The collateral competes with itself

Lend against a toll road and the road stays where it is. Nobody builds a wider, cheaper toll road beside it every twelve months, and certainly not the company that sold you the first one. Now lend against a machine whose manufacturer sells it at a seventy-five percent gross margin, sits among the lenders financing the purchase, reportedly guarantees part of its residual value — and decides, once a year, what the machine is worth by releasing a better one.

That is not a thought experiment. On 10 August 2026, Nvidia announced partnerships with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to build dedicated pools targeting five hundred billion dollars of financing for buyers of its chips, with the chips as collateral. Jensen Huang's phrase was that compute is now "an investable infrastructure asset." Three weeks earlier the credit market had pushed Nvidia's own default swaps to a record.¹ Both facts are true, and holding them together is the only honest way to read this.

This essay is not about a bubble. It is about a piece of collateral with an unusual property — it funds its own successor — and about what a lender should ask before treating it like a road.

Equipment finance is old and boring, and it works because the lender, the vendor and the collateral's future value are three different things. Here they are converging on one balance sheet.

The vendor is the lender. Not literally — Nvidia is careful to say it does not do vendor financing, and the Lucent and Nortel ghosts explain why. But it sells the chip at full margin, takes equity in the vehicles that buy it, anchors the consortium that lends against it, and is reported to cover up to a quarter of any shortfall if the chips fetch less than the books assumed. Huang's own answer to the circularity charge is that the scheme exists to bring independent long-term capital in. That is true. It is also exactly what the guarantee is for.

The successor sets the price. A GPU does not wear out; it is out-competed. Depreciation follows the release cadence, not the odometer — and the cadence is now annual: Hopper, Blackwell, Rubin this year, Rubin Ultra next. As Blackwell shipped, H100 on-demand rates fell from around eight dollars an hour at the peak to two, and secondary units from forty thousand dollars to the high twenties.² The first GPU securitisation, in 2024, was underwritten on chips holding half their value at three years. The party that decides whether that holds is the party that sold the chips — by deciding when to ship the next ones.

The loan attacks its own collateral. Residual values are propped up by scarcity — and scarcity is real: reserved H100 rates rose forty percent between October and March as reasoning models ate every available hour, and capacity coming online through September was booked before it existed.² But five hundred billion dollars of financing exists to build supply. If it works, rents soften; if rents soften, residuals fall; if residuals fall, collateral coverage thins. The programme is designed to remove the condition its own underwriting depends on. That is not malice. It is what "solving the compute shortage" means, seen from the collateral's side.

The template was written in August 2023, when a company few people had heard of borrowed $2.3 billion from Magnetar and Blackstone with H100s as security — the first time the chips had been used that way. Less than three years later CoreWeave's debt exceeds $21 billion, from under $8 billion in 2024.³

Read the structure and one clause tells you what the collateral really is. The vehicle pledges the GPUs — and the service contracts running on them, and the credit agreement requires those contracts to be with large, creditworthy counterparties whose payments cover the debt service. In other words, the loan was never secured by metal. It is secured by Microsoft's and OpenAI's willingness to pay, with the metal as a residual tail. That is a sensible structure, and it is also an admission: the lenders did not trust the chip's value either. They trusted the customer.

Which moves the question to the only place it can live: what happens when a contract ends and the chips must be re-let into whatever the market is that year? Right now the answer is reassuring. CoreWeave reports 2022-era H100 capacity rebooking at ninety-five percent of original pricing, and A100s — two generations old — contracted out to 2029 at full rates.⁴ Right now, demand for inference is absorbing every generation that exists. And the market prices the structure's fragility anyway: in the last week of July, CoreWeave's five-year default swaps traded past 855 basis points, while Nvidia's own set a record.¹ The credit market is not saying the chips are worthless. It is saying the re-let assumption is carrying more weight than it has ever carried before.

The bear case has been wrong for a year, and it is worth saying why. Chips have run longer than anyone modelled: Azure kept V100s in service for seven and a half years. Rental prices have not declined; they oscillate — down as Blackwell shipped, up forty percent in the winter squeeze, down again by June — and each swing had a cause other than demand destruction. Memory, not GPU dies, is the binding constraint, which props the price of every generation at once. And the accounting fight is genuinely unsettled: in the same quarter, Amazon shortened the useful life of its accelerators and Meta extended it.⁴

Every one of those points is a statement about demand. Not one is a statement about the collateral. The structure needs demand to exceed supply for the five or six years of the loan, across an annual release cycle, while the financing itself adds supply — and collateral exists for precisely the scenario in which that stops being true. In that scenario the risk is not that one cluster is worth less. It is that every distressed cluster reaches the market in the same quarter, into the same softer demand, all of it the same vintage. GPU collateral is correlated in a way that toll roads are not: they all depreciate on the same date, and the date is chosen in Santa Clara.

There is a sharper version of the obsolescence point, and it does not require demand to fall at all. In a power-constrained data centre the scarce input is not the rack; it is the megawatt. An H100 still works perfectly in 2027. But it occupies power that a Rubin would turn into three or four times the output, so its economic value is set by performance per watt relative to the newest chip, not by whether it runs.⁵ The moment the constraint is energy rather than silicon, every prior generation loses its slot even at full demand. That is the mechanism by which a better chip empties the value out of a working one — and it is the mechanism the residual assumptions have to survive.

Underwrite the re-let, not the metal. The contract is the security; the residual is a bet on the market at the end of it. Stress that bet on the vendor's release cadence, not the borrower's depreciation schedule — the two have been drifting apart for a year, and only one of them is decided by someone with an interest in the answer.

Price the slot, not the unit. A residual per chip assumes the chip is the scarce thing. Where power is the constraint, model the value of the megawatt it occupies against the newest generation's output per watt. That number falls every twelve months whether or not anyone stops renting.

Treat the guarantee as wrong-way risk. A residual-value backstop from the manufacturer pays out in exactly the world where the manufacturer's own demand has softened. It is not credit enhancement in the usual sense. It is a correlated exposure dressed as one — and it should be priced as if it might not be there when it is needed most.

The vendor on three sides. The successor setting the price. The loan funding the supply its own collateral needs not to exist.

None of this says the chips are worthless or that the loans will fail — the evidence today says the opposite, and a market that books A100s to 2029 is not a market in distress. It says something narrower and more durable: this collateral has a competitor, the competitor is the vendor's roadmap, and the roadmap is annual. A toll road does not get a better toll road built beside it every twelve months by the company that financed the first one.

The question for every lender in the five-hundred-billion-dollar pool is not whether the chip holds its value. It is who decides. And the answer, each spring, is the borrower's supplier.