What it takes to qualify for a GPU loan: the lender’s checklist

A GPU loan gets approved on the contract, not on the hardware. I see financing requests every week, and the ones that die mostly die on the first call, for the same few reasons. Here is the checklist I would hand a borrower before that call, in the order a lender works through it.

1. The offtake

GPUs lose value every year. The customer contract is what repays the loan, so lenders treat it as the primary source of repayment. They read it first and the hardware second.

We start with four questions. Who is the customer, and can they pay for the full term? How long is the contract? Is it take-or-pay, meaning the customer owes the money whether or not it uses the capacity? Has any of it been prepaid?

Counterparty credit sets everything downstream. CoreWeave’s $8.5 billion DDTL 4.0 facility was written to fund previously contracted services for a leading AI enterprise, and it became the first HPC delayed draw term loan to earn an investment-grade rating, A3 / A (low). A smaller operator will not get that. The principle holds at any size.

Term matters just as much. The cleanest loan matures inside the contract. CoreWeave has since closed a facility with roughly a five-year maturity against customer contracts averaging about three years, and says plainly that those lenders are taking renewal risk. That is a public company with billions in contracted revenue. I would not plan on the same treatment for a small operator. Those financings differ greatly in size, structure and credit quality from ours.

Lenders also size the loan on the contract rate, not the spot rate. Silicon Data’s B200 index stood at $5.69 per GPU-hour on September 7, and the 12-month term rate was $5.39, about 5% lower (our math). The sizing follows the lower number.

Source: Silicon Data B200 index (SDB200RT), September 7, 2026. The 5% gap is a GPU Capital calculation from Silicon Data figures.

Show data

The 12-month B200 rate sits about 5% below the spot index

B200 rate	$ per GPU-hour
Spot index	$5.69
12-month term rate	$5.39

A customer prepayment cuts the lender’s exposure and shows the customer means it. We wrote up how those deals are built in our post on long-tenor prepaid GPU contracts.

2. The hardware

Lenders value the asset as a whole node. An 8-GPU H100 system runs roughly $250,000 to $320,000 in 2026 price guides, and that is the figure a lender haircuts, not a per-card price.

Then the questions. Which generation is it? The H100 rental index reads $2.53 per GPU-hour today against $5.69 for B200, so an older node earns less and gets a shorter loan. How old is it, and what is the exact spec: memory, networking, CPU, HGX or PCIe? Where does it sit? The lender needs the country and the facility, and has to know it could repossess and re-lease the node there. Who maintains it? OEM warranty, spares on site and remote hands all matter, because a node that sits broken earns nothing.

Which generation to buy is a separate question. Our post on which GPU server pays back fastest covers it.

3. The sponsor

Lenders back operators who have run GPUs before. They ask how many nodes you have run, for how long, and for whom. They ask for uptime history, and they want it from your monitoring system and your customers’ SLA reports, not from a slide.

Then comes power. They want the power agreement and the colo agreement, signed, sized to the megawatts you need, and running at least as long as the loan. Without power there is no cash flow.

A first-time operator can still get financed. The lender will want protection somewhere else: a stronger customer, a larger prepayment, or a smaller advance.

4. The structure

Four terms do most of the work. The advance rate is the share of node cost the lender funds, and it moves with contract quality and node age. Amortization should repay the principal inside the contract term, so nobody is counting on a renewal. Reserves are usually a debt service reserve of a few months, held in cash. The cash sweep sends excess cash to the lender instead of the sponsor when a customer pays late or the contract nears its end. Borrowers push back on the sweep. A lender looking at an unlet tail will ask for one anyway.

Pricing depends on your tier, and our post on cost of capital by tier covers it.

Where borrowers lose deals

Three mistakes account for most of the losses I see.

The loan term is longer than the contract term. A borrower asks for five years against a three-year contract, and the lender is now underwriting a renewal. A large issuer can get that done. Most borrowers cannot.

There is no power agreement. A customer contract with no signed power behind it is a promise to deliver capacity you cannot yet run. Lenders see that at once.

One customer for everything. Concentration is fine when that customer is investment grade and take-or-pay. It kills the deal when the one customer is a startup with a thin balance sheet and a right to terminate.

The checklist

Exhibit 2. The lender’s checklist, in the order a lender works through it

What lenders check	The question behind it	Typical red flag
Customer credit	Can the customer pay for the whole term?	Unrated startup, no financials, no parent guarantee
Contract term vs loan term	Does the loan end before the contract does?	Loan outlasts the contract
Take-or-pay and termination	What happens if the customer walks?	Termination for convenience, usage-based billing
Prepayment	How much cash is already in hand?	None, or refundable on demand
Node generation, age, spec	Will the node still earn at the end of the loan?	Older node, mixed spec, thin networking
Location and maintenance	Can we repossess, re-lease and keep it running?	Hard-to-enforce jurisdiction, no OEM warranty or spares
Operating record	Has this team run GPUs for a customer before?	No uptime history, no SLA reports
Power and colo	Is the power signed, sized and as long as the loan?	Letter of intent only, shorter than the loan
Structure	Does the loan repay inside the contract?	Balloon payment, no reserve, no sweep

Source: GPU Capital.

What to have ready before the first call

Put these in one folder: the signed customer contract, including the termination and SLA clauses; the power and colo agreements; the hardware quote or purchase order with OEM spec and warranty terms; an uptime report for any fleet you already run; and a one-page summary of what you are buying, what you are asking to borrow, and what you are putting in yourself. If the customer is not rated, add whatever financial information you have on them.

The one thing to do this month

Write down three dates: when the customer contract ends, when the power agreement ends, and when the colo agreement ends. Your loan term should be no longer than the earliest of them. If that is shorter than the loan you want, you have found the problem before a lender did. Extend the weakest agreement or shorten the ask.