AI Infrastructure's Circular Financing Risk: a Hall of Mirrors With a Mortgage
The AI build-out's newest lenders hold GPUs as collateral, its biggest vendor is renting back its own chips, and its anchor tenant has never turned a profit. Follow the debt, and the demand signal starts to look like a funding signal.
Somewhere in a data hall sits a rack of graphics processors doing double duty as loan collateral. Each chip cost tens of thousands of pounds, its successor will be faster and cheaper per unit of useful work, and a lender has advanced hundreds of millions of dollars against the rack anyway. That rack is the cleanest window into the circular financing risk inside AI infrastructure, because the boom is no longer funded by equity and optimism alone. It is funded by debt secured against hardware whose economics point the wrong way.
How do you lend billions against a chip that gets cheaper every year?
The mechanics borrow from aircraft leasing and then invert them. The debt sits in a special-purpose vehicle, kept at arm's length from the operator's own balance sheet, and is sized against two things: the market value of the GPUs and the cloud contract signed against them. Repayments are scheduled across the silicon's useful life, roughly the gap between one Nvidia generation and the next. CoreWeave wrote the template, raising close to $10 billion of debt across 2023 and 2024 secured by its GPU fleet and the contracts attached to it, as its IPO filing sets out in detail.
Read that filing closely and a quiet truth falls out: the collateral was never really the metal. Lenders sized these loans against contracts with creditworthy tenants, and CoreWeave's prospectus discloses that 62% of its 2024 revenue came from a single customer, Microsoft. GPU lending is tenant lending in disguise. Which raises the uncomfortable question: where does the tenant's creditworthiness come from? Increasingly, from the chipmaker itself.
Why does circular financing make the tenant the collateral?
Trace the newest deals and the loop is brazen. Nvidia intends to invest up to $100 billion in OpenAI, paid out progressively as the lab deploys ten gigawatts of Nvidia systems, according to the companies' own announcement. AMD went a step stranger, pairing a six-gigawatt commitment from OpenAI with a warrant for up to 160 million of its own shares, vesting as the lab buys and deploys AMD silicon, per its regulatory filings: equity flowing out to a customer so the customer can afford the chips. Nvidia, meanwhile, has doubled its cloud spending commitments to $26 billion, including renting back access to its own GPUs from the providers it sells them to. A manufacturer paying to consume its own product is not booking demand. It is booking a mirror.
Every leg of this is legal, and every leg is individually defensible. But when a vendor's cash returns as the vendor's revenue, the revenue line stops measuring what enterprises want and starts measuring the vendor's appetite for writing cheques. Sequoia's David Cahn put the arithmetic plainly in his '$600 billion question': the revenue required to justify the build-out dwarfs anything the industry currently earns. Telecoms equipment makers ran the same experiment in the late 1990s, financing their customers' purchases and reporting the proceeds as insatiable demand. The demand evaporated the moment the financing stopped.
Is AI's demand curve really just a funding curve?
Follow the contracts and concentration does the rest. Oracle's remaining performance obligations hit $455 billion in its most recent reported quarter, up 359% year on year, per its own quarterly filing, after it signed some of the largest cloud contracts in its history. OpenAI has announced a half-trillion-dollar building programme with its infrastructure partners. My own tally, counting only commitments with a public price tag in a filing or a named company announcement, puts a single loss-making laboratory at the centre of well over $600 billion of promised spending; add its nearest rivals, none of them profitable either, and the frontier labs account for the bulk of the announced build-out. Treat that figure as a floor rather than a point estimate, since it excludes every deal without a disclosed number.
The mechanical point matters more than the total. If your anchor tenants lose money, your demand curve is their funding curve. Cloud growth that reads as broad enterprise adoption is, in large part, a handful of labs' ability to keep raising capital. Should the funding climate tighten, the tenant cannot make rent, and the landlord has already mortgaged the building to a lender who thought the building was the safe part.
Why would cheaper inference make the glut worse?
Here is the counterintuitive twist: every efficiency gain tightens the screw on the financing stack. Stanford's AI Index estimates that the cost of querying a GPT-3.5-class model fell more than 280-fold in the two years to October 2024, and each new hardware generation pushes costs lower still. Wonderful for adopters; miserable for a lender whose loan is serviced from revenue, because revenue is volume times price, and the price keeps falling. Volume can grow handsomely while the revenue pool shrinks beneath the debt.
The standard rebuttal is Jevons: cheaper compute gets used more, so demand absorbs the capacity. Probably true, eventually. But 'eventually' is precisely what a leveraged balance sheet cannot wait for. Fibre laid at the turn of the millennium carries traffic today; much of the money that financed it did not survive the wait. Being right about demand and early with the debt is the same trade as being wrong.
What should buyers do while the landlords sweat?
Separate the two risks, because they are not the same. The financing risk is real; the technology risk isn't. Models keep improving whoever holds the debt, so the correct response is positioning, not retreat.
Rent compute, don't buy it. If the build-out overshoots, the next few years become a buyers' market for capacity, and firms locked into long commitments at peak prices will end up funding their competitors' discounts. Spend the saving on the unglamorous work that decides whether AI pays: data quality, governance and use-case discipline. Our guide to getting AI-ready before you build anything covers the sequence, and a clear-eyed technical strategy settles which workloads justify dedicated capacity at all. Where you do deploy agents, keep them under practical human control, so a wobble in your vendor's funding never becomes a wobble in your operations.
Loops like this run longer than sceptics expect, and this one is fuelled by genuinely useful technology. But when the collateral, the tenant and the revenue all trace back to the same few balance sheets, what you have is not a property boom. It is a hall of mirrors with a mortgage, and the surveyor always turns up eventually.
Questions people ask
What is circular financing in the AI industry?
It's any arrangement where a supplier's own money funds purchases of its products: a chipmaker investing in a lab that spends the capital on that maker's GPUs, issuing stock warrants to a customer as a purchase incentive, or renting back capacity built from its own chips. Each transaction is legal and individually defensible, but the loop manufactures apparent demand and flatters reported revenue. Telecoms equipment makers ran a similar scheme in the late 1990s, and the 'demand' disappeared when the financing stopped.
Will the AI build-out end like the dot-com crash?
The closer parallel is fibre optic cable. The technology was real and the capacity was eventually used, but the companies that borrowed to build it didn't survive the wait for traffic. Expect the casualties to concentrate among whoever holds debt against the hardware rather than among the firms using the technology. Overcapacity, if it comes, is a crisis for lenders and a discount for everyone else.
Should businesses delay AI adoption until the market settles?
No, but they should change how they buy. Renting compute rather than owning it keeps you flexible if prices fall, and money saved on capacity is better spent on data quality, governance and staff skills, the things that determine whether AI earns its keep regardless of what happens to the labs anchoring today's build-out.
Related
- The Memory Oligopoly Behind AI's Cost Inflation
- The AI Subsidy Cliff: Your Vendor's Investors Have Been Paying Your Bill
- AI layoffs 2026: the jobs apocalypse is a sales pitch, and the backlash is the real risk
- AI & Automation
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.