Who Finances GPU Clusters
Who finances GPU clusters is the capital map on the Neoclouds layer — GPU-backed lenders, OEM lessors, residual offtakers, and customer prepay. Roles only. No tip sheet.
EXPLAINED · Players · Last reviewed: 2 Oct 2026 · Next review: after next GPU-backed facility print
Who finances GPU clusters is the roles map for the capital that turns accelerators into live halls: GPU-backed lenders who take the servers and the offtake as collateral, OEM lessors who finance the metal itself, residual-capacity offtakers who underwrite unsold hours, and customer-prepay counterparties who fund a slice of capex before the rack is energised.
In short. Who finances GPU clusters is a roles map on the Neoclouds layer: delayed-draw lenders clear against GPUs and contracted cash flows; OEM lessors finance the equipment; residual offtakers (a vendor buying leftover hours) backstop unused capacity; customers who prepay cash-in a slice of the build. The stuck step sits where the note, the lease, and the hall bind before billed hours clear — not a tip sheet.
The operator map is GPU cloud providers. Mechanism siblings: take-or-pay GPU contracts and remaining performance obligation. Versus: neocloud vs hyperscaler.
Who finances GPU clusters?
Who finances GPU clusters splits into four roles along one bind. GPU-backed lenders advance delayed-draw term loans secured by servers, related infrastructure, and customer contracts. OEM lessors and equipment financiers put the metal on a note without waiting for a public loan syndicate. Residual-capacity offtakers — often the GPU vendor — agree to buy unsold hours so the offtake clock still turns. Customer prepay and long-dated offtake pull cash in before, or as, the hall is built. Missing any one role leaves CapEx waiting — even when backlog looks full.
| Role | Job | Dated example (labelled) |
|---|---|---|
| GPU-backed lenders | Delayed-draw term loans secured by GPU servers, related infrastructure, and contracted cash flows — often in a special-purpose borrower | CoreWeave DDTL 4.0: $8.5B facility, first IG-rated GPU-backed financing, A3 / A (low); $2.837B outstanding as of 30 Jun 2026 (10-Q). Lambda $1.008B IG delayed-draw term loan closed 1 Oct 2026 |
| OEM lessors / equipment financiers | Finance the servers and licenses so the operator does not pay cash up front for every tray | CoreWeave OEM financing outstanding $4.8B as of 30 Jun 2026 (vs $3.8B at 31 Dec 2025); software-licence financing $347M (10-Q) |
| Residual-capacity offtakers | Buy leftover unsold hours so a GPU-backed loan still has a cash clock if the primary customer does not fill the hall | NVIDIA order form initial value $6.3B under the 10 Apr 2023 MSA: NVIDIA is obligated to purchase residual unsold capacity through 13 Apr 2032 (CoreWeave 8-K, 15 Sep 2025) |
| Customer prepay / offtake as credit | Prepaid reserved capacity, or a long-term investment-grade offtake, that lenders treat as the cash backbone of the build | Nebius: 70% of Q2 2026 deals included prepayments covering 50–60% of associated capex; ~$775M asset-backed facility in July 2026 at SOFR + 2.50% on the back of a long-term IG agreement (Q2 2026 letter) |
Who are the GPU-backed lenders?
GPU-backed lenders are not buying billed hours. They advance cash against collateral: the servers, related kit, contracted cash flows, and — in later facilities — the matching data-centre leases. CoreWeave’s Form S-1, filed 3 March 2025, said it had “pioneered GPU infrastructure-backed lending” and had raised over $14.5 billion in debt and equity across 12 financings, including $7.6 billion of committed GPU infrastructure-backed debt led by Blackstone and Magnetar.
That template scaled. On 31 March 2026 CoreWeave closed an $8.5 billion delayed-draw term loan (“DDTL 4.0”), rated A3 by Moody’s and A (low) by DBRS — company-described as the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract. The borrower is CoreWeave Compute Acquisition Co. VIII, LLC, non-recourse to the parent except for customary carve-outs. Pricing: floating at SOFR + 2.25% and fixed at about 5.9%. Maturity: March 2032. Initial capacity about $7.5 billion, with room to $8.5 billion as assets stabilize (CoreWeave IR, 31 Mar 2026; DDTL 4.0 overview).
The 10-Q for the period ended 30 June 2026 is the labelled stock. DDTL 4.0 outstanding was $2.837 billion ($1.4 billion floating-rate and $1.5 billion fixed-rate in the same note). Our check: $1.4 billion + $1.5 billion = $2.9 billion; $2.837 billion is the carrying print. Our check: $2.837 billion ÷ $8.5 billion ≈ 33% drawn. CCAC VIII collateral: $3.3 billion of non-current assets, primarily property and equipment, plus $155 million current. Our check: $3.3 billion ÷ $2.837 billion ≈ 1.16× — a labelled snapshot, not a recovery promise.
Lambda printed the same role on 1 October 2026: a $1.008 billion delayed-draw term loan, A (low) from Morningstar DBRS and Baa1 from Moody’s, 6.78% fixed, maturity 30 May 2033, secured by the GPU servers funded by the facility and the contracted cash flows, supporting three committed deployments with two investment-grade offtakers (Lambda). The delayed-draw, Lambda wrote, funds “only as infrastructure enters service.” The lender does not fund a dark hall.
Who are the OEM lessors?
OEM lessors and equipment financiers sit beside the public loan, not inside it. CoreWeave’s 10-Q (period ended 30 June 2026) reports OEM Financing Arrangements with an outstanding balance of $4.8 billion as of 30 June 2026, up from $3.8 billion at 31 December 2025. Software-licence financing sat at $347 million. Our check: $4.8 billion ÷ $3.8 billion = 1.26× over six months on the OEM book.
A hall landlord can sit in a related financing role. The same 10-Q describes DCSP Financing Arrangements: a data-centre service provider that designs, purchases, builds, and manages a site of up to 78 MW, plus a finance lease that commenced in April 2025 (14-year initial term; 13% imputed interest). Future contractual principal payments on that finance lease total $233 million before interest — a labelled lessor print on the shell, not the GPU tray.
The two lessor jobs stay separate. OEM finance is the metal. Hall finance is the megawatts and the cage. GPU-backed lenders often take both: DDTL 4.0’s collateral is “cloud infrastructure equipment,” the underlying customer contract, and “the corresponding data center leases” (DDTL 4.0 overview, Mar 2026).
Who underwrites residual capacity?
A residual-capacity offtaker buys the hours the primary customer does not take. On 9 September 2025 CoreWeave and NVIDIA entered a new order form under their 10 April 2023 master services agreement, initial value $6.3 billion. “In instances where the Company’s datacenter capacity is not fully utilized by its own customers, NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032,” subject to termination rights and delivery and availability requirements (CoreWeave 8-K, 15 Sep 2025). NVIDIA also supplies the GPUs and is a stockholder.
That is a credit enhancement on the same bind the GPU-backed lender is underwriting: if the hall is live and the primary offtaker does not fill it, a second cash clock is supposed to start. Delivery still gates it. An undelivered residual order does not pay a delayed-draw lender.
Who prepays — and why does offtake count as financing?
Customer prepay is cash-in before, or as, the operator performs. Nebius’s Q2 2026 letter (12 Aug 2026) said overall deal economics strengthened, “with 70% of deals including prepayments, covering 50–60% of the associated capex.” The same letter: “The $775 million secured facility we raised in July was on the back of one of these [long-term investment-grade] agreements,” at SOFR + 2.50% — the company’s first asset-backed financing. Long-term IG offtake is the credit file; the delayed-draw loan is the cash.
The balance-sheet print of that cash-in is deferred revenue. Nebius’s Q2 2026 results showed deferred revenue, current, of $979.4 million and deferred revenue, non-current, of $4,995.8 million as of 30 June 2026. Our check: $979.4 million + $4,995.8 million = $5,975.2 million, about $6.0 billion of prepaid capacity sitting as a liability until the operator performs.
Offtake that is not prepaid still finances the build when lenders treat the contract as collateral. Lambda’s 1 October 2026 facility is “backed by contracted cash flows from two hyperscale customers.” CoreWeave’s 10-Q is explicit about the other side of that trade: if counterparties do not perform, “we may have excess capacity and may remain responsible for expenditures for components, infrastructure, and data center leases and build-outs, as well as related financing that we have undertaken for which we may not receive corresponding revenue.” Take-or-pay relocates unused-hour risk after delivery. It does not delete the note.
Who gets paid — and where is the stuck step?
On the Neoclouds layer, who gets paid on the financing map tracks which role’s collateral is live.
| Step | Who clears | What has to be true |
|---|---|---|
| Signature | Offtake and residual orders print. Facility commitments print. Cash may or may not. | A contract a lender will underwrite |
| Draw | GPU-backed lenders and OEM lessors advance against eligible assets | Commissioning milestones / eligible GPU cost — Lambda’s delayed-draw rule; CoreWeave’s DDTL 4.0 borrowing conditions |
| Cash-in slice | Customer prepay becomes deferred revenue | Cash arrived; performance has not |
| Amortisation | Lenders get paid first out of contracted cash flows once the block is accepted | Energised hall, installed GPUs, accepted capacity |
| Residual | Vendor offtaker pays for unsold live hours | The hall is up; primary utilisation is short |
| Stuck step | Notes, leases, and halls wait on watts and trays | A dark hall does not amortise a GPU-backed loan |
Our check on CoreWeave’s SPV perimeter: as of 30 June 2026, debt at bankruptcy-remote subsidiaries that hold the financed infrastructure and related customer contracts was secured by $18.2 billion of non-current assets (primarily property and equipment) and $2.6 billion of current assets (10-Q). That is the collateral stack — labelled, not a recovery forecast.
The operator still has to deliver. CoreWeave’s committed contracts “start either on a fixed date or when we deliver the capacity specified in the contract” and were 98% of revenue in the quarter ended 30 June 2026 (10-Q). See take-or-pay GPU contracts and speed-to-power.
Common misconceptions
- People assume “who finances GPU clusters” means one bank writing a corporate revolver. Actually four roles share the bind: GPU-backed lenders, OEM lessors, residual offtakers, and customer prepay / offtake-as-credit. A facility headline is not the whole map.
- People assume a closed GPU-backed facility means the GPUs are paid for and live. Actually delayed-draw loans fund as clusters commission. CoreWeave’s $8.5 billion DDTL 4.0 had $2.837 billion outstanding at 30 June 2026 — about 33% drawn. Lambda’s 1 October 2026 facility funds “only as infrastructure enters service.”
- People assume take-or-pay offtake replaces the lender. Actually offtake is the cash clock the lender underwrites. If the customer does not perform after delivery, CoreWeave’s 10-Q says the operator may still be on the hook for infrastructure, data-centre leases, “as well as related financing.”
- People treat a league table of facility sizes as investment advice. Actually this page maps roles and stuck steps. Facility, OEM, residual, and prepay prints are labelled company figures — not buy/sell calls.
Not investment advice. Do your own research.
FAQ
Who finances GPU clusters?
GPU clusters are financed across four roles: GPU-backed lenders who take servers and contracted cash flows as collateral, OEM lessors who finance the equipment, residual-capacity offtakers who buy unsold live hours, and customers whose prepay or long-term offtake is the credit backbone. This is a roles map, not a ranking of “best” lenders.
What is a GPU-backed delayed-draw term loan?
A delayed-draw term loan advances as eligible GPU infrastructure is purchased and commissioned, secured by the servers and often the customer contract. CoreWeave’s DDTL 4.0 is an $8.5 billion facility (A3 / A (low)) with $2.837 billion outstanding as of 30 June 2026 (10-Q). Lambda closed a $1.008 billion IG delayed-draw term loan on 1 October 2026.
How do OEM lessors differ from GPU-backed lenders?
OEM lessors finance the equipment purchase itself — CoreWeave’s OEM financing outstanding was $4.8 billion as of 30 June 2026 (10-Q). GPU-backed lenders are usually a syndicate or institutional book taking a security interest in those assets and the offtake. One operator can use both on the same hall.
What is residual-capacity offtake?
A residual-capacity offtaker agrees to buy hours the primary customer does not use, once the hall is live. CoreWeave’s 8-K (15 Sep 2025) reports a NVIDIA order form with an initial value of $6.3 billion under which NVIDIA is obligated to purchase residual unsold capacity through 13 April 2032, subject to termination and delivery conditions.
Does customer prepay count as financing?
Yes, as cash-in before performance. Nebius’s Q2 2026 letter said 70% of deals included prepayments covering 50–60% of associated capex, and that a $775 million secured facility in July 2026 sat on a long-term investment-grade agreement. Deferred revenue (current $979.4 million + non-current $4,995.8 million as of 30 June 2026) is that cash-in until the operator performs.
Where is the stuck step?
The stuck step is delivery — energised megawatts, installed GPUs, accepted capacity. Delayed-draw lenders fund against commissioning. Take-or-pay bills after the on-switch. A signed offtake against a dark hall does not amortise the note. See take-or-pay GPU contracts and speed-to-power.
Is this a list of the best GPU lenders or stocks?
No. Second Order Players pages map roles and stuck steps. They do not rank lenders, give buy/sell calls, or treat facility-size tables as investment advice.
Related terms
- GPU cloud providers — operator roles: fleet, hours, offtake, power + halls.
- Take-or-pay GPU contracts — reserved hours that bill after delivery.
- Remaining performance obligation — contracted cloud not yet revenue.
- Speed-to-power — why watts gate the hall the note is waiting on.
Markets and stack doors
Layer: Neoclouds. Buyer layer: Hyperscalers.
Markets: US — CoreWeave and how neocloud contracts become cash · US speed-to-power. Stack: What Does CoreWeave Do? · What Does Nebius Do?.
Get the Wire — short notes when offtake, power, or GPU-backed facilities move.
How we check this
Figures come from company filings and company releases, each dated in the list below. Capacity and timing from press reports are labelled as reported, not guided. Where we work something out ourselves, the arithmetic is shown in full.
Last reviewed: 2 Oct 2026 · Next review: after next GPU-backed facility print
Spot an error? Tell us and we will correct it and note the change here.
Sources
- CoreWeave IR — DDTL 4.0 $8.5B, A3 / A (low), 31 Mar 2026
- CoreWeave DDTL 4.0 overview — collateral: equipment, contract, data-centre leases (Mar 2026)
- CoreWeave 10-Q — period ended 30 Jun 2026
- CoreWeave Form S-1 — GPU infrastructure-backed lending (filed 3 Mar 2025)
- Lambda — $1.008B delayed-draw term loan (1 Oct 2026)
- CoreWeave 8-K — NVIDIA $6.3B residual offtake through 13 Apr 2032 (15 Sep 2025)
- Nebius Ex. 99.2 — Q2 2026 letter; 70% prepay; $775M facility (12 Aug 2026)
- Nebius Ex. 99.1 — Q2 2026 results; deferred revenue as of 30 Jun 2026
S
Writes Second Order by TKN — who gets paid, where money sticks, and which step is stuck.
