Three tickers moved together this week. Nvidia. Broadcom. CoreWeave. The wire copy attributed the move to rising bond yields and "AI debt concerns."
The co-movement is the anomaly worth studying. A fabless chip designer with net cash on the balance sheet, an ASIC-and-networking franchise with contracted hyperscaler revenue, and a leveraged GPU landlord whose customer book you could count on one hand do not share a duration. They do not share a credit profile. They barely share an asset class in any rigorous sense. Yet the tape priced all three as a single instrument.
Markets behave this way when they are pricing a narrative rather than a cash flow. Narratives have no amortization schedule. Balance sheets do.
So I built the schedule. A simplified GPU-backed facility: six-year tenor, level amortization, 70% advance rate against appraised collateral, revenue indexed to prevailing GPU-hour pricing at 85% utilization. At inception, cash-flow debt service coverage lands near 1.3x — thin, but defensible underwriting. Then I changed exactly one input, the assumed economic life of the collateral, from five years to three, and left every cash flow untouched. Collateral coverage at the end of year three fell from 2.4x to 0.2x.
Same silicon. Same contracts. Same coupon. Different accounting assumption.
That assumption is the least examined number in the AI capital stack. And as of this quarter, it has an on-chain representation.
The three names are not the same trade
Nvidia designs accelerators and sells them. Gross margin north of 70%, negative working capital, effectively no debt. Its sensitivity to the long end of the curve is a pure equity-duration effect: a high-multiple growth stock is a long-dated cash flow stream, and when the risk-free rate moves, the denominator in every DCF moves with it. Nothing about Nvidia's operating business changes when the ten-year reprices. Its multiple does.
Broadcom is a different animal wearing the same label. Custom ASIC programs for hyperscalers, switching silicon, and a software line. Its revenue is contracted against customer capital budgets, and — this is the part the headline missed — it is structurally long the trend that pressures the neoclouds. Every hyperscaler that decides to build its own accelerator rather than rent third-party GPU capacity is a Broadcom design win. Tightening credit for GPU landlords pushes workloads back toward owned infrastructure. Broadcom's fundamental exposure and CoreWeave's fundamental exposure point in opposite directions.
CoreWeave is the actual subject of the story. A GPU cloud — a "neocloud" — buys accelerators, finances them with debt, and rents them by the hour. The asset is depreciating, the debt is amortizing, and the spread between rental revenue and debt service is the entire business model. That spread is a function of three variables: the interest rate on the debt, the utilization rate on the fleet, and the rental rate the market will bear. Two of those three are now moving against it simultaneously.
The transmission channels are cleanly separable, and conflating them is the analytical error at the center of this coverage. Rising yields hit equity valuations through the discount rate and hit leveraged compute providers through the cost of refinancing. Those are different mechanisms operating on different time horizons with different reversibility. A multiple compression is recoverable in a quarter. A refinancing wall is not.
What a GPU-backed facility actually is
Strip the marketing. A neocloud financing is an asset-backed loan where the collateral is a depreciating capital good with no liquid secondary market and no published price index. The lender advances some percentage of an appraised value, the borrower amortizes over a tenor, and the lender's protection is the residual value of the hardware at any point in the schedule.
That residual value is the whole credit. And it is not observable. It is appraised, which means it is modeled, which means it is an assumption wearing the costume of a fact.
Private credit funds have been the marginal buyer of this paper because the yield is attractive and the mark-to-model accounting is forgiving. Asset-backed securitization of GPU receivables has started to appear. The important structural detail is that these instruments are typically marked by the originator or by an affiliated valuation agent, on a monthly or quarterly cadence, against a schedule that was fixed at origination. The mark is a lagging indicator by construction.
That is not fraud. It is how private credit has always worked. But it matters enormously when the underlying asset has a three-year economic half-life and the reporting cadence is quarterly. Twelve reporting periods is the entire life of the collateral.
The rental rate is the true mark, and nothing prices it
Here is the core insight, and it is where the AI debt story diverges from every previous infrastructure financing cycle.
For a data center, the collateral is a building. Buildings depreciate slowly, they have comparable sales, and a lender can underwrite residual value against a century of transaction data. For a GPU cluster, the collateral is a compute-hour production function. Its value is not the silicon. It is the present value of the rental stream the silicon can generate before it is obsoleted by the next generation.
Which means the correct mark for GPU collateral is the spot rental rate per GPU-hour, times expected utilization, times remaining competitive life. Not the invoice price. Not the appraisal. The rental rate.
And the rental rate is not published. There is no GPU-hour index, no CME contract, no benchmark. There are broker quotes, private term sheets, and a handful of public cloud list prices that nobody actually pays. The market has built a multi-hundred-billion-dollar financing stack on top of an asset whose mark is unobservable and whose depreciation curve is a management estimate.
I have seen this exact failure mode before, in a different domain. When I audited the 0x v2 exchange contracts in 2017, the exploitable class of bug was never a broken signature check. It was a non-atomic relationship between two state transitions that the developers believed were simultaneous. The order was valid when it was signed and the fill was valid when it executed, and in between the world moved. Three race conditions in the matching logic, all of them variations on the same theme: an assumption that the price and the settlement share a clock.
The GPU credit stack has the same defect at a higher layer. The collateral's true value updates continuously with the rental market. The collateral's recorded value updates quarterly, on a schedule fixed at origination. The gap between those two clocks is not a rounding error. It is the entire risk.
The on-chain mirror has the same bug, with faster liquidation
Now the part that should concern anyone reading this from a DeFi seat rather than a credit seat.
Tokenized GPU receivables are already being structured. The pattern is familiar: an SPV holds the lease contracts and the hardware, the SPV issues a token representing a claim on the cash flow, and that token is wrapped in an ERC-4626 vault so it can be posted as collateral in a lending market.
The naive NAV function looks like this: