The risks in GPU-backed lending, ranked
The main risks in a GPU-backed loan, ranked, with what lenders do about each.
The risks, ranked
Updated October 17, 2026. Ranked by how often each appears in the public cases, then by what lenders lost.
| Risk | How it shows up | Severity | Mitigants, and who offers them |
|---|---|---|---|
| 1. The GPUs resell for less than the loanIn 13 of 24 cases | A new generation ships, rental rates fall and the appraised value of the financed model crosses the LTV covenant. Revenue and collateral usually fall together.Miner prices down 85% in 2022; RTX 3080 resale down 64% before the Ethereum Merge; H100 rentals from over $8 to $1-2 an hour by August 2024; IBM 4300, 1979. | High. Mining lenders that took the machines back, or a quick cash payoff, recovered 37% to 65% of the claim. |
|
| 2. The offtaker stops paying or walksIn 5 of 24 cases | Payments arrive later than terms, the customer asks to defer delivery or cut reserved capacity, or it runs short of cash. In most GPU loans the offtake contract is the repayment source.Stability AI, 2023-24; Celsius at Core Scientific, 2022; Microsoft lease walkaways reported by TD Cowen (Microsoft said it was pacing its build), 2025. | High. Stability AI’s cloud suppliers forgave about $100M owed plus $300M of future obligations in 2024. Nobody repossessed anything. |
|
| 3. The operator or the site failsIn 4 of 24 cases | Unpaid power or colocation bills, a host in trouble with its own debt, rising SLA credits, a repair backlog, or the head of operations leaving.Compute North, 2022; Cyxtera, 2023; Exodus, 2001. | High. When Compute North filed, its secured lender ended up owning the sites; a customer with $81M of deposits and notes wrote off $55.7M. |
|
| 4. The collateral is stayed, frozen or double-pledgedIn 3 of 24 cases | A bankruptcy petition freezes repossession; serial numbers stop matching the UCC-1 schedule; another creditor files a UCC-1; a tax or export investigation opens.Core Scientific automatic stay, 2022-24; Northern Data raid, 2025; First Brands double-pledging charges, 2026. | High. Core Scientific’s equipment lenders waited 13 months from the Chapter 11 filing to emergence. GPU lenders have no Section 1110 shortcut. |
|
| 5. Revenue starts lateIn 1 of 24 cases | Energization or delivery slips 60 days or more, or the offtaker sends a cure notice. Interest runs before the cluster bills.CoreWeave guidance cut, November 2025. | Medium. CoreWeave cut its 2025 revenue guidance when one developer fell behind; the customer kept the full contract. A one-site borrower may not get the same patience. |
|
| 6. The equipment is overpriced or the site cannot run itIn 0 of 24 cases | The purchase price sits above resale data on day one, or the site’s power, cooling or rack density falls short of the hardware.No public default yet; it shows up as a thinner cushion in the cases above. | Medium. Starts the loan with less equity than the model assumes, so every later risk bites sooner. |
|
| 7. The equipment is damaged or breaks downIn 0 of 24 cases | Fire, water or coolant leaks, power surges, theft, damage in transit, or a mechanical or electrical breakdown. The servers stop billing while the loan payments continue.None among the public credit cases; handled by insurance claims. | Low. Insurable: property pays at replacement cost. Named windstorm and earthquake are often separate or sublimited, with deductibles of 2-5% and 10-15% of insured value. |
|
Case counts are out of the 24 public stress cases in our compute credit recovery dataset, 1979-2026; a risk counts a case where it was the main or a contributing cause, so one case can count twice. Credit-support announcements and index benchmarks are excluded.
The short answer
The biggest risk is that the GPUs resell for less than the loan, and it rarely comes alone: in most real defaults the offtaker stops paying at the same time. Of the 24 public stress cases in our compute credit recovery dataset (bitcoin miners, GPU clouds and older equipment leasing), 13 involved a collapse in collateral value and 5 an offtaker that stopped paying. Next come operator or site failure, collateral frozen by a bankruptcy or a title problem, and revenue that starts late.
What lenders get back depends on whether the collateral keeps earning. In the 2022-2024 bitcoin-mining defaults, the closest public test of equipment loans on serial-numbered computers, lenders that took the machines back, or a quick cash payoff, recovered 37% to 65% of the claim. Lenders that stayed with a borrower that kept its power, site or contracts recovered 80% to 100%.
We found no public record of a GPU-backed loan being foreclosed and its GPUs auctioned as of October 2026.
The risk of lending to neoclouds
Concentration and correlation. Three buyers, Anthropic, Microsoft and Meta, hold 78% of the disclosed GPU contract value in our Offtake Tracker, and 43% rests on two AI labs without public credit ratings (Anthropic and OpenAI). Most neoclouds’ contracted revenue sits with one or two of these buyers, and the GPUs securing the loan lose value fastest in the same market that would push a weak buyer to walk. Where the operator kept its site and contracts, mining-era lenders recovered 80% to 100%; where they took the machines back to sell, 37% to 65%.
The other neocloud-specific risks are operational: the cluster earns only with power, a colocation contract and a team to run it, and in Chapter 7 a trustee usually does not operate it. That is why the lenders who recovered best kept the collateral earning: a site sale (Argo), machines left running under a hosting deal (Greenidge), or a stake in a company that kept its power (Core Scientific). See what happens to GPUs when a neocloud goes bankrupt.
Which neoclouds are financially stable: what the ratings say
Most neoclouds are private and unrated; CoreWeave, which has public company ratings, is rated S&P B+ / Moody's Ba3. Ring-fenced GPU or data center financings for IREN, CoreWeave, Lambda, Nscale and Hut 8 (neoclouds and data center landlords) carry at least one investment-grade rating, because of the customer and the structure more than the company: CoreWeave’s best ring-fenced financing is rated Moody's A3 / DBRS A (low). A rated deal says little about whether the operator itself is financially stable.
| Operator | Company rating | Rated financings | Investment grade | Best-rated financing | Backed by |
|---|---|---|---|---|---|
| IREN | None tracked | 1 | 1 | IREN Microsoft GPU financing: Fitch A / DBRS A (low) | GPUs + customer contract |
| CoreWeave | S&P B+ / Moody's Ba3 | 5 | 1 | CoreWeave DDTL 4.0: Moody's A3 / DBRS A (low) | GPUs + customer contract |
| Lambda | None tracked | 2 | 2 | Lambda fixed-rate DDTL: DBRS A (low) / Moody's Baa1 | GPUs + customer contract |
| Nscale | None tracked | 1 | 1 | Nscale Ward County DDTL: Moody's Baa1 | GPUs + customer contract |
| Hut 8 | None tracked | 2 | 2 | Hut 8 Beacon Point notes: Moody's Baa2 | AI data center lease |
| TeraWulf | None tracked | 1 | 0 | WULF Compute notes: Moody's Ba2 / S&P BB / Fitch BB | AI data center lease |
| Cipher Digital | None tracked | 3 | 0 | Cipher Black Pearl notes: Moody's Ba2 / Fitch BB- | AI data center lease |
From the Rated GPU Financings tracker, where each rating links to the agency release or filing. Investment grade uses the lower of split ratings. For an unrated operator, the Neocloud Credit Score grades the borrower and the deal on the same 0-100 scale as the Offtaker Credit Score.
Is the AI GPU bubble a credit risk?
For lenders whose repayment depends on a weak offtaker or on resale value, yes; less so where an investment-grade buyer is contracted for the term. Of 16 publicly rated GPU and AI data center financings we track, 8 ($50.7B of $67.9B) are investment grade. The risk is correlation: a glut would cut rental rates, resale prices and weak offtakers’ cash at once, as one IBM product launch did to computer lessors and their insurers in 1979.
The exposure is large: our GPU Financing Tracker counts 110 public GPU and AI data center financings since 2023, $197B of new money. The GPU Residual Value Report 2026 puts the conservative 2028 band for an H100 at 5-7% of its original price: a lender sizing a balloon on today’s resale prices is taking that risk.
Methodology and sources
Show the method and sources
The cases are the 24 distress events in our GPU and compute credit recovery dataset: the 2022-2024 bitcoin-mining defaults, GPU-era stress since 2022 and older asset classes, each from company filings, court records or press. Each risk maps to one failure mode; counts include contributing causes. Recovery ranges are the lender recoveries those sources quantify. Severity weighs those recoveries and the time to resolution. 19 of the 24 cases are listed, with their sources, in 16 rows on GPU and neocloud defaults and stress events; the other 5 are older precedents outside compute, in the recovery casebook. That page also lists 7 newer cases, checked in October 2026, that are not yet in these counts.
Further reading: What lenders recover when compute loans default; GPU loan default playbook; What happens to GPUs when a neocloud goes bankrupt; Risk management for GPU loans; Insuring a GPU cluster. Offtake concentration is computed from the Offtake Tracker.
Related
Discuss a transaction
Tell us about the deal and the risk that worries you most. We will come back with the structure and the cover that fit it.
Prefer email? hello@amcompute.com
GPU lending risk questions
- What are the biggest risks in GPU-backed lending?
- The GPUs reselling for less than the loan, usually while the offtaker stops paying: 13 of the 24 public stress cases in our recovery dataset involved a collapse in collateral value and 5 an offtaker that stopped paying. Next come operator or site failure, frozen collateral and late revenue.
- What is the risk of lending to neoclouds?
- Concentration and correlation: three buyers hold 78% of the disclosed GPU contract value in our Offtake Tracker, and the GPUs securing the loan lose value fastest in the same market that would push a weak buyer to walk.
- Which neoclouds are financially stable?
- Most neoclouds are private and unrated. Investment-grade ratings on ring-fenced financings for IREN, CoreWeave, Lambda, Nscale and Hut 8 follow the customer and the structure more than the company.
- Is the AI GPU bubble a credit risk?
- For lenders whose repayment depends on a weak offtaker or on resale value, yes: a glut would cut rental rates, resale prices and weak offtakers’ cash at once. Less so where an investment-grade buyer is contracted for the term.