GPU Loan Vs Lease: How to Choose for Your Next GPU Cluster
You have a signed offtake contract and a purchase order for the servers, and the question is how to pay for them. You can borrow to own the GPUs, or pay a lessor to use them for a set term. The GPU loan vs lease decision shapes your cash at closing, your balance sheet and who keeps the hardware's value when the contract ends.
GPU Loan Vs Lease: the Short Answer
Choose a GPU loan if contracted or documented customer revenue covers debt service over 3–5 years and you can fund at least 20% equity. A loan lets you own the servers and keep their residual value. Choose a GPU lease if your workload horizon is short or uncertain, or you want the lessor to carry residual-value and refresh risk. A lease also fits when you can't fund the equity and accept paying more for that flexibility.
A GPU loan is asset-backed debt: you buy and own the servers, and lenders hold a first-priority claim on the hardware until it's repaid. In a GPU lease the lessor owns the GPUs, and you pay to use them.
The deciding difference is who owns the hardware and its residual value when the term ends. Either structure becomes financeable through the customer contract that pays for it, and GPU financing through USD.AI is underwritten on that contracted cash flow.
GPU Loan Vs Lease at a Glance
The table compares both structures, using USD.AI's published loan terms on the loan side.
| GPU loan | GPU lease | |
|---|---|---|
| Who owns the GPUs | You, from closing; lenders hold a first-priority claim on the hardware | The lessor, for the full term |
| Upfront cash | At least 20% equity, plus a three-month debt-service reserve funded at closing | Usually lower, set by the lessor |
| Financing terms | Up to 80% loan-to-value; fixed or floating rate; 3–5 years with straight-line monthly amortization | Set by the lessor; payments price in its financing cost, margin and residual risk |
| Residual value | You keep it, and you carry the risk that it falls | The lessor keeps it and carries that risk |
| End of term | You own the servers outright | Return, renew, or buy at fair market value or a nominal price |
| Total cost over useful life | Usually lower in total dollars if you hold and monetize the hardware | Usually higher if you'd have held the hardware; lower if a new generation reprices it |
| Balance sheet | GPUs as an asset, the loan as debt | Right-of-use asset and lease liability under ASC 842; IFRS 16 uses a single lessee model |
| Tax depreciation | Typically the borrower, as owner | Typically the lessor, in a true lease |
| What the financier underwrites | The operator, the customer and the contract, with the GPUs as security | The lessee's credit and the lessor's own residual-value estimate |
Ownership and Residual Value: Who Carries the Risk of the Next GPU Generation
In a loan you keep the hardware and whatever it's worth when the term ends. In an operating-style lease the lessor keeps it, along with the risk that a newer generation reprices it.
Tony Trzcinka, senior portfolio manager at Impax Asset Management, told Reuters in October 2026 that "Banks typically underwrite GPUs over a 3-4 year depreciation schedule." He contrasted that with Nvidia, which he said "argues top-tier GPUs can earn revenue for a decade." The same report noted that Nvidia has said some deals could carry a residual-value guarantee of no more than 25%.
The value of older GPUs has held up better than lenders' 3–4 year schedules assume. On SemiAnalysis Weekly in October 2026, Jordan Nanos said one-year H100 deals are "approaching the prices that they were signed in the first half of 2023 today, like three bucks an hour." He said buyers are renewing contracts on four-year-old H100s at about double their earlier price, and that some providers who signed one-year deals near $1.30 an hour in late 2025 have since renewed at about $2.50.
Residual value is upside on top of the contract. Daniel Nishball said on the same episode that 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," and called what comes after the term "like a bonus." In a loan that bonus belongs to the operator, and in an operating-style lease it belongs to the lessor.
The loan term is how a loan manages that risk. Straight-line monthly amortization over 3–5 years pays principal down every month, so the balance falls while the hardware ages. A shorter term, inside the 3–4 year view Trzcinka described, leaves you owning whatever value remains at payoff.
A lease handles the same risk by pricing it. The lessor builds its residual-value estimate into your payments. With a purchase option, you buy the servers at the end of the term on the stated terms. Shifting refresh risk to the lessor has real value when the next generation arrives mid-contract.
Cost and Cash: What You Pay Upfront and over the Term
A loan needs more cash at closing, which on USD.AI's terms means at least 20% equity plus a three-month debt-service reserve. A lease usually needs less upfront. In our view it costs more in total dollars if you'd have held the hardware for its useful life. Its payments carry the lessor's financing cost, margin and residual risk.
The GPU financing process starts with customer revenue, a hardware purchase order, colocation and the equity contribution, all lined up before you apply. Every deal is then scored by Compute IQ on the operator, the customer's credit and the strength of the offtake contract.
Consider an illustrative deployment of 20 eight-GPU NVIDIA B300 servers (specs on our NVIDIA B300 reference page), with assumptions that are ours. At about $600,000 per server, our Q4 2026 pricing assumption, server cost comes to $12.0 million. A loan of up to 80% covers $9.6 million, and your equity of at least 20% comes to $2.4 million.
On a three-year term, the short end of the range, straight-line amortization repays about $267,000 of principal a month plus interest. The three-month reserve then holds about $800,000 of principal, plus three months of interest. The example covers server cost only, and interest depends on the rate, so we don't state it.
Rental pricing rewards commitment, and that sets the revenue behind either structure. The same provider's B200 term curve put September 2026 rentals at $5.69 per GPU-hour spot, $5.39 for 12 months and $5.03 implied a year out. The customer contract you sign sets the revenue that services either a loan or a lease, so match the financing term to the contract term.
Under IRS Publication 946, depreciation goes to the property's owner, which is typically the borrower in a loan and the lessor in a true lease. Guidance from the IRS confirms the permanent 100% additional first-year depreciation deduction, known as bonus depreciation, for eligible property acquired after January 19, 2025. The same publication caps Section 179 at $2,560,000 for tax years beginning in 2026, reduced once qualifying property placed in service exceeds $4,090,000. At cluster scale bonus depreciation is the relevant tool; confirm how it applies to your structure with a tax adviser.
Balance Sheet and Structure: How Each Lands on Your Books
Neither option keeps GPUs off the balance sheet under current US GAAP. Under ASC 842, FASB requires lessees to recognize lease assets and lease liabilities for all leases, finance and operating, other than short-term leases. The IFRS Foundation's IFRS 16 uses a single lessee model that puts leases over 12 months on the balance sheet.
Under the same standard, a lease that transfers ownership, or carries a purchase option you're reasonably certain to exercise, is a finance lease. That puts lease-to-own closest to a loan. A sale-leaseback, where you sell hardware you own and lease it back, isn't treated as a sale when the leaseback is a finance-type lease. Classification is judgment-based, so bring your auditor in before you sign.
Two October 2026 deals show how lease structures use special-purpose vehicles (SPVs), separate companies set up to hold the chips. Reuters reported that Amazon was in talks to move about $8 billion of Nvidia Grace Blackwell chips into an SPV and lease them back. The report cited the Financial Times and described talks with outside investors.
Bloomberg reported that Broadcom arranged a $63 billion deal in which a new SPV buys Broadcom chips and leases them to Anthropic. Broadcom provides residual value support. These structures rest on very large, highly creditworthy lessees and sponsors. Most operators in the Compute Middle Market, the GPU cloud operators between startups and hyperscalers, don't fit that profile.
An asset-backed loan uses the same ring-fencing and leaves ownership with you, which is how a USD.AI loan is structured. With financing through USD.AI, an SPV under your holding company holds the GPUs, the customer contracts and the revenue accounts. Lenders hold a first-priority claim on the GPUs, contracts and revenue.
Loan capital is held in escrow until the servers are installed and independently verified. The GPUs stay deployed and earning while they secure the loan, and a dedicated servicer manages each loan for its lenders for the full term. Every loan carries property and casualty insurance on the hardware once escrow releases, and loans with non-investment-grade or on-demand revenue also carry a residual-value warranty.
When to Choose Which
Many operators rent capacity first and move to ownership as contracted demand grows. In a GPU loan vs lease decision, contract quality decides where you sit in that cycle. Compute IQ scores the operator, the customer's credit and the strength of the offtake contract.
- A signed multi-year customer contract on current-generation GPUs: Take the loan. Contracted revenue services the debt, 3–5 year amortization fits the contract, and you keep the residual value.
- Short, uncertain or burst demand: Lease or rent and let the lessor carry refresh risk, because debt amplifies losses when demand is weak.
- Capital tied up in a running cluster: A sale-leaseback releases cash but hands over ownership. Refinancing releases capital and keeps the GPUs yours.
Financing for neocloud operators through USD.AI covers new deployments and refinances GPUs already in service, which stay deployed and earning. On-demand revenue still supports a loan through GPU-as-a-Service financing, priced to reflect that the revenue isn't contracted. Either way, the plan to finance a GPU cluster starts with the customer contract.
FAQ
What is GPU lending?
GPU lending is a loan used to buy GPU servers, secured by the hardware and repaid from the revenue the GPUs earn. The borrower owns the GPUs, while lenders hold a first-priority claim on them until the loan is repaid.
Is financing more expensive than leasing?
Financing usually costs less if you hold and monetize the hardware for its useful life, since lease payments carry the lessor's margin and residual risk. A lease can cost less when you'd otherwise be left holding hardware that a new generation has repriced.
What is the downside of a lease to own?
A lease to own typically costs more in total than an equivalent loan and commits you for the full term. You also decide on the purchase when the GPUs may already be a generation behind. If it transfers ownership or carries a purchase option you're reasonably certain to exercise, ASC 842 makes it a finance lease, booked much like debt.
Can I refinance GPUs I already own?
Yes, financing through USD.AI also covers GPUs already deployed and earning. Refinancing releases capital while you keep ownership and the residual value, which a sale-leaseback hands to the buyer. It needs the same customer contract and ownership records as a new deployment.
How much of a GPU deployment can a loan cover?
Through USD.AI, loans cover up to 80% of deployment cost, with the operator contributing at least 20% as equity. Rates are fixed or floating, terms run 3–5 years with straight-line monthly amortization, and financing starts from about $5 million. Collateral sits in Tier III or Tier IV data centers.
