Skip to content
TEN Brief Ten verified stories a day 2026.11.11 KO

이 기사는 한국어로도 읽을 수 있습니다 →

Economy · 8 min read · Reference

What is GPU-backed debt — loans secured by AI chips and rental contracts, priced anywhere from 5.9% to 15%

GPU-backed debt is a loan secured by AI chip servers and by the customer contracts that rent those servers. The borrower is usually a special-purpose vehicle set up by a GPU rental company, known as a neocloud, and the lenders are private credit funds, banks and insurers. Repayment comes from the monthly fees customers pay, not from selling chips, so pricing tracks the credit quality of the customer. The first big facility of CoreWeave in 2023 cost about 15% a year; its investment-grade-rated $8.5 billion loan in March 2026 was fixed at about 5.9%. The weak point is that chips age far faster than houses or aircraft. Semafor reported on November 9, 2026, that investors were trying to sell such loans at steep discounts.

Wrapped server racks on pallets at a sunlit data center loading dock, workers seen from behind

The three lines

  • Structure — an SPV buys the chips, pledges them plus the rental contracts, and repays from customer fees
  • Pricing — SOFR+2.25–3.00% with investment-grade customers, SOFR+4.00–5.50% without (public 2025–2026 deals)
  • Risk — fast-aging collateral and no foreclosure precedent; Semafor reported discounted loan sales on November 9, 2026

Key questions

what is GPU-backed debt
**A loan secured by AI chip servers and the contracts that rent them out, repaid from customer fees.** | Item | Detail | |---|---| | Borrower | An SPV set up by a GPU rental company (neocloud) | | Lenders | Private credit funds, banks, insurers, sometimes hardware vendors | | Collateral | GPU servers plus contracted customer cash flows | | Repayment source | Monthly customer fees | | Common forms | Delayed-draw term loans (DDTL), sale-leasebacks, vendor financing |
GPU-backed loan interest rates
**Public deals range from about 5.9% fixed to about 15% floating, and the gap comes from customer credit, not the chips.** | Deal | Date | Size | Rate | |---|---|---|---| | CoreWeave DDTL 1.0 | 2023 | $2.3B | SOFR+9.62%, about 15% | | CoreWeave DDTL 4.0 (A3) | March 2026 | $8.5B | SOFR+2.25% or about 5.9% fixed | | CoreWeave DDTL 5.5 | August 2026 | $2.6B | SOFR+5.50% | | Lambda (Baa1) | October 2026 | about $1.0B | 6.78% fixed |
risks of GPU-backed loans
**The chips may age out before the loan matures, repayment depends on one or two customers, and no lender has yet publicly foreclosed on and auctioned GPUs.** | Risk | Detail | |---|---| | Depreciation | New chip generations cut rental rates and resale values of older ones | | Concentration | One loan often rests on one or two customer contracts | | Mismatch | Loans run five to six years; some contracts average about three | | Recovery | No public GPU foreclosure and auction as of October 2026 | | Floating rates | Higher benchmark rates raise interest costs directly |

GPU-backed debt is a loan secured by AI chip servers and by the contracts of the customers who rent them. It works a little like a mortgage: a GPU rental company pledges its chips to borrow the money to buy them. But the cash that repays the loan does not come from the chips. It comes from the monthly fees a customer pays to use them. That is why the price of these loans depends less on the hardware than on who has signed up to use it.

The structure became widely known in 2023, when US GPU rental company CoreWeave borrowed $2.3 billion this way, and within three years single deals reached several billion dollars. On November 9, 2026, Semafor reported that investors who had rushed to fund the data center buildout were now trying to sell whole loan portfolios at steep discounts. One person told Semafor that GPU financings in particular were being hit, on the worry that loan terms may outlast the useful life of the chips. This explainer sets out how the loans are built and where they are fragile.

1. Who borrows, who lends

RoleWhoWhat they do
BorrowerAn SPV set up by a neocloud such as CoreWeave, Lambda, Nebius, Nscale or IRENBuys and owns the GPU servers
LendersPrivate credit managers, banks, insurers and asset managersFund draws within a committed limit
CustomerAI model developers and big tech companiesRents the GPUs on a multi-year contract
SupplierChip and server makersSell the hardware, sometimes with vendor credit or backstops

A neocloud is a newer cloud company that buys GPUs in bulk and rents them to AI firms. Unlike Amazon or Microsoft, it lacks large cash flows from an existing business and must borrow most of the purchase price.

A special-purpose vehicle (SPV) is a separate legal entity created for one financing. In CoreWeave's $8.5 billion loan of March 2026, the borrower was not the parent but a subsidiary, CoreWeave Compute Acquisition Co. VIII, and the collateral was substantially all of that subsidiary's assets. The chips and the customer contract sit in one box, and only the box is pledged.

The lender base has widened. Early deals went to private credit funds at high rates. The $8.5 billion facility was anchored by Blackstone Credit and Insurance, with MUFG and Morgan Stanley as structuring agents and Goldman Sachs and JPMorgan as additional arrangers, alongside banks, asset managers and insurers, according to the company. Lambda marketed its October 2026 loan to insurers and fixed-income investors.

FormHow it worksPublic example
Delayed-draw term loan (DDTL)A limit is committed up front and drawn as servers are installedCoreWeave DDTL 1.0 to 5.5, Lambda, Nscale
Sale-leaseback or leaseA financing entity owns the chips and leases them to the userSpaceX, via an Apollo affiliate (about $40 billion of principal, per Bloomberg)
Vendor financingThe server maker is paid over timeCoreWeave OEM financing, about $4.2 billion at June 30, 2026 (CCIR)

2. Is the collateral the chip or the contract?

LayerWhat it isWhat lenders assess
HardwareGPU servers and networking gearDepreciated cost; whether it can be re-rented or sold
ContractA multi-year customer agreement and its cash flowsCustomer credit rating, term, termination rights

The weight sits on the second layer. CoreWeave described its $8.5 billion loan as the first investment-grade-rated financing secured by high-performance computing infrastructure and an associated customer contract. Moody's rated it A3 and DBRS A (low), while CoreWeave itself is rated below investment grade. The same company gets two very different ratings because the thing being rated is the customer contract inside the box.

LoanDateSizePricingCustomer credit
CoreWeave DDTL 1.02023$2.3BSOFR+9.62% (about 15%)Undisclosed
CoreWeave DDTL 2.0May 2024up to about $7.6BSOFR+6.00% for investment-grade contracts, SOFR+13.00% for othersMixed
CoreWeave DDTL 3.0July 2025$2.6BSOFR+4.00%OpenAI
CoreWeave DDTL 4.0March 2026$8.5BSOFR+2.25% or about 5.9% fixedOne investment-grade customer
CoreWeave DDTL 5.0May 2026$3.1BSOFR+4.50%Other
CoreWeave DDTL 5.5August 2026$2.6BSOFR+5.50%Other
LambdaOctober 2026about $1.0B6.78% fixedTwo investment-grade customers

SOFR is the overnight benchmark for US dollar loans. American Compute, a GPU financing firm that tracks public deals, puts 2026 floating-rate loans backed by investment-grade customers at SOFR+2.25–3.00% and those without at SOFR+4.00–5.50% since 2025. The Token Dispatch attributes the one-point gap between CoreWeave's May and August 2026 loans, which carry the same ratings, to contracts averaging about three years against a loan of roughly five. When the contract ends before the loan does, the remaining years rest on aging chips alone.

3. How much, how long, and what happens when value falls

Loan-to-value figures are rarely disclosed. What can be confirmed:

TermLevelBasis
Advance rateCoreWeave DDTL 2.0: 90% of depreciated cost for investment-grade contracts, 60% for othersCCIR (from filings)
MaturityAbout five to seven years: March 2028 (DDTL 1.0), March 2032 (DDTL 4.0), May 2033 (Lambda)Company releases, CCIR
RepaymentQuarterly amortization, or fully amortizing (Lambda)CCIR, Global Data Center Hub
Rate typeMostly floating; newer loans require swaps on at least 95% of floating balancesCCIR
Coverage testDebt service coverage of 1.15x to 1.40xCCIR

Two features differ from a mortgage. Sizing is based on depreciated cost and contracted cash, not a market appraisal, because there is no deep record of used GPU prices. And principal is repaid quickly, since a loan that shrinks more slowly than the chip's value sees its LTV rise on its own. The accounting debate over whether chips last three years or six is a separate subject; here it matters only that lenders bear the loss if the estimate proves wrong.

There is no daily margin call of the kind found in stock-backed lending, because chips have no daily price. Instead the contracts use staged triggers:

StageTriggerConsequence
1ScheduleQuarterly principal payments
2Coverage ratio below the thresholdCash is trapped inside the SPV
3Contract value below a multiple of the loan (2.40x in DDTL 5.5)Excess cash sweeps to repay the loan
4Customer gains a right to terminate, or stops paying for monthsEvent of default; DDTL 1.0 requires a lost contract to be replaced within 90 days
5DefaultLenders take the chips and the contracts

Most triggers are tied to the customer contract, not chip prices. Falling chip values show up elsewhere: in the secondary market, where a holder who wants out must find a buyer, and a buyer who doubts the collateral bids below face value. That is the scene Semafor described, though its report gives no discount figures and names no sellers. The dated events around it:

Date reportedEventSource
October 29, 2026Ares Management says on its earnings call it does not know how to underwrite the depreciation curves of GPUs and prefers physical infrastructureDealStreetAsia, Semafor
November 4, 2026Goldman Sachs and Apollo begin sounding out buyers for SpaceX's chip financing, about $40 billion of principal, asking some for up to $500 millionReuters, citing Bloomberg
November 5, 2026ING seeks bids on a dozen data center loans totaling about $500 million, including $25 million tied to a Stargate site in New Mexico; bids due November 19Bloomberg, via Reuters
November 9, 2026Semafor reports steep discounts on AI loan portfoliosSemafor

No lender has publicly foreclosed on and auctioned GPUs as of October 2026, according to American Compute. Its closest precedent is the 2022 bitcoin-mining bust, when an index of mining-machine prices fell 85% in a year. By its count, lenders who took machines back or settled for discounted cash recovered roughly 35 to 65 cents on the dollar, while those whose borrowers kept their sites, power and contracts recovered 80 to 100 cents.

4. How it compares with mortgages and aircraft finance

MortgageAircraft financeGPU-backed loan
CollateralHouse and landAircraftGPU servers plus customer contracts
Asset lifeDecades20 years or moreFive to six years on the books; real life disputed
Value pathCyclical but durableGradual declinePossible step-downs with each chip generation
Price evidenceSales records, appraisalsAppraisers, resale historyThin public record
Repayment sourceHousehold incomeAirline lease paymentsFees from one or two customers
On defaultStatutory foreclosureIn the US, repossession 60 days after an airline bankruptcy filing (Section 1110)No dedicated regime, no auction precedent
RedeploymentNoneTo another airlineTo another customer, if a data center with matching power and cooling exists
Typical term20–30 yearsAround 10 yearsAbout five to seven years

The similarity is that the collateral earns money and can be moved to another user. The differences are a short life, thin price history and concentrated customers: there are hundreds of airlines, but only a handful of buyers for tens of thousands of GPUs at once.

5. Why these loans exist

Neoclouds need billions of dollars of chips per site and lack the cash. Issuing shares dilutes owners, and ordinary corporate bonds are priced on the whole company: CoreWeave's unsecured notes carry coupons of 9% or more and traded at yields near 15% in October 2026. Put a strong customer's contract in an SPV, pledge only that box, and the same company can borrow at 5.9%. GPU-backed debt lets a weaker company borrow against a stronger customer's credit.

The cost appears as interest. CoreWeave's interest expense was $640 million in the second quarter of 2026, and the company guided to $860–940 million for the third, against guided adjusted operating income of $200–260 million. CCIR counts $35.55 billion of principal at June 30, 2026, of which only about $3.7 billion is non-recourse to the parent; the rest is guaranteed. An SPV does not, by itself, remove the parent from the hook.

FAQ

How is GPU-backed debt different from a corporate bond? A bond is priced on the whole company and usually repaid at maturity. A GPU-backed loan is secured on specific chips and contracts and repaid in installments.

Does a fall in chip prices trigger a margin call? No. Public agreements test contracted cash flow against debt payments, then trap or sweep cash if the test fails.

Has any lender seized and sold GPUs? No public case had been identified as of October 2026.

Why does this matter outside the US? The servers these loans fund also contain high-bandwidth memory and server DRAM, much of it made in South Korea. A tighter loan market could slow chip orders, though no such effect has been confirmed.

Still open: the size of the discounts, which the ING auction and the SpaceX syndication may reveal; CoreWeave's third-quarter figures; and what a seized GPU actually fetches. Our earlier explainers on GPU depreciation, AI circular financing, corporate bonds and AI capital expenditure cover the neighboring ground. This article explains a financing structure and is not a recommendation to buy or sell any asset.

Sources

  1. Semafor — AI chip financing hits a wall (2026-11-09)
  2. CoreWeave IR — CoreWeave Closes Landmark $8.5 Billion Financing Facility (2026-03-31)
  3. CCIR Research — CoreWeave's Debt, Entity by Entity (as of the June 30, 2026 10-Q)
  4. American Compute — GPU loan spreads and interest rates (2026)
  5. American Compute — What Lenders Recover When GPU and Compute Loans Default
  6. Global Data Center Hub — Lambda $1.008 Billion Delayed Draw GPU Term Loan (2026-10)
  7. Global Data Center Hub — Is CoreWeave's $8.5B Deal the GPU Asset Class Moment?
  8. Reuters via SRN News — Goldman, Apollo seek buyers for $40 billion SpaceX AI chip deal, Bloomberg News reports (2026-11-04)
  9. The Standard — ING puts $500 million of data-center loans up for sale, Bloomberg News reports (2026-11-06)
  10. DealStreetAsia — Ares wary of AI chip financing over GPU depreciation risks (2026-10-29)
  11. The Token Dispatch — Why Chip Lenders Now Sell Their Loans
  12. Tech Times — CoreWeave Q3 Earnings: 159% Revenue Growth vs $900M Quarterly Interest (2026-11-09)

Verification

Published
Last modified
Cross-check
Checked against 12 independent sources.
Unverified
  • The Semafor report gives no discount size, sale price or seller names. The claim that GPU financings are being hit hardest comes from one unnamed person, and we found no second source with specific discount figures.
  • CoreWeave was due to report third-quarter 2026 results after the US market close on November 10, 2026; we could not confirm the actual figures. The $860–940 million interest figure is company guidance from August.
  • Advance rates (90% for investment-grade contracts, 60% for others) and coverage tests (1.15x–1.40x) rely on one secondary analysis of CoreWeave filings, CCIR Research, not on the credit agreements themselves.
  • CoreWeave's total debt differs by definition: CCIR counts $35.55 billion of principal at June 30, 2026; earnings previews cite about $51.6 billion including leases.
  • Recovery rates on 2022 mining-equipment loans are American Compute's own compilation from company filings; the firm is a GPU financing business and may have an interest.
  • The SpaceX financing and the ING loan sale are Bloomberg reports relayed by Reuters and not confirmed by the parties.
Authoring
Reviewed by a person before publication. The full process is described in the Editorial.

Ten stories, once each morning

We send the three-line summaries only; the full pieces stay on the site. One-click unsubscribe, any time.

Related