How lenders size a GPU loan: two tests, and the smaller one wins

Ask a GPU lender how much they will lend and the answer comes in two parts. The first is an advance rate, a percentage of what the borrower pays for the hardware. The second is a debt service coverage ratio, or DSCR: how many times the contracted cash flow covers each payment of interest and principal. The loan is the smaller of the two results.

Test one: the advance rate

The advance rate caps the loan against cost, so the sponsor always has equity at risk. The public deals show a wide range:

• IREN Hardware 3: 95% of project costs for each tranche.
• Nscale Ward County and North Carolina: 90% of GPU capex.
• GMI Cloud: 80% of server capex, with debt also capped at 55% of the value of its NVIDIA contract.
• Crusoe Iceland (launch terms): 75% loan-to-cost.
• USD.AI on-chain frameworks: up to 70% against tokenised GPU receipts.

The highest advance rates sit next to the strongest customers. A 95% advance works only because the contract, not the hardware, carries the credit.

Test two: the sizing DSCR

The second test starts from the contract. The lender projects cash flow available for debt service (contracted revenue less operating costs such as power, colocation fees and maintenance) and sets the debt service the project can carry at a fixed cover. Nscale's US loans size at 1.30x. IREN sizes at 1.20x. Lambda's Compute I loan amortises at a 1.20x minimum DSCR, according to DBRS.

A worked example

Take a simple, illustrative project: US$1,000m of GPUs, a five-year contract and US$260m a year of cash flow available for debt service. Assume a 7% all-in interest rate and a loan that fully amortises over the five years with annual payments.

• Advance rate test at 90%: US$900m.
• DSCR test at 1.30x: the project can pay US$200m a year (US$260m ÷ 1.30). Five annual payments of US$200m at 7% support a loan of about US$820m.
• Loan size: the lesser, about US$820m. The DSCR test binds, and the sponsor funds the remaining US$180m.

Raise the contract cash flow to US$300m a year and the DSCR test supports about US$946m, so the 90% advance rate binds instead, at US$900m. In practice the contract price, the tenor and the interest rate decide which test bites. These figures are illustrative only.

After closing: traps and covenants

Sizing sets the loan on day one. A ladder of lower DSCR tests then polices it. In the Nscale and IREN structures a DSCR below 1.10x traps cash, and if the trap lasts six months the trapped cash prepays the loan. The financial covenant sits at 1.05x.

The gap between those levels is the borrower's headroom. A loan sized at 1.30x can absorb a fall in cash flow of about 15% before it reaches a 1.10x trap, and about 19% before a 1.05x covenant. A loan sized at 1.20x has about 8% before the trap and 12.5% before the covenant.

CoreWeave's facilities use a different design. DDTL 3.0, 5.0 and 5.5 carry maintenance DSCRs of 1.35x to 1.40x, and DDTL 4.0 sits at 1.15x with a minimum liquidity test. DDTL 5.5 adds a contract test: cash is trapped if the projected value of the contracts falls below 2.40x of the loans.

A third test: what the GPUs are worth

Some lenders add a value test on top. IREN's Hardware 3 caps the loan at 65% of the depreciated cost of the hardware, tested monthly, on a six-year straight-line life. It is the closest the GPU market has come to the loan-to-value tests used in aircraft finance.

Questions for the next term sheet

• Which costs sit above the line in cash flow available for debt service, and who bears an increase in power costs?
• Is the DSCR historical, projected or both, and over how many months is it averaged?
• How long does a cash trap run before it turns into a prepayment?
• Is there a cure right, and can the sponsor cure with new equity?