GPU-backed financing: asset-backed debt, equipment loans and leases

No operator's cash flow can pay for the AI build-out as it happens. A large GPU campus costs far more than its owner earns in a year, the chips must be bought before they earn anything, and the fleet has to be replaced every few years as new generations arrive. The scaled neoclouds spend several times their annual revenue on capital expenditure. Even at a steady state, spending to sustain the fleet roughly matches depreciation, so outside capital is a permanent part of the model, not a phase of growth.

The industry's main financial achievement of the past three years has been to build a full stack of capital to carry that spending. This piece explains its layers, the structures used to lend against GPUs, and what lenders are actually taking a view on when they do.

The capital stack

Capital reaches an AI-infrastructure platform through several layers, each priced off a different risk.

• Growth and pre-listing equity, which backs the team, its contracts and its access to scarce chips and power.
• Vendor and strategic support, such as chipmakers' commitments to rent back capacity, or backstops from large technology companies that stand behind a customer's payments.
• Convertible notes, which give lenders a share in any rise in the listed equity in return for a lower coupon.
• Contract-backed GPU debt, investment grade where the customer is investment grade, and below it where the contract is shorter or the customer weaker.
• Project debt for powered shells, long bonds against long leases, which looks like conventional infrastructure finance.
• Customer prepayments, cash paid upfront that part-funds the build.
• Equipment loans, leases and sale-leasebacks, which finance the hardware or recycle capital already spent.

The layers are priced by who stands behind the cash flows. The closer a loan sits to an investment-grade customer's promise to pay, the cheaper it is.

Contract-backed GPU debt: the template

The defining design choice of the modern GPU loan is that lenders and rating agencies underwrite the customer, not the silicon. In the template transaction, a separate, bankruptcy-remote company, a special-purpose vehicle, owns the GPUs and holds the right to the payments under a take-or-pay contract with an investment-grade customer. The customer's payments flow into the vehicle and pay the lenders first, through a fixed order of priority. The GPUs are there as a floor for recovery if things go wrong, not as the main source of repayment.

These loans are usually delayed-draw term loans: the facility is agreed in full, and the operator draws on it as GPUs are delivered and installed. The loan can be rated well above the operator itself, because what is rated is the customer's obligation. CoreWeave's investment-grade facility of March 2026 marked the moment rating agencies accepted that a contracted GPU fleet can carry investment-grade debt, and facilities for IREN, Lambda and Nebius followed during the year.

Asset-backed securities

Asset-backed securities take the same idea to the bond market. A pool of assets that pay predictable cash, such as contract receivables or leases, is placed in a vehicle, which issues notes in layers of seniority. The senior notes are paid first and carry the highest rating; the junior notes absorb the first losses in return for a higher yield. Investors look at the quality of the customers, how much the cash flows exceed what the notes require, and whether the notes repay before the assets lose their value.

Datacentre leases have long been financed this way. For GPUs the question is the life of the asset: a building lasts decades, a GPU generation a few years, so the notes must amortise quickly enough that they are repaid while the chips still earn. That is why asset-backed structures for GPUs lean so heavily on the contract behind them.

Equipment loans and leases

Not every financing runs through a contract vehicle. The server makers that assemble GPU systems also offer financing alongside support and logistics, and specialist lenders make equipment loans secured on the hardware itself. Leasing is common too. In a finance lease the operator effectively owns the equipment and records it as an asset and a liability; in an operating lease the lessor keeps the risk of what the equipment is worth at the end of the term. Sale-leasebacks let an operator transfer equipment or a building it already owns to a buyer and lease it back, freeing capital for the next build.

In each of these the lender or lessor is taking a view on the economic life of the GPUs: how much they will be worth when the loan or lease ends, and whether someone will still pay to rent them.

What lenders underwrite

Across all of these structures, lenders look at the same handful of things.

• The customer. An investment-grade customer under a long take-or-pay contract is the strongest collateral there is. Concentration matters: a backlog dominated by one customer carries that customer's risk.
• Tenor. Debt that outlives its contracts turns lenders into underwriters of renewal risk, and they now charge for it. Matching the debt to the contract is part of the design.
• Contract cover. Contracted backlog divided by debt is a crude but telling test, and a falling ratio is a warning sign.
• Depreciation. Lenders tend to assume shorter GPU lives than the accounts do.
• Prepayments. A customer that pays upfront cuts the debt needed and the all-in cost of capital.

For operators, balance-sheet design has become product design. Those who can fund large programmes at a low cost of capital can sign contracts that others cannot.

When rates move

The stack is sensitive to interest rates, and unevenly so. When the Federal Reserve raised rates on 16 September 2026, the weakest layers repriced most. Unsecured notes of leveraged operators fell sharply, new contract-backed loans came to market wider than their first price talk, and some issuers turned from secured debt to convertible notes. Floating-rate borrowers felt the rise through their coverage tests unless they had hedged. The investment-grade, contract-backed layer moved least, which is the point of building it.

Reading financings in our data

We record financings as they are announced and filed, each with its date, its instrument and its source. The Capital Markets page shows the last thirty days of deals with headline terms to Compute Open readers; subscribers see the full terms and the closing calendar, and Plus adds the full financing tape and the archive. The deals register explains what we count and how.

Sources

• AI Infrastructure Primer 2026, Compute Terminal: the chapters "Financing the buildout" and "The value chain", with their dated sources.
• Company filings (8-K, 10-Q, 10-K and 6-K, SEC EDGAR) for the facilities named, as recorded in the Primer and our financing tape.
• Compute Terminal deals register and financing tape; Capital Markets.