GPU-backed debt: How AI chips and customer contracts became loan collateral

GPU-backed debt is a loan where the lender’s security is a pile of AI chips, usually together with the customer contracts that pay to use them. It lets a company buy graphics processing units (GPUs) now and repay from the rent those chips earn later.

The idea is simple to state and hard to price, because nobody is sure what a used AI chip will be worth in a few years.

A car loan for a rack of servers

Think of a taxi company that borrows to buy cars. A bank lends against the cars, and the company repays from fares. If the company fails, the bank can sell the cars.

A GPU cloud firm works the same way. It borrows to buy servers full of Nvidia chips, rents them to AI developers, and uses the rent to repay. The twist is that lenders have been told to look at the rent more than the cars.

Putting the chips and the contracts in one box

According to a venture-capital write-up by Wing, the template was set on August 3, 2023, when CoreWeave closed a $2.3 billion loan co-led by Magnetar and Blackstone Tactical Opportunities. It was a delayed-draw term loan (DDTL): the borrower takes the money in installments as chips arrive, instead of all at once.

The key design, per the same write-up, was a bankruptcy-remote special purpose vehicle (SPV). That is a separate legal entity that holds the chips and the assigned customer contracts, so lenders are paid from that contracted cash flow and are not dependent on the parent company’s other troubles.

The intended outcome is that the loan is repaid before the hardware becomes obsolete, so the lender never has to sell a used GPU.

From a risky bet to an investment-grade rating

Early pricing was steep. Wing reports the 2023 loan cost roughly 11% to 14% all-in. Another newsletter, Global Data Center Hub, describes a rate near 15% and says the loan was secured only by resale value of the hardware.

Later deals leaned on the customer. Global Data Center Hub reports that a CoreWeave facility closed March 31, 2026 ($8.5 billion, backed by GPUs and contracted Meta payments) was rated A3 by Moody’s and A (low) by DBRS, and priced at the Secured Overnight Financing Rate (SOFR) plus 2.25 percentage points. It argues the agencies effectively rated the collateral and Meta’s credit rather than CoreWeave itself.

Investment grade means a rating of BBB- or higher at S&P or Fitch, or Baa3 or higher at Moody’s, according to the deal-software firm Datasite’s glossary. Anything lower is often called high yield. Wing reports a later CoreWeave facility tied to weaker customers was rated Ba2/BB+ and priced at SOFR plus 4.5 points, using broadly similar hardware. The gap suggests the customer’s credit, not the chip, sets the price.

Where our SpaceX story fits

Elon Musk wrote on X that SpaceX closed $40 billion of debt for GPUs and that it could because SpaceX is investment grade with Starlink and launch revenue. We covered his claim in Musk: Debt markets have shut AI data center lenders out, bar SpaceX.

That post does not say what secures the money, who lent it or on what terms, so it is not confirmed to be a GPU-backed structure of the kind above. It does show the same dividing line. A borrower with large income outside AI data centers can raise debt on its own credit. A pure GPU renter must lean on the chips and the contracts.

Bloomberg, as listed by Techmeme, traced how the Blackstone-led 2023 CoreWeave loan became a blueprint for chip-backed debt, a market it says is growing fast.

The question nobody can answer yet: What is a used chip worth?

The risk sits in the collateral. Wing notes that firms depreciate identical hardware differently: CoreWeave over six years, Nebius over four. It also reports that GPU rental rates swung from about $8 an hour in early 2024 to $1.70 in October 2025 and about $2.35 in March 2026.

Wing cites a dataset of publicly recorded used-GPU sales totaling only about $26.3 million over three years. With so few trades, there is little price history to test claims. Wing gives estimated recovery of 50% to 70% of new price for two- to three-year-old hardware in an orderly sale and 30% to 50% in a distressed one, but these are estimates, not market quotes. It adds that critics such as Jim Chanos and Michael Burry argue GPUs have an economic life of only two to three years, a view it says is hard to test.

Other risks named in the two write-ups:

• Customer concentration: a loan can depend on one buyer’s commitment.
• Renewals: Wing says the 2023 to 2024 contract vintage begins rolling off in 2027, and if contracts do not renew, lenders fall back on a thin used-chip market.
• Narrow ratings: Global Data Center Hub argues the investment-grade label covers one ring-fenced tranche, not the whole company.

What to watch

Three signals will show whether this market is durable. First, whether lenders keep pricing chip loans off the customer’s credit as stress grows. Second, whether contracts signed in 2023 and 2024 get renewed from 2027. Third, whether used-GPU trades become frequent and public enough to give lenders a real price.

Note that the figures above come from industry newsletters and a venture firm’s analysis, not from loan filings, and some reports on specific deals differ on terms. For any single deal, the credit agreement and rating reports are the primary record.