Home / Insights / Compute

Explainer

GPU financing structures explained

How leases, loans, sale-leasebacks, vendor and milestone financing, and contract-backed facilities pay for GPU systems, and what lenders check before they commit.

Published by
Rillor
Published
Reading time
5 min read

A Supermicro server with eight B200 GPUs was listed at $384,228 at Newegg, out of stock, when viewed on 9 October 2026 [27]. GPU cloud operators that file with the SEC depreciate such equipment over five to six years [8, 13, 14]. Financing spreads that cost across the years the hardware earns.

This explainer covers six structures. They differ in who owns the hardware, what the lender relies on and who carries the value of the equipment when the term ends.

Structure Who owns the hardware What the lender or lessor relies on At the end of the term
Fair-market-value lease Lessor Lease payments and the residual value Buy at fair value, renew or return
$1-buyout lease Lessor until the final payment Lease payments covering the full cost The lessee owns the systems
Equipment loan Borrower, from the start Loan payments and a security interest in the hardware Loan repaid, lien released
Sale-leaseback Lessor, after buying from the operator Lease payments and the hardware Depends on the leaseback terms
Vendor and milestone financing Buyer; the vendor may hold a security interest The vendor's terms, or funds released at delivery and commissioning milestones Repaid within the vendor's term
Contract-backed financing Borrower, often a dedicated subsidiary The customer's contract payments and the hardware Repaid from the contract's payments

Sources: [1, 2, 3, 5, 6, 8, 13, 17].

Leases: fair market value or $1 buyout

The Equipment Leasing and Finance Association describes a fair-market-value lease as a lease to use equipment and a $1-buyout lease as a lease to own it [1]. Payments on a fair-market-value lease are lower because the lessee does not pay for the full cost of the equipment; the lessor expects to recover the rest by selling or re-leasing it [1]. A $1-buyout lease finances the whole cost and gives the lessee ownership for tax purposes [1].

The difference matters in law and in accounting. Under the Uniform Commercial Code, a lease the lessee cannot cancel, with an option to buy for nominal consideration, creates a security interest [2]. Under U.S. lease accounting, a lessee records a finance lease if any of five tests is met. Among them: ownership transfers, a purchase option is reasonably certain to be exercised, the term covers a major part of the asset's economic life, or the payments amount to substantially all of its fair value [4].

Equipment loans

With a loan, the borrower owns the systems from day one and the lender holds a security interest, perfected in most cases by filing a financing statement [3]. Original equipment manufacturers lend this way too. CoreWeave reported equipment financing from manufacturers with a notional balance of $4.8 billion at the end of 2025, secured by the financed equipment, with terms of one to three years [8].

Sale-leaseback

An operator that already owns systems can sell them to a lessor and lease them back. WhiteFiber's filing describes selling 96 AI servers, equal to 768 GPUs, to a third party and leasing them back for three years to finance a three-year customer agreement [7]. Accounting decides what the deal is. If the seller holds a fixed-price option to buy the systems back, the transfer is not a sale: the seller keeps the equipment on its books and records the proceeds as a financial liability [5, 6].

Vendor and milestone financing

Nscale generally pays suppliers in installments: a deposit with the purchase order, progress payments before delivery and a final payment on delivery [13]. Lenders can follow the same milestones. Lambda's October 2026 facility releases funds as clusters reach commissioning milestones, so capital is drawn only as infrastructure enters service [17]. Suppliers can also support demand. NVIDIA agreed in September 2025 to buy CoreWeave's residual unsold capacity through April 2032 under a $6.3 billion order [20].

Contract-backed financing

Here the loan rests on a signed customer contract. CoreWeave's March 2026 facility of $8.5 billion sits in a subsidiary, is secured by its assets and equity, carries a parent guarantee limited to specified bad acts, and requires a debt service coverage ratio of at least 1.15x [10]. Rated A3 by Moody's, it priced at SOFR + 2.25% [11]. Nscale's August 2026 facilities mature on the date the customer makes its final scheduled payment, if that comes first [13]. Lambda's August 2026 term loan amortizes fully on a schedule aligned with the contracted cash flows and the useful life of the GPUs [16].

What lenders underwrite

Five things recur in the filings and credit research.

  • Customer credit. CoreWeave's 2024 facility charged 6.00% to 6.50% over SOFR on investment-grade contracts and 13.00% on others [9]. IREN's financings reported in August 2026 cost 6.0% with an investment-grade customer and 9.0% with non-investment-grade customers [15].
  • Term against life. Nscale's facilities mature by the customer's final scheduled payment, and Lambda's loan amortizes with the contracted cash flows and the GPUs' useful life [13, 16].
  • Prepayment. IREN's investment-grade financing and customer prepayments together fund 96% of the related GPU spending [15].
  • Coverage. SemiAnalysis reported that lenders to backstopped operators look for coverage of at least 1.3x, typically a loan-to-value ratio of 70% to 80% [21].
  • Collateral value. CoreWeave's earlier facilities required quarterly paydown to the lower of two limits, one based on projected contracted cash flows and one on the depreciated purchase price of the GPU servers [9].

Residual value

The open question in every structure is what the hardware earns after the first term. Useful lives for GPU servers in operators' filings run from five years at Nebius [14] to six at CoreWeave [8]. Amazon shortened the life of a subset of its servers from six years to five from January 2025, citing the pace of technology development in AI [25]. KBRA framed the credit question as whether customers keep using the capacity at an attractive price after the original contract, which goes beyond whether the GPUs still work [23]. Bloomberg reported in October 2026 that GPU loans made up about 15% and 20% of the collateral in two equipment securitizations [24].

Read the full report

The full report, with the legal and accounting tests, a table of GPU facilities from 2024 to 2026 and the complete source list, is available as a PDF.

The full report

GPU financing structures explained

The research behind this article, with every source listed. 17 pages.

  1. Summary
  2. Background
  3. Method
  4. Findings
  5. Implications
  6. Limits of this analysis
  7. Sources

Notices

Rillor is not a registered investment adviser or commodity trading advisor and does not provide investment or trading advice. Research, forecasts, data and software described on this site are for research and engineering use. Nothing here is an offer or recommendation to buy or sell any security, commodity interest or digital asset. Past or simulated results do not indicate future results.

NVIDIA, AMD and related product names are trademarks of their owners and are used only to identify products. No affiliation or endorsement is implied.

Ask about the data behind this report.

Talk to us