore. It might be a converted mining site run by Core Scientific. Inside, electrical gear and cooling come from Vertiv $VRT, Eaton $ETN, Schneider Electric $SU.PA, nVent $NVT, Legrand $LR.PA and the data-centre cooling business of Modine $MOD. AI racks run so hot that liquid cooling has moved from novelty to necessity, which is part of why these vendors' order books have grown.

From building to grid

Then the building needs power, and here the timetable changes. The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, roughly 1.5% of world consumption, and in its base case projects around 945 TWh by 2030.³¹ A later IEA update puts 2025 at about 485 TWh and 2030 at about 950 TWh. These are different forecast vintages, so they should not be averaged.³² The IEA also stresses that a data centre can be built in two or three years, while the energy infrastructure behind it takes longer to plan and build.³¹

That gap is the financing problem in physical form. A campus can be finished, with its loans accruing interest, while it waits in an interconnection queue, the list of projects waiting for grid studies and upgrades. A place in the queue is not power. A letter of intent with a utility is not a binding contract. And rising electricity use, though a sign of real physical demand, says nothing about whether the computing it powers is profitable.

The power companies themselves are a mixed group. Regulated utilities such as Dominion $D (through Dominion Energy Virginia), Southern $SO (through Georgia Power), American Electric Power $AEP, Entergy $ETR and NextEra $NEE earn allowed returns on approved investment. Entergy is unusual in disclosing $8.6 billion of contracted power revenue backlog as of mid-2026.²⁹ Constellation $CEG, Talen $TLN and Vistra $VST own merchant generation, including nuclear plants that run almost continuously. Constellation's fleet ran at a 93% capacity factor in the quarter to June 2026.²⁹ Equipment makers such as GE Vernova $GEV and Siemens Energy's grid business $ENR.DE sell turbines and transformers. Further out are fuel suppliers Cameco and Centrus $LEU, reactor-component maker BWX Technologies $BWXT, and small-reactor developers Oklo $OKLO and NuScale $SMR. Oklo reported no revenue for 2025, and NuScale's revenue was $31 million.²⁹ For them, AI is optionality on future deployment, not current income. PSEG $PEG owns nuclear plants that are relevant here, though it discloses no comparable figures for that unit.

From power to customer

Finally the powered rack is rented, by a hyperscaler's cloud or a GPU cloud, to a model lab, an enterprise or an internal team. Only now does revenue start. If the contract is long, firm and with a creditworthy customer, the chain's cash comes back. If it is short, cancellable or with a customer that is itself burning cash, the chain is betting on renewal.

Where does the money pile up along the way? The reported numbers give a clear answer for the listed companies. Combined operating margin for the chip-and-server layer was 47.2% in the quarter to June 2026. Underneath that, Nvidia's operating margin for its last fiscal year was 60.4%, while Quanta's was 4.1%, Wistron's 3.6% and Wiwynn's 6.7%.²⁹ The scarce parts, accelerators and memory, capture the profit. The assemblers capture the volume and the working capital. Facilities grew more slowly, at 15.4% for the layer in the latest quarter, at margins near 19%.²⁹ Many of these are company-wide figures. Eaton, Schneider and Legrand sell far more than data-centre gear, and most of these companies do not disclose how much of their revenue comes from AI. Where a company discloses nothing, that is the honest answer.

The chain, then, has winners on paper. Whether the leaders can keep their lead, and whether the borrowers can carry theirs, is the contest that follows.

↑ Back to Top

Chapter 7: The contest between scale and strain

Two numbers frame the present. In the quarter to June 2026, Empor's specialised chip and server layer grew revenue 99.9% at a combined operating margin of 47.2%. The specialised GPU-cloud layer grew faster, at 168.0%, with a combined operating margin of −114.8%.²⁹ Both layers sell the same AI boom. One turns it into profit. The other turns it into assets that it hopes will become profit. The deciding question in every contest below is the same: how much of the orders and backlog becomes collected cash after power, replacement hardware, leases and interest.

Designers and bottlenecks against assemblers

Nvidia leads on one specific thing: profit captured per dollar of AI hardware. Its return on capital employed was 74.7% in its last fiscal year, and its operating margin in the quarter to July 2026 was 66.2%, up from a year earlier.²⁹ The lead rests on CUDA and the habits of a decade of developers, on complete systems rather than loose chips, and on priority access to TSMC's manufacturing and packaging. TSMC, the nearest bottleneck, is widening its own margins: 60.3% operating margin in the latest quarter, more than ten points higher than a year before. Broadcom's rose to 54.3%, helped by custom chips for hyperscalers.²⁹

The strongest evidence against a lasting Nvidia lead is that its biggest customers are building alternatives. Google, Amazon and Meta $META all design their own accelerators, and Broadcom's sharp revenue rise, 85.5% in its latest quarter, shows that the custom route is getting real orders.²⁹ Leaders in chips have lost leads before when customers decided to make their own. For now the record narrows the claim without rejecting it. Nvidia leads in merchant accelerators, and the share of work that moves to custom silicon is the open question.

Memory has had the most dramatic turn. Micron's revenue for the fiscal year to September 2026 rose 256% to $133 billion, with an operating margin of 75%. SK hynix's latest quarter showed revenue up 257% and a 76% operating margin.²⁹ Memory has always been the industry's most cyclical business, and its record of booms followed by gluts is the main reason to treat these margins as a peak to test rather than a new floor.

The assemblers tell the other side. Wistron's revenue more than doubled in 2025, yet its free cash flow margin was negative, and Super Micro's free cash flow margin in its latest fiscal year was −17.8%.²⁹ When an assembler grows fast it must buy components and carry inventory and receivables before customers pay, so growth consumes cash. A slowdown would release that cash, but only if the inventory can be sold.

The printed-circuit-board makers sit in between, and their margins are widening. Elite Material's operating margin reached 26.9% in the latest quarter. The market has noticed: its market value rose more than five-fold in a year.²⁹ That is the clearest example in the theme of a point worth keeping in mind. A layer can be right about demand while the price already assumes it.

Cash-rich clouds against a borrower

Among the hyperscalers, the contest is about who can pay for the build-out from their own pockets. Microsoft, Alphabet and Meta generated free cash flow margins of roughly 18–23% in their latest fiscal years, with net debt below one times EBITDA, and Amazon's was positive but thin at 1.1%.³³ Oracle is the outlier. In its fiscal year to May 2026, capital spending equalled 82.6% of revenue, free cash flow margin was −35.2%, and net debt was 3.7 times EBITDA.³³ Oracle has published a plan to raise equity and debt in calendar 2026, which describes intended issuance rather than money already raised.³⁴

Oracle's case rests on its contracts: $664 billion of remaining performance obligations as of August 2026.²⁹ That is enormous for a company with $67 billion of annual revenue. But it is not all AI, it is not cash, and much of it is reported to depend on a small number of very large customers whose own funding comes from outside investors. The stronger the RPO, the more capacity Oracle must build before collecting it. A long order book is visibility, and it is also an obligation. Microsoft's RPO, at $678 billion, is backed by a company that generated $67 billion of free cash flow on Empor's figures.²⁹ The two numbers look alike, and they carry very different risks.

The hyperscalers' combined capital spending reached $182 billion in the quarter to June 2026, 87% more than a year earlier.²⁹ Their free cash flow margins have fallen as a result. Alphabet's dropped from 18.2% for 2025 to 9.5% on a trailing basis by June 2026.²⁹ Even the strongest balance sheets are now spending a much larger share of their income on the build-out.

Borrowers against landlords and toolmakers

For the listed GPU clouds, capital spending ran at between about 2.8 and 7.7 times revenue in their latest fiscal years, and free cash flow margins were deeply negative.²⁹ The median company in the layer increased its diluted share count by 180.5% over three years, the clearest sign of how much of the build-out shareholders have funded.²⁹ Small revenue bases distort ratios like these, so treat them as an index of dependence on outside money, not as a solvency verdict. Nebius, the fastest-growing of them, increased revenue more than four-fold in its latest quarter. IREN's revenue fell 23%, as it moves from mining to AI.²⁹

For this group, share-price swings are a poor guide to risk. What matters is cash plus undrawn committed facilities against the next twelve months of obligations; debt maturities against customer receipts; how concentrated the customers are; whether contracts can be cancelled; and what the GPUs would fetch in a sale. Empor does not have complete maturity schedules or utilisation figures for these companies, so no defensible solvency ranking is possible. CoreWeave leads on one measurable thing, the scale and sophistication of its disclosed financing. Whether that makes it safer or simply larger is unproven.

The equipment makers carry less of this risk. Vertiv's EBITDA margin rose to about 23% in mid-2026, and it spends only around 2–4% of revenue on capital investment, so it is paid for products without owning the campuses.²⁹ The landlords sit in the middle. Equinix spent nearly half its revenue on capital investment in 2025, and its returns on capital are in single digits.²⁹ They own long-lived buildings and power rights, which are valuable if tenants stay.

Money managers

Then there are the lenders. Blackstone, Blue Owl $OWL, Brookfield $BAM, Ares $ARES, BlackRock, Apollo $APO, KKR $KKR, DigitalBridge $DBRG and Carlyle $CG all have digital-infrastructure businesses. Their published financials cover their entire firms. None of them discloses realised returns or losses on AI-related loans in a comparable form, and assets under management are not the same as credit exposure. The market seems wary all the same. The layer's share price return over the past year was −19.8%, and Blue Owl's market value fell by more than half in the year to June 2026.²⁹ Prices cannot say whether that wariness concerns AI loans, private credit generally, or something else. The managers mostly earn fees, while their fund investors bear the credit risk.

One more caution. In their latest results, 15 of 17 chip and server companies beat analysts' revenue forecasts, as did all five hyperscalers.²⁹ Beating a forecast shows execution. It does not show that the cash will come back. And it does not settle the question that briefly shook the whole market in early 2025: what if computing got much cheaper?

↑ Back to Top

Chapter 8: When cheaper compute helps—and when it hurts

In January 2025 the Chinese lab 深度求索 DeepSeek published a paper on its R1 reasoning model.³⁵ Reports that it had built competitive models at a fraction of the usual cost prompted investors to question whether the hyperscalers' spending plans and AI service pricing still made sense.³⁶ The reported training cost was attributed and disputed. It left out wider research spending, and the full cost base was never independently verified.³⁶ But the question it raised was the right one. If each AI task needs far less computing, does the world need less infrastructure, or will cheaper answers mean far more of them?

The optimistic chain

In the optimistic version, each link holds. AI moves from pilots into repeated, paid work. Falling inference costs widen the uses that make economic sense. Utilisation rises in the clusters already built, and collections follow. Contracts become longer and firmer, which lets lenders accept lower spreads. Projects get power on schedule, and the hardware earns its cost before it is retired. In that world cheaper computing is fuel, the way cheaper bandwidth eventually filled the fibre left over from 2001.

There is evidence for the first link. Stanford's AI Index reports that 88% of surveyed organisations used AI in at least one function in 2025 and 70% used generative AI.³⁷ The same report notes that deployment of AI agents, the multi-step automated tasks that would generate heavy recurring inference, remained in single digits across nearly all business functions.³⁷ Using AI somewhere is a long way from paying for it at scale.

The pessimistic chain

In the pessimistic version, efficiency outpaces use. Models need fewer chips per task faster than paid demand grows. Rental prices for GPU time fall. Workloads stay stuck in pilots. Labs need more equity, or more support from their strategic partners, to keep paying for computing. Contracts expire without renewal, lenders cut how much they will advance against a GPU, and borrowers face refinancing or are left with stranded inventory. Custom chips from Google, Amazon, Meta and Broadcom could speed that up by shifting work off merchant GPUs, which would cut the resale value of the collateral already pledged.

The record so far fits the optimistic chain better. Hyperscaler capital spending accelerated after DeepSeek rather than slowing, as the 87% rise shows. But spending is an input. The pessimistic chain would show up first in prices and renewals, which are mostly not disclosed.

What the links show

Empor's tests of how the layers move together add a useful, if limited, picture. Chip and server revenue tends to move with hyperscaler capital spending, usually within the same quarter, and that pattern holds for 15 of the 17 companies. Super Micro is a notable exception.²⁹ The link to facilities is weaker and later: data-centre landlords such as Digital Realty respond most strongly about a year after hyperscalers spend. The link to power is looser still and arrives around two quarters later, holding for only 7 of 11 companies. Cameco and NextEra move against the expected direction.²⁹

The most revealing result concerns the GPU clouds. Their operating margins tend to fall about a quarter after hyperscaler spending rises, across all five companies tested.²⁹ It is consistent with a worry the independents cannot escape: when the giants build their own capacity, they also compete with the specialists who rent to the same customers. Interest rates, by contrast, show no reliable link to margins in any layer. The private-credit managers' margins did not widen with higher rates as expected, and GPU-cloud net margins did not shrink with them.²⁹ A few years of data moving together is evidence, not proof. Product cycles, acquisitions and accounting changes all move these numbers too.

The lesson of DeepSeek, then, is that efficiency changes who wins more than whether there is a build-out. Financial innovation is not resilience either. An announced loan or partnership can carry a project through construction. It cannot make an unprofitable workload profitable. If it does not become profitable, someone eventually has to take the loss.

↑ Back to Top

Chapter 9: Who is left with the bill?

Go back to the rack from Chapter 6, but put it on a lender's workout desk. The customer's contract has ended. The GPUs still have some market value. The lender now has to find out whether the hardware can be moved, powered, supported by current software and rented again, and at what price.

Who paid, and who loses

The build-out is paid for by many hands. Hyperscalers pay from cash flow and bonds. Private equity and infrastructure funds own projects. Project lenders and GPU-backed credit facilities finance equipment. Landlords sign long leases. Suppliers commit to production, and customers sometimes prepay. Venture investors fund the labs whose demand sits at the end of much of the chain.

If returns stall, losses do not disappear. They follow the contracts. Equity holders lose first: the shareholders diluted to fund GPU clouds, and the venture investors in labs. Lenders come next, in an order set by collateral, guarantees, recourse and contract rights. That is why a five-year loan against three-year contracts matters more than any share price. Suppliers can be left holding inventory and purchase commitments, as Nvidia's H20 charge showed on a small scale. Landlords can be left with powered buildings and no tenants. Utilities can be left with grid investments waiting for recovery through regulated rates, which means through other customers' bills. Customers can be left with capacity they reserved and no longer need, and with a negotiation over who absorbs the difference.

The hyperscalers sit at both ends. They fund their own capacity, and they buy from independent clouds. In a downturn, their large, diverse businesses can absorb losses that would finish a single-purpose borrower. That is the core of the answer: the risk does not spread evenly. It collects where fixed obligations outlast the revenue that was supposed to pay them.

The numbers to watch

Five signals should move before revenue does.

GPU-cloud liquidity against next year's bills. This measures cash plus genuinely available committed facilities against obligations due within twelve months. It moves early because funding dries up before revenue falls. It would settle whether the independents can keep building without emergency capital. The source is quarterly 10-Q filings. The latest reading: CoreWeave said, as of 30 June 2026, that its resources covered the year ahead, with about $10 billion undrawn across specified facilities.²⁴ Shrinking undrawn lines while obligations grow would break the story.

Contract term against debt maturity. This measures how long customer contracts run, and whether they can be cancelled, compared with when the loans fall due. It moves at every new facility, well before defaults. It would settle whether lenders are underwriting contracts or renewals. It is disclosed when facilities close. The latest reading: about three years of contract against five years of debt for DDTL 5.5 in August 2026.²³ A widening gap without more customers would weaken the story.

Receipts and utilisation as capacity comes online. This measures whether newly commissioned capacity is billed and paid for. No comparable public series exists across providers. The nearest proxy is operating cash flow against investing cash flow in quarterly filings. CoreWeave's first half of 2026 showed $14.9 billion invested against $14.0 billion raised.²⁴ A rising share of investment funded by operations would confirm the story.

Supplier commitments against shipments and inventory. This measures whether reserved production turns into delivered, paid-for goods. It moves before revenue because orders are deferred before they are cancelled. It appears in annual and quarterly commitments notes. Nvidia reported $279 billion of commitments in Empor's latest reading.²⁹ Commitments and inventory rising while shipments slow would break the story.

Energised, customer-accepted capacity against announced megawatts. This tracks how far announced gigawatts become working power. It moves before revenue because delays add interest before they cut sales. There is no reliable common series, and Empor's own power and capex tables are too sparse to rely on.²⁹ Utility filings and company project updates are the best available sources. Repeated delays after financing has closed would break the story.

The next readings arrive soon: TSMC reports on 15 October 2026, and Microsoft, Alphabet, Meta and Amazon report at the end of the month, though dates can move.²⁹ None of them will reveal a lender's recovery rate. What they can show is whether the money is still flowing down the chain, and whether more of it is coming back up as cash.

↑ Back to Top

Glossary

An asset-backed loan is credit secured by equipment, property or contract cash flows. Recovery depends on enforceable rights and on what the assets are worth. Capex is money spent on long-lived equipment, buildings and infrastructure. Customer backlog is a company's estimate or disclosure of future work. The contract terms determine how secure it is. A delayed-draw term loan is a committed loan that the borrower can draw later, often as construction or equipment spending proceeds. A GPU is a processor that performs many similar calculations in parallel. It is useful for AI, but it is not a complete computing system by itself. Inference is running a trained model to produce an answer or other output. An interconnection queue holds projects waiting for review and connection to the grid. A queue request is not operating power. Residual value is what an asset may be worth after use, or in a default. RPO is contracted revenue not yet recognised under accounting rules. It is not cash already collected. A take-or-pay contract requires a minimum payment even if the customer uses less than the capacity it reserved. Training is the process of fitting a model using data and computing power. Working capital is cash tied up in inventory and unpaid customer invoices, net of credit from suppliers.

References

1. Order No. 888 — Federal Energy Regulatory Commission, 1996 ↩

2. Investor FAQs — Digital Realty ↩

3. Will business investment bounce back? — Federal Reserve Board, 2003 ↩

4. WorldCom investigation report — SEC filing, 2003 ↩

5. ImageNet Classification with Deep Convolutional Neural Networks — Krizhevsky, Sutskever and Hinton, 2012 ↩

6. CUDA Programming Guide: Introduction — NVIDIA ↩

7. NVIDIA launches world's first deep learning supercomputer — NVIDIA, 2016 ↩

8. Attention Is All You Need — Vaswani et al., arXiv, 2017 ↩

9. Introducing ChatGPT — OpenAI, 2022 ↩

10. Microsoft invests in and partners with OpenAI — OpenAI, 2019 ↩

11. Amazon Web Services launches S3 — Amazon, 2006 ↩

12. Our Origins — Amazon Web Services ↩

13. OpenAI LP — OpenAI, 2019 ↩

14. CoreWeave Form 10-K for fiscal 2025 — SEC ↩↩↩

15. Microsoft and OpenAI extend partnership — Microsoft, 2023 ↩

16. Company scorecard and trend tables — Empor, 12 Oct 2026 ↩

17. BlackRock, Global Infrastructure Partners, Microsoft and MGX launch new AI partnership — BlackRock, 2024 ↩

18. Announcing the Stargate Project — OpenAI, 2025 ↩

19. Stargate advances with partnership with Oracle — OpenAI, 2025 ↩

20. Building the compute infrastructure for the Intelligence Age — OpenAI ↩

21. CoreWeave SEC filing describing DDTL 1.0 — SEC, 2026 ↩

22. CoreWeave secures $7.5 billion debt financing facility led by Blackstone and Magnetar — CoreWeave, 2024 ↩

23. CoreWeave DDTL 5.5 exhibit — SEC, August 2026 ↩↩↩

24. CoreWeave Form 10-Q for the quarter ended 30 June 2026 — SEC ↩↩↩↩↩

25. Investor information — Lambda ↩

26. Crusoe announces Series F funding — Crusoe, September 2026 ↩

27. Rumble closes acquisition of Northern Data — GlobeNewswire, 17 June 2026 ↩

28. NVIDIA Form 10-Q for Q2 fiscal 2027 — NVIDIA ↩

29. Theme pulse, scorecard, trends, links and results tables — Empor, 12 Oct 2026 ↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩

30. NVIDIA Form 10-K for fiscal 2026 — SEC ↩

31. Energy and AI: Energy demand from AI — International Energy Agency, 2025 ↩↩

32. Key Questions on Energy and AI: Executive summary — International Energy Agency, 2026 ↩

33. Hyperscaler capital intensity and debt sustainability — Empor, 12 Oct 2026 ↩↩

34. Oracle announces equity and debt financing plan for calendar year 2026 — Oracle, 2026 ↩

35. DeepSeek-R1 — DeepSeek-AI, arXiv, 2025 ↩

36. Big Tech faces heat as China's DeepSeek sows doubts on billion-dollar spending — Reuters via Investing.com, 27 January 2025 ↩↩

37. 2026 AI Index Report: Economy — Stanford HAI, 2026 ↩↩