What Lenders Look for in GPU Infrastructure Projects

Why the questions come in this order

GPU hardware can lose value quickly, so a lender usually reads the project from the cash flow backwards: who pays, under what contract, into which entity, from which site, for what hardware. The hardware and its resale value come in as support, not as the foundation. That ordering is general market practice. Specific lenders weigh the points differently, and none of what follows is an offer or a promise of terms.

A note on numbers. Advance rates, pricing and coverage ratios vary by deal and are not reliably public. A commercial guide to lender data rooms says as much about advance rates. We therefore do not publish a table of typical ratios here, because we have not found a source we can stand behind.

1. The customer and the contract

This is where lenders start. They look at the customer's credit quality, the length of the term, whether there is a minimum commitment, what the customer can terminate for and what remedies apply if the operator fails to deliver capacity. They also look at concentration. CoreWeave's annual report lists dependence on a limited number of customers as a risk factor and reports that approximately 67 percent of its 2025 revenue came from its largest customer. It is a public illustration of why a single large counterparty is analysed carefully, not a benchmark for your project.

For the contract mechanics, see Understanding take-or-pay GPU offtake agreements.

2. The borrower and the ring-fence

Lenders ask whether the hardware and the contracts sit in a separate, ring-fenced company, often a special purpose vehicle, and who owns it. A vendor guide to lender data rooms calls this the first question and says that nothing downstream matters if the assets are not ring-fenced. Expect documents on formation, governance and ownership, and identity checks on the ultimate owners.

3. The site and the power

The hardware has to run somewhere. Lenders read the colocation or site agreement for committed power, the date the space is ready for service, the term and what happens if the provider is late. Morgan Lewis notes that the interplay of liability caps, service credits and remedies for late delivery is among the most negotiated parts of a colocation agreement, and that providers often seek multi-year commitments, sometimes up to a decade.

The site can also be a constraint on the operator. Data Center Knowledge reported in April 2026 that colocation providers screen neocloud tenants for credit strength, and that some cap their exposure or ask for guarantees or letters of credit. A lender will therefore ask whether the site agreement itself is secure.

4. The hardware

Lenders want to see the vendor purchase agreement or quotation, the model, the quantity and the delivery schedule. Delivery is often a condition for drawing the loan, and proof of title and serial numbers matter for security. On value, Bird & Bird warns that rapid technological change can accelerate depreciation and leave collateral worth less than the loan, and Pinsent Masons notes that lenders usually assume chips lose value within three to five years, while some structures assume longer. A lender will test whether the loan still repays under a shorter life and a lower resale value than you expect.

5. The sponsor and the operator

Who will run the cluster, with what track record, and how much of their own money is going in? Lenders generally expect meaningful equity from the sponsor. If the sponsor is an investor and someone else operates the hardware, describe both. Lenders look at the operator as well as the owner. A first-time operator is not excluded, but should expect closer scrutiny of the contract and the partners.

6. The financial model

The model should show contract revenue, operating costs such as power and hosting, debt service and the result if utilisation, price or hardware life turn out worse than planned. The question behind it is whether principal and interest are repaid within the contract term. Keep assumptions explicit, and label what is contracted, what is expected and what is hoped for.

7. Insurance, accounts and the practical file

Lenders usually ask for property and business-interruption insurance and for arrangements that direct customer payments into controlled accounts. A vendor guide lists these items in the security package and the data room. One commercial diligence provider also ranks unpaid power or colocation bills, operator distress and late start-up among the main risks, and names collateral access agreements with the colocation provider as a typical mitigant. Treat that as industry commentary from a party with a commercial interest.

What to do next

Our financing overview and the page on contract-backed financing set out how we describe this to operators. For the build sequence, see Financing new GPU clusters. Lenders and other capital providers can register interest on the For lenders page.

Related guides

• How Contract-Backed GPU Financing Works
• Financing New GPU Clusters: Hardware, Colocation and Customer Contracts
• Understanding Take-or-Pay GPU Offtake Agreements
• All insights

Sources

Each source below was opened and read on 30 October 2026. Titles are in the original language. Vendor and opinion sources are marked and used only for qualitative points.

1. [1] GPU Cluster Financing Data Room: What Lenders Read Before the Term Sheet (2026). Peony (data room software vendor), 2026. Commercial vendor. https://www.peony.ink/blog/gpu-cluster-financing-data-room
2. [2] CoreWeave, Inc. Form 10-K for the fiscal year ended 31 December 2025. CoreWeave, Inc. (filing with the U.S. SEC), 2026. Primary document. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm
3. [3] Activating the Colo Model: What Really Matters in Data Center Colocation Agreements. Morgan Lewis (DataCenterBytes), 2026-05. Law firm publication. https://www.morganlewis.com/blogs/datacenterbytes/2026/05/activating-the-colo-model-what-really-matters-in-data-center-colocation-agreements
4. [4] Neocloud Storm Gathers as Data Center Deals Stall Over Credit Risk. Data Center Knowledge, 2026-04-06. Trade or financial press. https://datacenterknowledge.com/cloud/neocloud-storm-gathers-as-data-center-deals-stall-over-credit-risk
5. [5] GPU-Based Financing in the Global Data Center Market: A New Standard for Large-Scale Investment Structures. Bird & Bird, 2025. Law firm publication. https://cm.twobirds.com/en/insights/2025/germany/gpubased-financing-in-the-global-data-center-market-a-new-standard-for-largescale-investment-structu
6. [6] AI chip financing: can a leasing model solve the GPU supply problem?. Pinsent Masons (Out-Law), 2026-08-17. Law firm publication. https://www.pinsentmasons.com/out-law/analysis/ai-chip-financing-can-a-leasing-model-solve-the-gpu-supply-problem
7. [7] The risks in GPU-backed lending, ranked. American Compute (commercial provider of diligence services), 2026. Commercial vendor. https://www.amcompute.com/gpu-risk