The machine rents high. The product keeps getting cheaper.

Nvidia is negotiating a $500 billion financing plan, and the whole disagreement fits into two numbers. The company has asked lenders to treat its GPUs the way they treat aircraft — as collateral that holds its value for a decade or more. The banks answered with a depreciation schedule of three to four years, plus higher rates and stronger guarantees, and Nvidia has begun rewriting some deals around those demands. ^[1]

AI compute is a productive, durable and fungible asset that can support long-term financing. — Nvidia

The two figures do not look like an argument. They are the argument, set down in a term sheet before any market opened, and everything else this year is that gap moving out into the world. The gap matters because, for a rented-out chip, the value of the collateral and the revenue that pays the debt are the same variable. A GPU is worth what it can earn in rent. Rent is set by how scarce compute is. So a lender who writes "this asset depreciates in four years" and a borrower who writes "it lasts a decade" are not arguing about hardware. They are arguing about how long today's prices hold — which means the underwriting is the assumption about the shortage. The providers' own paper shows the mechanism running in both directions. CoreWeave has drawn roughly $35 billion of asset-level debt against a $104.2 billion contracted backlog, debt secured by revenue that is itself a promise to keep renting compute at today's rates. And the company's filings flag the downside in a single clause: margin compression "if the supply of compute normalizes." ^[2] On the input side, the shortage is real and measured. Global data center supply reached 8.9 gigawatts operational last year against 21.1 gigawatts of demand, a 12-gigawatt deficit. ^[3] The memory every AI chip needs is expected to stay short through 2028, and possibly to 2030. ^[4] These are the numbers that make "a decade of scarcity" feel like a safe bet, and they are the numbers Nvidia is lending against. Then, in one week in May, the other half of the bet moved. On May 23, DeepSeek made its flagship model's 75 percent price cut permanent — on Huawei processors, not Nvidia's — and priced the result 12 to 19 times cheaper than GPT-5.5 or Claude Opus 4.7 for an equivalent task. ^[5]

This is why the price cut is permanent rather than promotional. — Sanchit Vir Gogia

The arithmetic of the pinch is not subtle. Morgan Stanley ran a fully optimized data center on the newest Nvidia chips and found it would cost about $25 billion a year to rent while producing $23 billion of output — and that its 31 percent return on invested capital holds only if the hourly price stays at $8.50. ^[6] Every model priced below that is a model eating the margin the debt was written against. The credit market has begun to vote, in the clipped, dry way credit markets do. By the second quarter, the hyperscalers' combined capital spending had crossed above their combined operating cash flow; the buildout is now funded off the balance sheet rather than the income statement. ^[7] Alphabet posted its first negative free cash flow on record, and Meta's quarterly figure fell 91 percent. ^[8] In Texas, the power futures that price the physical input to all this have already cooled — the forward market for electricity no longer believes the most extreme data center projections. ^[9] The people who underwrite the debt say it plainly. The cost of insuring a hyperscaler's bonds against default has widened about 60 basis points relative to a bank's since October 2025. ^[10] Apollo's chief economist describes what is being repriced in one sentence.

What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets. — Torsten Slok

Now set the bill beside the revenue that has to pay it. OpenAI's widely quoted $50 billion annualized revenue figure turned out, when pressed, to be a projection rather than a number anyone had billed. ^[11][12] Bain reckons the hyperscalers need $4.2 trillion of new revenue within five years to fund the buildout; Goldman and Sequoia put the break-even at $300 billion to $1 trillion a year. ^[13]

The question is whether the applications arrive in time to pay for it — Bain & Company, Inc.

What realized AI revenue looks like at one of the world's largest IT firms is a useful yardstick: Tata Consultancy Services, up 15 percent on profit, reported annualized AI revenue of $3.1 billion — real, billed money, and orders of magnitude smaller than what is being spent on the buildout. ^[14] The bull case needs no fence around it; it is one horn of the same fork, and it sharpens rather than dissolves the contradiction. The shortage half is genuine, and the people holding the chips say so — Nebius reports demand still outstripping supply. ^[15] The hedge fund manager Gavin Baker points to how long the scarcity has already outlived the skeptics.

I don't think anyone in '24 or '25 thought that the prices of old GPUs would still be going vertical in 2026. — Gavin Baker

That is the shortage winning. It is exactly what makes the other number — the product's price — so strange. If compute is this scarce, the output ought to cost more, not less. DeepSeek did not wait for the shortage to clear. The industry's own executives concede the stakes without much prompting. Vesora's Sandra Rivera argues the spending is finally crossing over from training to inference, with enterprise demand unbroken; but she adds the caveat herself.

there is always a period of, you know, irrational exuberance, if you will, and then where things modulate. — Sandra Rivera

The argument is about to acquire instruments. Later this year, the first futures markets on chip-rental prices arrive — Silicon Data with CME Group, Ornn with Intercontinental Exchange, a Bermuda exchange, and ETFs already filed to track the contracts. ^[16][17] For the first time, what a rented GPU will be worth next year becomes a price you can look up rather than a term you negotiate. That is the fork, held open rather than resolved. If rents normalize, the providers' own filings have already conceded what comes next: margin compression. If rents hold, the bill lands on a revenue base that, on today's billed numbers, is still almost entirely a projection. One answer is already written into a depreciation schedule. The other is already written into a price cut. The market simply has not had to choose between them yet.