GPU-backed loans: contracts, equipment and repayment
How GPU collateral and customer cash flows support lending, and where the two can diverge.
A GPU-backed loan finances eligible compute equipment or associated contracted cash flows, with the lender taking the security and recourse actually agreed. Debt capacity depends on usable cash, enforceable rights and a credible exit; the GPU model alone does not establish any of them.
- 01Equity & deposits
- 02Eligible draw
- 03Customer collections
- 04Debt service & exit
Start with the borrowing entity
Draw an entity map before sizing debt. The hardware owner, service provider, customer contracting party and site occupier may differ. Identify which party borrows, which parties guarantee, and which assets are pledged. Keep equipment security separate from a claim on customer payments and from permission to continue operating at the site.
Size the loan twice
First test operating repayment: collected revenue less power, hosting, support, tax, essential replacement spending and working capital. Then test the remaining principal against net recovery at contract expiry or default. Use delayed acceptance and a lower re-contracting price as separate stresses. A profitable service can still have a cash deficit during its ramp.
Read funding availability as a condition
A delayed draw commitment can match equipment delivery, but each draw may require eligible contracts, accepted equipment, equity and documentation. Reconcile deposits paid before eligibility with cash available outside the loan. Calculate interest on drawings and fees on the specified undrawn balance; the headline commitment is not the interest-bearing amount.
Keep comparisons on the same basis
A SOFR margin, a fixed coupon and a convertible coupon describe different financing exposures. The register preserves dated observations rather than ranking them as interchangeable borrowing costs. CoreWeave DDTL histories and IREN fixed/floating facilities provide useful starting points, but each has distinct support, security and payment mechanics.
a GPU delayed draw term loan (DDTL) is a committed facility that a neocloud draws in stages, each draw funding a batch of GPUs for a specific customer contract. Each draw is sized on that contract's projected cash flow, priced on the customer's credit and repaid over the contract term. It is the main financing product of the GPU cloud market.
Why "delayed draw"
A neocloud signs a customer contract, then waits months for GPUs to be delivered, installed and accepted. It cannot borrow the full amount on day one because the assets and cash flow do not yet exist. A DDTL gives it certainty of funding (the commitment) while lenders only fund against assets that have arrived (the draws).
The life of a GPU DDTL
- Commitment. Lenders commit a maximum amount for an availability period. The borrower pays a commitment or undrawn fee on the unused portion: 0.50% a year in CoreWeave's DDTL 3.0, 4.0, 5.0 and 5.5.
- Customer contract signed. The SPV signs, or is assigned, a capacity contract.
- Draw request. As GPUs are delivered, the borrower requests a draw. Typical conditions include: the contract is in force, the customer has signed a direct agreement, the GPUs have been delivered or accepted, the sizing DSCR is met and no default is continuing.
- Funding. The draw is sized on the projected contracted cash flow for that batch. Some facilities charge an upfront fee per draw; CoreWeave's US$3.0bn DDTL 2.1 tranche charged 1.50% upfront per draw and a 3.00% commitment fee.
- Operations. Customer payments go to controlled accounts and pay interest, principal and reserves through a waterfall.
- Repayment. Each draw amortises over the remaining contract term.
Worked example: CoreWeave's six facilities
| Facility | Date | Size | Pricing | Main point |
|---|---|---|---|---|
| DDTL 1.0 | 2023 | US$2.3bn | SOFR + 8.75% (repriced to 9.62%) | First large private credit GPU loan |
| DDTL 2.0 | 2024 | US$7.5bn | SOFR + 6.00% / 6.50% / 13.00% by customer credit tier | Pricing tiered by customer rating |
| DDTL 3.0 | 2025 | US$2.6bn | SOFR + 4.00% | Single contract; 1.40x DSCR and 0.85x contract realisation tests |
| DDTL 4.0 | 2026 | US$8.5bn | SOFR + 2.25%; fixed option Treasury + 2.00% | Bank-led; bad-acts parent guarantee only |
| DDTL 5.0 | 2026 | US$3.1bn | SOFR + 4.50% | Multiple contracts plus data centre leases; full parent guarantee |
| DDTL 5.5 | 2026 | US$2.6bn | SOFR + 5.50% | Rated Ba2 by Moody's (reported); full parent guarantee |
The trend is clear from the documents: as customer quality and market familiarity improved, margins fell by more than six percentage points in three years, and the most tightly priced facilities needed the least parent support.
Fixed or floating
Most DDTLs float over SOFR. Some offer a fixed-rate tranche for borrowers whose customer contracts are fixed price. CoreWeave's DDTL 4.0 included a fixed-rate option at Treasury + 2.00%. Nscale's Ward County (US$1.85bn) and North Carolina (US$1.2bn) facilities each combine a floating tranche at three-month SOFR + 2.375% with a fixed-rate tranche at the SOFR swap rate + 2.375%.
Prepayment
Lenders want protection against early repayment, especially as margins fall and borrowers look to refinance. CoreWeave's DDTL 5.0 carries a make-whole in year one, 2% in year two and 1% in year three.
The key negotiation points
- Draw conditions: delivered, installed or accepted? Who certifies acceptance?
- Sizing: which cash flows count, and at what DSCR?
- Customer tiers: is pricing fixed, or does it move with each customer's rating?
- Parent support: full, limited or bad-acts only.
- Concentration limits: how much can be drawn against one customer?
- Prepayment protection and refinancing flexibility.
Questions to take to the documents
- Which entity can collect the customer cash after a borrower default?
- Can a replacement operator access the hardware, software and facility?
- What principal remains at the first contract expiry, after a delayed launch?
Explore the disclosed examples
Open a register record for its dated amount, classification and source limitations. A link is not certification of all terms.
GPU financing: the full map
- GPU-backed loans and DDTLs
- Asset-level loans to an SPV, drawn as GPUs are delivered and sized on customer contracts. The core product, covered in this guide.
- GPU ABS
- Rated, securitisation-style debt on GPU capacity contracts and hardware. See GPU ABS and GPU loan vs GPU ABS.
- Borrowing bases
- Revolving or delayed draw facilities whose size moves with eligible contracts and hardware value. See GPU borrowing bases.
- Convertibles and high yield
- Group-level corporate debt for neoclouds and ex-miners. See neocloud convertible finance.
- Residual value support
- Guarantees of hardware or building value that let debt run longer than the contract. See residual value guarantees.
- Hyperscaler and chipmaker support
- Backstops, capacity commitments and equity that improve the credit. See hyperscaler support and Nvidia in the capital stack.
- Outside the US
- Security over GPUs and enforcement differ by country. See GPU financing outside the US and GPU security by country.
Questions clients ask
- What is a delayed draw term loan?
- A term loan that is committed up front but drawn in stages over an availability period. In GPU financing each draw is made as a batch of GPUs is delivered and accepted under a specific customer contract.
- How are GPU DDTLs priced?
- As a margin over SOFR that depends mainly on the customer's credit. Public CoreWeave facilities range from SOFR + 8.75% in 2023 to SOFR + 2.25% in 2026 for a facility backed by a single strong contract.
- When does a GPU DDTL repay?
- Usually over the customer contract term, from contracted cash flow, so the debt is repaid before the GPUs reach the end of their economic life.
For in-house counsel
Take it to the documents
The filed wording behind this piece, for the lender’s or the borrower’s side. Previews are open; the full text is free with InfraCompute Weekly.
- GPU loan summariesFiled GPU credit agreements side by side: advance rates, covenants and the collateral package.
- Financing clause libraryCash traps, coverage ratios, cures, step-in and GPU security, quoted from credit agreements and indentures.
- Issues list: GPU credit agreementBuild it from the lender’s or the borrower’s side, with where filed deals land.
- The legal desk for in-house counselContract library, issues lists, checklists and English law notes.
Sources for the examples
Selected public sources support the examples. The analytical framework is InfraCompute editorial work; source links do not imply a complete agreement or legal review.
- CoreWeave closes $2.6bn DDTL5.5 ↗Disclosure date: 2026-08-10
- IREN Microsoft GPU financing Form 8-K ↗Disclosure date: 2026-05-29
- CoreWeave DDTL 2.0 credit agreement (S-1 Ex 10.13) ↗
- CoreWeave DDTL 3.0 credit agreement (Ex 10.1) ↗
- CoreWeave DDTL 4.0 credit agreement (Ex 10.1) ↗
- CoreWeave DDTL 5.0 credit agreement (Ex 10.1) ↗
- CoreWeave DDTL 5.5 credit agreement (Ex 10.1) ↗
- Nscale Ward County GPU facility credit agreement (S-1 Ex 10.23) ↗
- Nscale North Carolina GPU facility credit agreement (S-1 Ex 10.24) ↗
