Stranded GPUs and the SPV Problem

When S&P Global Ratings put a rating on CoreWeave's $1.5 billion of senior unsecured notes due 2031, it attached a recovery rating of "5," meaning the agency expects holders to recover only 10% to 30% of their money in a default. In the same body of work it named, alongside customer concentration, "the uncertain residual value of GPU collateral" as a binding constraint on the credit.¹ That is a notable thing for a rating agency to write about a fast-growing cloud-computing company. Residual value is the swing risk an analyst flags for an aircraft lessor or a shipping line, not usually for a business whose product is compute.

The reason it appears here is the subject of this piece. A growing share of GPU infrastructure capex is financed through special-purpose vehicles (SPVs) secured by the GPUs themselves and by the customer contracts that run on them. The structures are well built and the senior lenders are sophisticated. But they hold a real financial risk, the risk that the collateral loses economic value faster than the debt is repaid and faster than the books recognize, and they hold it in a form that does not show up cleanly on first inspection. Part of the risk sits in private credit that files nothing. Part of it sits in the gap between what a GPU is carried at on the balance sheet and what it is economically worth. And part of it has been moved, by design, to the layers and counterparties where disclosure is thinnest. The thesis of this article is not that a collapse is coming. It is that the risk is genuine, it is mispriced into invisibility, and the place to look for it is precisely where the structure is built to keep it out of plain sight. The piece starts at the most concrete layer, the project-finance structures funding the clusters, with CoreWeave as the case that can actually be read in primary filings.

The Structure, and Why It Is Hard to See

CoreWeave is the only sizable GPU lessor that files with the US Securities and Exchange Commission. Its debt stack is anchored by a Blackstone-led delayed-draw term loan, the facility announced in May 2024 at $7.5 billion and later expanded, called DDTL 2.0 in the filings, which followed a $2.3 billion predecessor and was joined by further facilities in 2025.²³ Each facility sits in a dedicated bankruptcy-remote subsidiary, named in the filings as CoreWeave Compute Acquisition Co. IV, V and VII LLC, secured by the GPU hardware in that subsidiary and the customer contracts it serves, and priced early on at roughly 11% on a variable basis.³

The clearest illustration is the OpenAI agreement. The Q3 2025 Form 10-Q, filed on 13 November 2025, describes a strategic customer agreement with committed value of up to $11.9 billion through October 2030, with the infrastructure held in an SPV intended to incur its own debt. In the company's own words, "in the event of default, OpenAI has a lien and security interest in the equity of the SPV."⁴ The vehicle holds the hardware, borrows against it, and pledges its own equity to the buyer of the compute.

A first point of precision, because it is where a careless version of this argument goes wrong. CoreWeave consolidates these SPVs, so the debt is on CoreWeave's own balance sheet, not hidden from it: $14.0 billion of total debt against $20.7 billion of property and equipment at 30 September 2025, up from $7.9 billion of debt a year earlier.⁴ The "hard to see" quality is not that CoreWeave conceals the borrowing. It is, first, that CoreWeave is the exception. For the private neoclouds that have copied the structure, there is no consolidated balance sheet to read at all. Second, the terms that would let an outsider judge the risk, the advance rate against the hardware, the residual-value curve the lenders assume, the covenant thresholds, are redacted from the filed agreements even in CoreWeave's case. Third, and most consequential, the assets are carried on a depreciation schedule that may be slower than their economic decay, which is the part of the risk that can sit inside a balance sheet that looks orderly. Those three points are the spine of this piece.

A note on the most-cited stress metric, to keep the argument honest. CoreWeave's Q3 2025 operating income of $52 million covered only about 6% of its $311 million quarterly interest expense.⁴ A bull would rightly object that the company is in a heavy growth-capex phase, so a single quarter's operating income against interest is not the steady-state picture; the contracted backlog, not the current income statement, is what the facilities are really lending against. That is fair. The point of the figure is narrower: it shows the structure depends entirely on the contracted backlog converting as written, with no current earnings cushion if it does not.

How Big, and How Much of It Is Visible

The structure has been replicated across the neocloud sector, almost entirely in private credit where terms are not disclosed. Lambda secured up to $500 million from Macquarie in April 2024, structured as a special-purpose GPU financing vehicle.⁵ Crusoe layered a $225 million facility from Upper90 with a $750 million facility from Brookfield in 2025.⁶ In January 2026 Apollo led a $3.5 billion capital solution for Valor, the entity financing GPU infrastructure for xAI, secured against GB200-class hardware, which closes off any argument that CoreWeave's structure was a one-off.⁷

The aggregate can only be estimated, and the reason it can only be estimated is itself the point. PitchBook reports that AI startups absorbed 24.9% of all venture-debt dollars in 2025, roughly $22.9 billion, and names BlackRock, JPMorgan and Carlyle among the lenders active in GPU-backed loans.⁸ Aggregating the named bilateral facilities with that venture-debt slice puts a conservative floor for GPU-collateralized debt somewhere in the range of $50 billion to $70 billion, and the true figure is unknowable because most facilities are private and disclose nothing. Jim Labe, co-CEO of TriplePoint Capital, named the concern precisely in the PitchBook feature: a chip "worth more than its sticker price one quarter, then outclassed by a next-gen model the next."⁸

What the Structure Is Protected By, and What It Is Not

These facilities are not naked bets on the resale price of used chips, and any serious version of the bear case has to start by saying so. Three protections are real. The loans amortize against contracted, often take-or-pay revenue from named counterparties, so a large share of principal is repaid from the contract regardless of what a used H100 fetches. They are non-recourse to the parent beyond a narrow guarantee covering fraud and certain insolvency events, so a residual shortfall is ring-fenced inside the vehicle. And the senior lender advances only a fraction of the hardware cost, so there is equity beneath the senior debt to absorb the first loss. Together these are why the honest framing of this market is a grind rather than an imminent blow-up, and why sophisticated credit investors have been willing to fund it at around 11%.

The mistake would be to read those protections as removing the residual and obsolescence risk. They relocate it. The contract covers the loan only for as long as the contract runs and only if the counterparty performs; the residual value still has to carry the uncontracted tail of the asset's life, and it carries more of it for the weaker, more private borrowers whose contracts are shorter and whose terms are not public. Non-recourse does not make the loss disappear; it assigns it to the equity and the unsecured layer, which is exactly where S&P's recovery rating of "5" already prices a 10% to 30% outcome.¹ And a conservative advance rate protects the senior lender only if the assumed residual is right; if the economic value of the collateral falls faster than assumed, the equity cushion that looked conservative at origination thins from below. Each protection works by moving the risk to a place that is harder to see than the parent's headline balance sheet: into private vehicles, into junior tranches, into the tail. That relocation is the mechanism, not a reassurance.

The Aircraft Benchmark

No lender believes a GPU is an aircraft, and the point of the comparison is not that anyone confused the two. It is that aircraft leasing is the largest and longest-running example of this exact financing form, residual-dependent asset finance housed in bankruptcy-remote vehicles, which makes it the benchmark against which the GPU version's missing infrastructure is most visible. AerCap, the largest aircraft lessor, depreciates passenger aircraft straight-line over 25 years to a 15% residual value, in the words of its 2024 Form 20-F: residual values are "generally estimated to be 15% of the manufacturer's estimated realized price for the aircraft when new."⁹ Air Lease Corporation uses the same 25-year life and 15% residual and adds the risk logic explicitly, owning aircraft "during the first third of their estimated useful life" and selling "to mitigate residual value risk."¹⁰

What makes the aircraft residual trustworthy is a set of supporting facts, and the comparison is useful precisely because the GPU market has none of them. An airframe is technologically stable across decades. There are published ownership registries, certified appraisers under the International Society of Transport Aircraft Trading, two large publicly listed pure-play lessors, and a secondary market with observable transaction prices. The lessor can rotate out of an asset in the first third of its life because a deep market exists to sell into. GPU-backed credit has one SEC filer, redacted agreements, no liquid secondary market for used clusters at scale, and a "first third" of a claimed six-year life that lasts about two years, which is roughly when the next NVIDIA generation re-prices the collateral. The structure was borrowed; the infrastructure that makes the structure safe was not.

Why the Residual Decays: the Refresh Cadence

The mechanism that erodes the residual is NVIDIA's product cadence. Hopper, the H100 generation, began shipping in late 2022. Blackwell was announced eighteen months later, on 18 March 2024, with NVIDIA claiming up to 25 times less cost and energy consumption than its predecessor for trillion-parameter inference.¹¹ Blackwell Ultra followed exactly twelve months after that, on 18 March 2025, with a claimed 1.5 times the performance of the prior generation and a stated 50 times revenue opportunity for AI factories relative to Hopper; the same keynote confirmed Rubin for 2026.¹² The multiples are NVIDIA's own marketing and should be read as such, but the direction and the roughly annual cadence are not in dispute.

The rental market has tracked that cadence, and here the distinction between price and value has to be made carefully, because it is where this argument is most often overstated. The one-year rental rate for an H100 fell from a peak near $8 per hour in 2023 to roughly $1.70 by October 2025.¹³ That figure is a rental rate, not a sale price, and it overstates the fall in the asset's resale value: part of the $8 was scarcity pricing during the 2023 shortage, and part of the decline since is supply catching up rather than the chip degrading. A used H100 still sells for a meaningful fraction of its replacement cost, and it cascades down a usage curve from training into inference, where it earns for years. The narrow, defensible claim is this: a cash-flow-secured facility lends against the asset's earning power, and earning power is what the rental curve measures. Whatever the resale value, the rate at which an H100 can be rented, which is the cash that services the loan, has fallen by roughly three quarters in two years, while the debt that funded it amortizes over five to six.

The Risk the Balance Sheet May Not Show

This is where the financial risk becomes least visible and most worth stating. A GPU is carried on the balance sheet at cost less accumulated depreciation, and depreciation runs on an assumed useful life. CoreWeave depreciates its technology equipment, including GPUs, straight-line over six years, having lengthened the assumption from five years at the start of 2023.¹⁴ That choice did not start at CoreWeave: Alphabet extended the useful life of its servers from four years to six effective January 2023, a change that by itself reduced depreciation expense by $3.9 billion and lifted net income by $3.0 billion in fiscal 2023.¹⁵ The lengthening flattered reported earnings across the sector and set the six-year figure now treated as the industry-standard accounting life.

A straight-line schedule writes the asset down evenly. The economic value of a GPU does not fall evenly; it falls faster in the early years, as each new generation re-prices it, and then flattens toward an inference-use floor. The result is a gap. For a stretch in the middle of the asset's life, the carrying value on the balance sheet sits above the asset's economic value, and nothing in the financial statements makes that gap visible until an impairment or a sale forces it into the open. A reader looking at an orderly PP&E line and a six-year schedule sees nothing wrong. The gap is real all the same.

The most informative evidence that the gap is not hypothetical is a reversal by one of the operators with the deepest view of the hardware. Amazon, in its fiscal 2024 10-K filed on 7 February 2025, shortened the useful life of a portion of its servers from six years back to five and recorded roughly $920 million of accelerated depreciation, citing the pace of technological change.¹⁶ The figure is small against Amazon's capex and applies to a mix of servers rather than to GPUs specifically, so it should not be oversold. But the direction is the signal: when a company with that vantage point revisits the assumption, it shortens the life rather than lengthening it. Whether the six-year life embedded across the sector, and referenced when GPU facilities are described as using "industry-standard" accounting, will hold for the GPU-heavy slice is the open question. The honest position is that the lenders' own residual curves are private and cannot be checked, so the publicly visible depreciation schedule is the best available proxy, and the one recent move in it points the wrong way for the optimistic case.

What to Watch

Three things, with concrete markers, will resolve the open question over the coming quarters.

The first is the appearance of a real secondary-market price for used clusters. As long as residual value is observable only through rental indices, the coverage in these facilities rests on assumption rather than on a clearing price. A liquidation of a sizable cluster, through a distressed neocloud or a lender exercising remedies, would produce the first hard recovery data point. Watch for any disclosed sale of H100-class hardware at scale and the implied cents on the dollar against original cost.

The second is the depreciation footnote. Amazon has already shortened useful life on part of its fleet. Watch for any further shortening of server or GPU useful lives, at the hyperscalers or at CoreWeave, in fiscal 2026 or 2027 filings. A move from six years toward five or four would pull the carrying-value gap into the open and tighten every coverage ratio that references EBITDA.

The third is the pricing of the next wave of GPU-backed issuance and the recovery ratings attached to it. S&P's positive outlook on CoreWeave rests partly on its contracts holding. Watch the recovery rating on any new unsecured issuance, and watch whether facilities backed by counterparties weaker than Microsoft and OpenAI can price at all. The first GPU-backed facility that fails to clear, or that clears only with a parent guarantee stapled on, would mark the point at which the market stopped taking the residual assumption on trust.

Personal View

The body above is analytical. What follows is my own view, kept brief.

The feature of this structure that stays with me is how little of it is visible to anyone outside the deal. The borrowing sits in special-purpose vehicles, much of the lending is private credit that files nothing, and the part that is disclosed has its key terms redacted. The ordinary saver, whose pension fund or insurer may sit somewhere in the chain of buyers for this paper, has no realistic way to see the residual-value bet being made on their behalf.

In my view, the right frame is a financing machine that runs on momentum. As long as demand for compute keeps rising and each NVIDIA generation keeps justifying the last, the collateral can be refinanced, the contracts renewed, and the residual question deferred. And momentum does not merely sustain the structure, it compounds it. Every larger capex round is financed the same way, so the absolute size of the residual-and-obsolescence bet grows with each cycle, and because the contracts and the debt are long-dated, the moment at which that bet is settled keeps moving further out. Growth does two things at once: it enlarges the exposure and it postpones the reckoning. That is comfortable on the way up and the wrong shape to be holding if the direction changes.

The mechanism rhymes with the pre-2007 mortgage structures in one specific respect: risk that is hard to see is easy to keep funding while the underlying values are still rising, and almost nobody spends the up-phase modelling the down-phase. It is worth being precise about where the rhyme breaks. These are contracted cash flows from real, profitable counterparties, not loans against household income that was never there; the losses are ring-fenced by non-recourse structures; and the demand is, for now, genuine rather than speculative.

So the interesting question is not whether the structure works today. It plainly does, and on current evidence the momentum is still building, which means the bet is still growing rather than unwinding. The interesting question is the one the hype leaves unasked: what the consequences would look like if that momentum breaks, how large the position will have compounded to by then, and how far through a largely invisible chain the damage would travel.

Later in this series: Part 2 reads the vendor-financing loop between NVIDIA and its largest customers against the 1999 telecom precedent; Part 3 unpacks the quarter-trillion of compute commitments sitting in 10-K footnotes; Part 4 closes the series with the on-balance-sheet credit profile of the hyperscalers absorbing it all.

END · 01