GPU Loan Underwriting: What Lenders Actually Score Before They Fund

Summary

• GPU loan underwriting is the credit review behind a GPU-backed loan. It tests whether contracted revenue can repay the debt and whether the GPUs hold enough value while it does.
• The customer's credit and the offtake contract carry most of the review, roughly 70% by our published weighting for USD.AI's process. The operator and the hardware make up the rest.
• The outcome sets the loan's size, pricing and structure. Disclosed market deals show spreads moving with the credit of the customer behind the contract.

Introduction

You have a customer, a data center slot and a purchase order, and you need capital for the servers. GPU loan underwriting decides whether that capital arrives and on what terms. More of the AI buildout now runs through credit markets: J.P. Morgan Asset Management estimated in August 2026 that the full data center buildout could cost roughly $5 trillion through 2030.

In our view, GPU financing rewards preparation: an operator who knows what underwriters score can bring a deal that is ready to score. We write this from the deals USD.AI reviews.

This article covers the four things GPU loan underwriting scores and how underwriters handle depreciation. It then shows how the review becomes a term sheet and what to prepare before you apply.

What GPU Loan Underwriting Evaluates

GPU loan underwriting is the credit review a lender runs before financing GPUs. It assesses four things: the customer buying the compute, the offtake contract, the operator running the cluster, and the hardware and facility securing the loan. The offtake contract is the agreement under which a customer buys the operator's compute. Repayment comes from that contracted revenue, and the GPUs are the collateral if the revenue fails.

A GPU-backed debt facility is a cash-flow loan with hard collateral behind it. Underwriters start with who pays and finish with what the hardware is worth if payments stop. The weighting follows that order, which is why the strongest applications lead with the contract. For the mechanics end to end, see how a GPU-backed loan works.

The customer and the offtake contract

The customer's credit and the contract's legal strength account for about 70% of a GPU loan review. We published that weighting in August 2026 in "How GPU Financing Works at USD.AI." The customer is the party whose payments repay the loan, so underwriters read its financials, backers, cash runway and leverage first.

Then they read the contract. Term length, payment structure and exits decide how much of the loan the revenue actually covers. A take-or-pay contract means the customer pays for committed capacity whether or not it uses it.

In our view, take-or-pay terms and fixed monthly payments give underwriters the most to lend against. Exit clauses cut the other way: a customer that can terminate early shortens the revenue the loan can count on. A contract that runs at least as long as the loan, with payments sized to cover debt service, is the cleanest case.

Public pricing shows how much the customer matters. CoreWeave's DDTL 2.0 is a delayed-draw term loan, drawn in pieces as hardware is bought. Before a September 2025 amendment, it charged term SOFR plus 6.00–6.50% on investment-grade loans.

Non-investment-grade loans paid 13.00% plus term SOFR, per CoreWeave's October 2025 filing. The same borrower paid roughly double the margin when the customer behind the draw was weaker.

On-demand revenue doesn't disqualify a deal in Compute IQ, and it is priced for its shorter commitment. Compute IQ's illustrative sample deal covers 256 NVIDIA B200 GPUs on a 36-month AI lab contract. The customer earns 34 points and the contract earns 36.

The operator behind the cluster

The operator earns 18 points in that same illustrative sample. Three years running clusters adds 8, two clusters in service add 6 and a 25% equity contribution adds 4.

Underwriters look at operating track record, deployments in service, financial position, ability to deliver the deployment and the equity you put in. Delivery is the practical test: can your team receive the servers, install them on schedule and keep them online for the customer?

Track record counts, but it is one input among several. A first deployment qualifies when the customer and contract are strong, because those carry most of the score. Newer operators financing a first cluster with a signed customer contract are a core part of GPU financing for neoclouds.

The hardware and where it runs

The hardware has to be current-generation, serialized data center GPUs from the OEM or an authorized supplier, housed in a Tier III or Tier IV facility. The purchase order sets the hardware, the price and the delivery timeline, which gives underwriters a verified cost basis for the collateral.

Uptime Institute defines Tier III as a facility where every capacity component and distribution path can be removed on a planned basis. That covers maintenance or replacement without impacting operations.

Tier IV goes further: an individual equipment failure or distribution path interruption won't impact operations. The jurisdiction must support reliable title, security interests and enforcement. In financing through USD.AI, the hardware carries property and casualty insurance once escrow is released.

Lenders hold step-in rights at the data center, meaning they can access and take control of the hardware if a loan defaults. The colocation provider waives its liens on the hardware before escrow releases.

Generation matters too. In September 2026, NVIDIA's HGX lineup listed the B300 with eight Blackwell Ultra GPUs and the B200 with eight Blackwell GPUs. Rubin was listed as next. For system specs, see our NVIDIA B300 hardware reference.

How Underwriters Handle GPU Depreciation

The loan has to amortize faster than the hardware loses value, and the rental rate drives both sides of that test. GPU depreciation and repayment capacity move together because both follow what the chips earn per hour. Silicon Data put its H100 neocloud index at $2.94 per GPU-hour on Nov 2, 2026. Silicon Data's B200 index read $6.64 per GPU-hour on Oct 31, 2026.

H100 pricing has held up longer than many lenders assumed. On SemiAnalysis Weekly (Ep. 037, Oct 9, 2026), Jordan Nanos addressed one-year H100 deals. He said they are "approaching the prices that they were signed in the first half of 2023 today, like three bucks an hour."

Residual value sits on top of the contract as upside. On the same episode, Daniel Nishball said operators "underwrite it to make money on a five to six-year period. So they're not actually banking on the residual value at all." He called what comes after the term "like a bonus."

Server accounting lives run five to six years at Amazon, Meta and CoreWeave. Amazon cut the useful life of a subset of its servers and networking equipment from six years to five. The change took effect in January 2025, and Amazon cited the pace of AI development.

Meta moved most servers and network assets to 5.5 years, effective Jan 1, 2025. CoreWeave's 2025 annual report lists technology equipment at six years.

Underwriters model two curves against those lives. The first is the GPUs' expected resale value, month by month. The second is income stability: the offtake contracts, counterparty risk and expected utilization.

The loan works when both curves clear the debt schedule. Loans financed through USD.AI amortize straight-line monthly over 3–5 years, inside the accounting lives above. Lambda described its August 2026 term loan as having an amortizing schedule aligned with contracted cash flows and useful life.

How Disclosed Deals Price GPU Loans

The output of GPU loan underwriting is the term sheet, and disclosed deals show pricing moving with the credit behind the contracts. The table lists separate transactions, each with its own borrower, collateral and counterparties, so it doesn't form a market index. Facility sizes are "up to" commitments, and "n/a" marks a term the cited sources don't state.

The Valor row is a lease, included because it finances GPUs against a single named customer.

The pattern across the loans is consistent. The tightest disclosed margin, SOFR + 2.50% at Nebius, rests on an investment-grade customer. CoreWeave's non-investment-grade draws carried the widest. In our view, the customer's credit moves price more than any hardware detail in these deals.

How Financing Through USD.AI Is Scored

Compute IQ scores every financing originated through USD.AI on one 100-point scale, and that score sets the terms. It draws on more than $25 billion of GPU financing data points as of Oct 8, 2026.

The Compute IQ underwriting system scores the operator, the customer and the contract against USD.AI's underwriting standards. The score sets financeability, one of three pricing tiers, and the loan's size and structure. Deals scoring below 60 aren't financeable, and when a deal falls short, Compute IQ shows which gaps to close.

A passing score isn't a financing commitment. The remaining conditions are verified and final approval is given before a loan closes. The terms for GPU financing through USD.AI are set out below.

• Size. Loans cover up to 80% of the deployment cost (loan-to-value), with the operator contributing at least 20% as equity.
• Rate. Rates are fixed or floating.
• Term. Loans run 3–5 years with straight-line monthly amortization.
• Reserve. A three-month debt-service reserve is funded at closing.
• Range. Financing starts from about $5 million, with larger facilities available.
• Collateral. GPUs sit in Tier III or Tier IV data centers.

Every financing uses the same structure. A bankruptcy-remote special-purpose vehicle (SPV), a separate company that owns only the deal's assets, holds the GPUs, offtake contracts and revenue accounts. Lenders hold a first-priority claim on the GPUs, contracts and revenue, and a dedicated servicer manages each loan for its lenders for the full term.

Capital is held in escrow until the servers are installed, burned in and independently verified. That protects both sides: lenders fund working hardware, and you draw against a cluster that is ready to earn.

A published example shows the structure at scale. In June 2026, a $98.1 million facility originated through USD.AI financed 2,304 NVIDIA B300 GPUs for Duos Edge AI. On Oct 1, 2026, it became the inaugural loan sale through the GLX institutional marketplace, where seasoned, performing financings may be sold to institutional buyers.

What to Prepare Before Underwriting Starts

Five items start every GPU loan underwriting review, and the same package applies to every deal. Each one maps to a part of the score above, so a complete package lets the review start with the substance.

• Offtake contract. Bring an executed offtake contract or documented on-demand revenue.
• Hardware purchase order. Bring a purchase order from the OEM or an authorized supplier, or details of the deployed GPUs for a refinancing.
• Colocation agreement. Bring an executed colocation agreement with a qualifying Tier III or Tier IV data center.
• Equity contribution. Show an equity contribution of at least 20% of the deployment cost.
• Ownership and security. Show ownership and security interests that hold in the data center's jurisdiction.

Refinancing deployed GPUs follows the same process, starting with the contract. The GPUs keep running for your customers while they secure the loan.

Deals stall when one of these is missing: uncontracted capacity without documented revenue, a facility that can't be insured, or equity that isn't in place. When a deal falls short, Compute IQ shows which of these to close before you resubmit. For the full sequence from application to funding, see the GPU financing process.

Conclusion

GPU loan underwriting asks four questions: who pays, under what contract, who runs the cluster, and whether amortization outruns depreciation. The customer and the contract answer most of them, and disclosed deals at CoreWeave, Nebius and Lambda price that credit directly.

USD.AI is built for the Compute Middle Market, the GPU cloud operators between startups and hyperscalers. Compute IQ scores every deal the same way, and every financing closes on the same structure. Capital stays in escrow until the servers are installed and verified.

Start with your offtake contract. Talk to USD.AI about the deployment you're planning or the GPUs you already run.

FAQ

What is GPU-backed debt?

GPU-backed debt is a loan secured by GPUs, together with the related customer contracts and revenue accounts. It is repaid from contracted compute revenue, and the GPUs are the collateral if that revenue fails.

Can you borrow against GPUs you already own?

Yes. Deployed GPUs qualify for refinancing through the same GPU loan underwriting used for new deployments, starting with the offtake contract. The GPUs keep running for your customers while they secure the loan.

How much will a GPU loan cover?

Financing through USD.AI covers up to 80% of the deployment cost (loan-to-value). The operator contributes at least 20% of the cost as equity.

Does on-demand revenue qualify?

Yes. Documented on-demand revenue qualifies for GPU financing through USD.AI, and Compute IQ prices it to reflect its shorter commitment compared with a term contract.

Can a first-time operator qualify?

Yes. A first-time operator qualifies when a creditworthy customer and a well-drafted offtake contract support the deal. The hardware, equity and data center requirements still apply.