Three Clocks, One Asset: Who Carries the Re-Lease Risk in AI Compute

One sentence in a loan announcement

On 10 August 2026 CoreWeave announced that it had closed a $2.6 billion delayed-draw term loan. The release says the facility has an approximate five-year maturity and that the underlying customer contracts average approximately three years in length¹. Nothing in the announcement presents this as a problem, and it may not be one. But it is the rare case of an issuer putting two of the clocks that govern AI compute side by side, and the gap between them is the subject of this analysis.

The popular debate is about a third clock. Are GPUs depreciated over too many years? That argument has a loud advocate and a long comment thread, and the evidence for it is thin. The more useful observation is that AI capacity is being financed in layers, each with its own duration, and that the layers do not line up. For anyone who buys or commits to compute capacity, the decisive variable is what happens after the first contract ends, and who is paid to carry that risk.

This is a conditional structural claim. It is not a forecast of write-downs, and the evidence below does not support one.

Three clocks, one asset

The accounting clock is the depreciable life that a company assigns to servers. The hyperscalers have moved it repeatedly. Microsoft lengthened server and network lives from four to six years from its 2023 fiscal year. Alphabet made a similar move in January 2023. Meta extended most servers and network assets to five and a half years from January 2025. Amazon went from five to six years in January 2024, then cut a subset of its equipment back to five in January 2025⁴⁵. The direction is not one way, and a company that shortens a life reduces its own earnings.

The financing clock is the debt tenor. For the newest CoreWeave facility, the filing-based reporting gives a maturity of 1 September 2031, with a debt-service coverage covenant of at least 1.35 times, a guarantee from CoreWeave and security over substantially all the assets of the borrowing unit².

The commercial clock is the customer contract, which the issuer puts at about three years¹.

Two cautions on the table. Useful lives are management estimates, and nothing we opened shows an auditor or regulator challenging the five-to-six-year range; the notes to the relevant filings were not opened, and the dollar effects of the life changes differ between the secondary sources we read, so they are not used here⁴⁵. And the three-year contract against five-year debt is documented for one facility only. Another reading of the same sources says that earlier CoreWeave facilities were backed by contracts that ran to the maturity of the debt³; that is the analyst's account of company statements, and we have not seen the originals.

Where the risk lands

Take a deliberately simple case. A neocloud signs a three-year capacity contract and finances the hardware with five-year debt. Index the monthly contract revenue to 100 and suppose, for illustration, that the debt needs 5,000 of cash over its life. Ignore interest, operating cost and any residual value; the real facility will amortise on its own schedule, which we have not seen.

The contract delivers 36 months at 100, or 3,600. That covers 72 per cent of what the debt needs. The remaining 1,400 must come from months 37 to 60, which means the capacity has to be renewed or re-leased at about 58 per cent of the original price (1,400 divided by 24 months, divided by 100) just to break even. At renewal pricing of 75 per cent the structure clears with room to spare. At 50 per cent it falls short.

This is composite arithmetic, not CoreWeave's numbers. Its use is to show where the sensitivity sits. The break-even is not at the first contract, where the revenue is committed. The sensitivity sits in the second contract, which has not been priced yet.

CoreWeave's release is open about this. The facility gives the company the option to renew existing contracts or re-lease capacity to other customers at the end of the initial contracts, subject to criteria in the credit agreement¹. The release does not describe those criteria. The credit agreement is the document that would, and we have not opened it.

The decisive question is not how long a GPU lasts. It is who is contractually exposed to the price at which its second life is let.

What the rental market says

If obsolescence were already compressing second-life pricing, rental rates for older GPUs would be falling now. The record is more mixed than that, and readers should hold it loosely.

H100 rental rates fell from roughly $7 to $10 an hour at launch in 2023 to about $2 to $4 by late 2025, according to a publication from a mining and high-performance-computing firm, which also notes a short-lived rebound recorded by Silicon Data in early 2026⁷. A consumer technology outlet reports H100 rates moving above $2 and approaching $3 an hour, from just under $2 at the start of 2026, without naming its data provider⁸. A GPU-compute vendor reports one-year contract rates rising from $1.70 to $2.35 between October 2025 and March 2026, citing SemiAnalysis, and an on-demand market median near $3.39 in September 2026, citing an aggregator⁹. We did not open those underlying sources, two of the three publishers sell or are close to compute, and the series share underlying data.

The defensible reading is narrow. Rental prices for this generation fell, then recovered. That cuts against the simple claim that the market is already testing the accounting lives. It does not tell us what a 2024 contract will fetch when it rolls in 2027, because rising rates on a supply-constrained market are what a lender would expect before new capacity arrives.

The strongest case that this is nothing

A sceptic would say that tenor gaps are ordinary in asset-backed lending. Ratings agencies looked at this structure and rated it Ba2 and BB+¹. A covenant, a guarantee and security over the borrower's assets protect lenders². Older chips are reportedly heavily used for inference⁴. Amazon's own record shows that useful lives move in both directions⁴. On this reading the debate is commentary, and a refinancing in year four is a normal feature of the business.

Much of that is right, and it changes the paper in three ways. First, it removes any claim that lenders are unprotected or that neocloud debt is distressed; no source we opened documents a covenant breach or a loss on a GPU-secured facility. Second, it moves the point of interest from the asset to the contract, because protection is attached to the structure and to the borrower, not to the hardware's future price. Third, it supplies the test. If renewal pricing on expiring contracts holds near the original levels, the mismatch is benign and the central tension of this analysis disappears. That is the falsification test: compare renewal prices on expiring first-generation contracts with the prices they replaced.

One pricing signal sits uneasily with the benign reading but does not contradict it. A 100 basis point difference separates the May facility at SOFR plus 450 and the August facility at SOFR plus 550, at the same ratings¹²³. One analyst reads that gap as the market's first price for the tenor mismatch³. That is an inference. The two facilities differ in other respects, and the August terms were first marketed tighter and widened in negotiation, according to reports citing people familiar with the offering².

What enterprise buyers can do with this

The enterprise reader is usually not the lender. But a buyer's capacity commitment is the revenue that supports the debt, so the buyer holds negotiating position that the lender is relying on without knowing it.

The practical questions are these. Who holds the re-lease risk in our contract, and does it become ours at renewal? What useful life and renewal assumptions does our provider use in its own financing? How much of our supply sits with a counterparty whose refinancing falls due in the same window as our renewal? The scale of the sector makes the questions more than academic: the four largest US hyperscalers purchased roughly $434 billion of property and equipment in the four quarters to March 2026, on one analyst's summation of reported figures⁵, and one hyperscaler's cash capital expenditure alone rose from $48.1 billion in 2023 to $77.7 billion in 2024 while operating cash flow rose from $84.9 billion to $115.9 billion⁶. Capital intensity is rising faster than any assumption about residual value has been tested.

Three structures follow. Stagger contract terms so that no single renewal decides the year. Negotiate renewal options with price caps while the provider needs the anchor customer. Spread capacity across providers, so that one counterparty's refinancing does not constrain supply. For boards and audit committees, the control point is disclosure: ask for the provider's financing structure and asset-life assumptions as part of any major capacity decision, and include provider credit in third-party risk review.

Where the argument stops

The argument does not apply where contracts match asset life, where offtakers are investment-grade and committed for long terms, or where an enterprise owns its hardware for its own use. It would be weakened by disclosed renewals at or above original pricing, by sustained Hopper-class rental prices after Blackwell-class volume arrives, or by operators disclosing high utilisation on older fleets.

What remains uncertain is the central fact: renewal and re-lease economics for first-generation contracts, which begin to expire between 2027 and 2029. No source we opened provides renewal pricing. We also did not verify the useful-life notes in most hyperscaler filings, customer identities behind the August facility are only reported, and debt totals for CoreWeave differ by source. The Q1 2026 principal of $25.15 billion is the figure we can attribute, to the company's first-quarter filing as reported².

The question that outlasts the useful-life debate is easy to state. When a contract that financed a GPU runs out, what is the second price, and whose balance sheet is written to receive the difference?