The GPU Financing Process, Step by Step: from Customer Contract to Funded Cluster
How the GPU Financing Process Works
A GPU loan is repaid by the customer contract that rents the hardware, so the GPU financing process starts with that contract. From there it runs through seven steps, beginning with assembling the contract, purchase order, colocation agreement and equity, then applying and passing origination review. Compute IQ scoring, documentation, and escrow and funding follow, and servicing runs for the life of the loan. The GPUs secure the debt, and the contracted revenue pays it down.
GPU financing, asset-backed debt secured by the GPUs and underwritten on the cash flow of the contracts that rent them, is conventionally arranged deal by deal. GPU financing through USD.AI uses the same sequence and document set for every deal. An operator on its first cluster goes through the same steps as one on its tenth, with the same documents at each stage.
In July 2026, Morgan Stanley's AI financing analysis projected that a handful of the largest builders could add more than 30 GW of capacity over two years, driving roughly $2 trillion of cash capex. A gigawatt needs roughly $12 billion for the shell and often more than twice that for chips and racks, so most of the spending goes into the hardware that GPU financing covers.
This guide is for GPU cloud operators, from a first deployment with a signed contract to multi-site operators adding capacity. The process is built for the Compute Middle Market, the GPU cloud operators between startups and hyperscalers, and larger operators follow the same steps. Comparisons of financing types, the loan structure and data center requirements are covered in other articles.
The Seven Steps in the GPU Financing Process
Each step has an owner and a deliverable, and each depends on the last: review needs the step 1 package, and escrow needs signed documents and confirmed equity. Every financing uses one structure: a special-purpose vehicle holding the GPUs and contracts, escrow until the servers are verified, and servicing for the full term. Our explainer on how a GPU-backed loan is structured traces each transfer inside that structure.
| Step | Owner | Deliverable |
|---|---|---|
| 1. Line up revenue, hardware, colocation and equity | Operator | Customer contract, purchase order, colocation agreement and equity |
| 2. Apply | Operator | Financing request with the step 1 package |
| 3. Origination review | Origination team | Confirmed requirements and a list of anything missing |
| 4. Compute IQ scoring | Compute IQ | Score that sets financeability, pricing tier, size and structure |
| 5. Documentation and approval | Counsel and origination team | SPV, security interest, standardized loan documents and final approval |
| 6. Escrow and funding | Escrow agent | Payment to the supplier after verified installation |
| 7. Servicing | Servicer | Debt service and monitoring for the full term |
1. Line up customer revenue, hardware, colocation and equity
You assemble four items before you apply, and every later step checks the deal against them:
- Customer revenue: an executed offtake contract, where a customer commits to set volumes at an agreed price and term, or documented on-demand revenue.
- Purchase order: a purchase order from an OEM or authorized supplier, or records of GPUs already in service for a refinancing.
- Colocation agreement: an executed agreement with a Tier III or Tier IV data center.
- Equity: at least 20% of the deployment cost.
The customer contract comes first because current-generation rental prices are expected to fall. In September 2026, Silicon Data's B200 rental data priced B200 capacity at $5.69 per GPU-hour spot and $5.39 on a 12-month term. The implied rate a year out was $5.03, so the curve already points down. An operator selling capacity by the hour relies on revenue holding up against that curve, while a signed contract fixes the price for its term.
The H100 shows how far rental prices can move within a single loan term. The SemiAnalysis H100 index, a market index, fell from $6.62 per GPU-hour in the second half of 2023 to about $2.80 by July 2025. It held near that level through March 2026, while one-year contract prices recovered from $1.45–$1.95 in October 2025 to $2.10–$2.70 in April 2026. The index more than halved, then firmed as on-demand capacity sold out in early 2026. That swing is why the deals we review put the contract before the hardware.
The hardware side of the package gives the system configuration, quantity, supplier, and power and cooling readiness at the site. The systems have to be on the approved collateral list, which as of October 2026 covers NVIDIA B200, B300, GB200, GB300 and Vera Rubin (VR200) systems and AMD Instinct MI455X and MI430X. A complete 8-GPU NVIDIA B300 server costs about $600,000 at Q4 2026 pricing, about $75,000 per GPU, before networking, racks, installation and tax. Our NVIDIA B300 hardware reference covers financing for that system.
2. Apply
You submit the financing request with the step 1 package, and the borrower is the business that owns, or will own, the GPUs. A hardware vendor, host or marketplace can start the deal through introductions from infrastructure partners, and the GPU owner then applies. That keeps borrower identity clear in hosted and marketplace setups, where the site and the hardware can have different owners.
Financing starts at about $5 million, with larger facilities available. An application opens a review, and the terms follow Compute IQ scoring.
3. Origination review
The origination team confirms the four step 1 requirements: an approved purchase order, confirmed equity, an executed colocation agreement, and an offtake contract or qualifying on-demand revenue. Anything missing is identified at this stage, before the file moves to scoring. Anti-money-laundering and know-your-customer (AML/KYC) checks on every ultimate beneficial owner, the people who ultimately own or control the company, also run at this stage.
Most deals drop out at review: of 394 deals worth $13.0 billion we reviewed by June 2026, 244 worth $6.0 billion did not qualify. The other 150, worth $7.0 billion, qualified. Of those, 56 worth $2.7 billion were sent a term sheet, and 26 worth $817 million signed one.
The deals that close have performed: as of October 2026, 28 loans have been originated through USD.AI with zero defaults. Together they financed 5,056 GPUs, for about $420 million of cumulative origination.
4. Compute IQ scoring
Compute IQ, USD.AI's agentic underwriting system, scores every deal on a 100-point scale across three pillars: the operator, the customer and the contract. The operator is assessed on its track record, GPU deployments to date, financial strength, cash runway and equity backing. The customer is assessed on its cash runway, EBITDA, funded liquidity, leverage and operating history, and the contract on its tenor against the loan term, its termination protection and whether payment is unconditional (take-or-pay). The score sets the facility size, pricing tier and structure, and a deal that scores below 60 isn't financeable.
Scoring follows review because it uses the documents that review confirms. Our page on how Compute IQ scores deals covers the method in detail.
5. Documentation and approval
Counsel documents every financing on the same standardized set. A special-purpose vehicle (SPV), a company set up only to hold the financed assets, holds the GPUs, the customer contracts and the revenue accounts. The SPV is bankruptcy-remote, so its assets stay out of reach of your operating company's creditors. The lender group holds step-in rights at the data center and a first-priority security interest, a senior legal claim on those assets.
Loans cover up to 80% of the deployment cost, and you contribute at least 20% as equity. Rates are fixed or floating, and terms run 3–5 years with straight-line monthly amortization. A three-month debt-service reserve is funded at closing.
You can't unilaterally redirect purchase funds, substitute collateral, move equipment or divert revenue. Those controls keep the loan, the hardware and the contract that repays it tied together for the full term.
Before the loan is signed, the origination team checks the remaining conditions and gives final approval. At signing, the capital is committed and held in escrow.
Illustrative example. Assume 100 NVIDIA B300 servers at about $600,000 each at Q4 2026 pricing, financed on a three-year term, the short end of the range. The figures show how the terms size a loan, and they cover server cost only.
| Item | Assumption | Amount |
|---|---|---|
| Server cost | 100 servers at about $600,000 each | $60 million |
| Loan | Up to 80% of cost | Up to $48 million |
| Equity | At least 20% of cost | At least $12 million |
| Monthly principal | Three-year straight-line amortization | About $1.33 million, plus interest |
| Debt-service reserve | Three months, funded at closing | About $4 million of principal, plus three months of interest |
The $34M facility in Sweden that USD.AI published in June 2026 uses this structure for NexGen Cloud's NVIDIA B200 GPUs over three years.
6. Escrow and funding
The SPV places the order, and the committed loan funds sit in escrow while the OEM builds and ships the servers to the data center. There the servers are installed, burned in and independently verified online, and escrow then pays the OEM or supplier. Your equity is confirmed before any capital releases. If the servers aren't verified online within a 120–150 day longstop, the deadline set for verification, the escrowed capital is returned.
For you, the money and the hardware move together: the supplier is paid for running servers, so the collateral exists on the day capital moves. Loan capital is never paid out against hardware that hasn't arrived.
7. Servicing for the full term
Servicing stays in place for the full term of the loan. Your customer pays into the SPV's controlled revenue account, the SPV pays principal and interest, and remaining revenue is available to your operating company. Telemetry tracks uptime and server health throughout.
A three-year loan is fully repaid inside every common measure of GPU life. Epoch AI's chip lifespan data puts the median time leading AI chips stay in frontier training at about 3.9 years as of December 2025. CoreWeave's FY2025 annual report depreciates its GPUs over six years. Meta's FY2025 10-K filing uses 5 years to 5 years 6 months for servers and network assets, and both company figures are accounting lives.
Terms longer than three years reach past the training median. They are matched to the tenor of the offtake contract, so contracted revenue covers repayment for the whole term.
Once seasoned and performing, a financing may be sold to an institutional buyer through GLX, and nothing changes for you as the borrower. GLX connects institutional capital to compute, and the capital returned from a sale funds new financings.
What Slows a GPU Financing Down
Almost every delay starts with a gap in the step 1 package, because review moves only as fast as the documents in front of it. The most common gap is the customer contract: a missing or unsigned contract stops the file. Letters of intent without committed volumes don't show the revenue that repays the loan, so they hold the file in the same way.
An unsigned colocation agreement also holds the file at review, since the executed agreement is one of the four requirements. Sites that depend on new on-site generation can wait years for power. Wood Mackenzie turbine research put gas turbine lead times at six years in April 2026, with order books committed through 2027.
Purchase orders slow a deal when they aren't itemized, or when they include soft costs such as software and subscriptions, which aren't financeable hardware. Equity has to be confirmed before escrow releases capital, so unconfirmed equity holds funding even after the documents are signed.
Through June 2026, deals under our original structure averaged 121 days from intro call to close: 36 to term sheet and 85 to close. Under the current escrow structure and secured loan agreement, the average was 60 days, with 19 to term sheet and 41 to close. These averages describe our past deals and set no timeline for yours.
Seven deals worth $427 million expired after signing a term sheet, before closing, over the same period. A complete step 1 package is what keeps a signed term sheet from expiring.
FAQ
What is GPU-backed debt?
GPU-backed debt is a loan secured by GPUs and repaid from the revenue they earn under customer contracts. The lender group holds a security interest in the hardware for the full term.
What is the typical approval timeline for GPU-backed loans?
Deals under USD.AI's current structure averaged 60 days from intro call to close through June 2026. A deal's own timeline depends mostly on how complete its step 1 package is.
How much of a GPU deployment can be financed through USD.AI?
Financing through USD.AI covers up to 80% of the deployment cost, with at least 20% from you as equity. Compute IQ scoring sets the final loan size within that limit.
Can I finance GPUs that are already deployed?
Yes. Financing through USD.AI covers new deployments and refinances GPUs already in service. Records of the deployed GPUs and proof of ownership take the place of a purchase order.
Who is the borrower when a host or marketplace is involved?
The borrower is the business that owns, or will own, the GPUs. A host, marketplace or hardware vendor can introduce the deal, and the owner applies.
If you have a customer contract and a deployment plan, we can take you through each step for your cluster. Talk to USD.AI
