The GPU cloud you rent inference from is probably a debt instrument. Dedicated GPU clouds like CoreWeave and Nebius buy Nvidia hardware using loans that the hardware itself secures, and the hourly rate you pay has to service that debt. A multi-year commitment is a bet on a refinancing cycle whose cost basis can shift before the term ends.
What is the GPU debt loop?
The loop runs in three legs: capital enters a GPU cloud provider from equity investors and lenders, that provider borrows against the Nvidia GPUs it buys to fund yet more GPUs, and enterprises rent the resulting capacity as inference, with the hourly rate servicing the debt. Each leg reinforces the next. Nvidia’s position is what makes the loop tight: it is the dominant supplier of the silicon that collateralizes the debt, with $215.9 billion in FY26 revenue and more than 80% of the market for AI-training and inference GPUs.
The mechanism that makes the loop turn is collateralization. When the GPU itself secures the loan that bought it, the lender’s recovery depends on the GPU holding resale value. That tie converts a hardware purchase into a debt-backed bet on Nvidia’s product roadmap. Capital expenditure and the debt to carry it sit on the providers’ balance sheets, while Nvidia retains its chip margin.
How does CoreWeave collateralize its GPU fleet?
CoreWeave is one of the largest dedicated GPU clouds, recognized by Gartner as a Visionary in the 2026 Magic Quadrant for Cloud AI Infrastructure and marketing itself as an AI-native platform on Nvidia silicon, with engineering for the Vera Rubin NVL72 rack and claims of 10x faster inference spin-up. The business model underneath those claims is the relevant part: a fleet of Nvidia GPUs, financed with debt, with the GPUs as the collateral.
The collateral pool is not static. A provider that starts by pledging hardware can, over time, pledge the customer contracts sitting on top of that hardware instead. That shift matters because it changes whose asset the lender depends on. Hardware collateral ties recovery to resale value; contract collateral ties it to a tenant’s willingness to keep paying. A buyer signing a multi-year commit may discover that the commitment itself has become security for the provider’s next debt raise.
How is Nebius financed?
Nebius is the second dedicated AI cloud worth comparing. Based in Amsterdam and listed on Nasdaq, with Arkady Volozh as founder and CEO, it competes with CoreWeave on the same Nvidia hardware that both rent out. Nebius’s public material is thinner than CoreWeave’s: the about page lists leadership, the origin of the company name, and equity stakes in adjacent businesses, but offers no throughput figures or cost claims a buyer could benchmark.
The asymmetry between the two providers is visible even without balance-sheet access. CoreWeave publishes more operational metrics and courted a Gartner placement; Nebius publishes less. A quieter provider is not necessarily safer. It may simply be one whose collateral structure is harder for a tenant to inspect.
Why does your hourly GPU rate embed a credit-risk premium?
An inference hour has to cover three costs: the electricity and depreciation of the GPU, Nvidia’s margin on the silicon, and the financing charge on the debt that bought the GPU. On a debt-heavy balance sheet, the financing charge is not a rounding error. The rate you sign today has to service debt that may be refinanced at a different spread tomorrow.
This is the gap in how most teams comparison-shop GPU clouds. A spreadsheet that ranks providers by H100 hourly price treats the rate as a commodity quote. It is also a credit instrument. Two providers renting the identical H100 can quote different hourly rates not because the silicon costs them different amounts, but because the financing charge on that silicon differs. One rate embeds the cost of servicing collateralized debt against a fleet whose resale value tracks Nvidia’s release cadence; another embeds the cost of funding expansion before revenue catches up. The cheaper hourly rate may belong to the balance sheet carrying more debt.
A steeper-than-market discount on a long commit can be a signal that the provider needs bookable revenue, not that capacity is cheap. Once customer contracts are pledgeable collateral, committed revenue has financing value to the provider independent of its margin value. The buyer offering that commitment is, in effect, supplying the asset that secures the next debt raise.
Vendor marketing does not surface this. CoreWeave emphasizes speed of inference spin-up and optimal TCO on its platform; Nebius’s about page offers no comparable operational metric. Both are vendor-reported figures without independent corroboration, and neither is accompanied by balance-sheet disclosure. The buyer is asked to compare throughput and ignore gearing.
What happens when GPU collateral gets revalued?
Collateralized lending works until the collateral’s market value moves. GPUs pledged in one year at one valuation face a different recovery picture once Nvidia’s next generation reaches volume, because the resale value of an older GPU is set by what buyers will pay for it, not by what the lender booked. CoreWeave’s own product page now leads with engineering for the Vera Rubin NVL72, which is the kind of cadence reset that revalues everything pledged before it.
A collateral revaluation cuts two ways for the enterprise tenant. If the lender marks the fleet down, the provider faces a margin call or a refinancing squeeze, and the cost of carrying that fleet rises; that cost is recovered through the rate the tenant already locked, or through the provider’s willingness to honor committed capacity. If the provider survives the squeeze by raising new debt on worse terms, the new cost basis is inherited by whoever signs the next contract. Either way, the tenant’s “fixed” rate is fixed only on its own invoice.
The refinancing cadence runs inside the term of a typical three-year enterprise commitment. If customer contracts themselves become collateral, then a provider under refinancing pressure has an incentive to protect the contracts it can pledge and shed the ones it cannot, which may not be the contract you held.
How should teams compare GPU clouds beyond price per hour?
A procurement review that stops at hourly price is incomplete. The checklist that actually reflects the risk:
- Debt-to-equity ratio and gearing: a provider’s annual report predicts mid-contract price behavior better than its pricing page. A debt-heavy balance sheet signals that the rate you sign has to carry financing costs that can move.
- Collateral structure: ask what secures the provider’s debt. Hardware-only collateral carries one risk profile; customer-contract collateral carries another, because it means your contract is an asset on someone else’s balance sheet.
- Customer concentration: a provider whose revenue depends on one or two large tenants is one renegotiation away from a capacity reallocation. Concentration may be disclosed in filings; ask for it.
- Equity-backer alignment: if the GPU supplier itself holds equity in the provider, the provider’s incentive aligns with the supplier’s shipment cadence, not with the tenant’s cost stability. New GPU generations help the provider’s collateral and hurt the tenant’s locked-in rate.
- Vendor-reported performance: treat CoreWeave’s platform speed and TCO claims, and any comparable Nebius metrics, as marketing until independently benchmarked.
None of these items appear on a provider’s pricing page, and most are in public filings.
What contract protections should GPU buyers demand?
The rate embeds credit risk the buyer did not underwrite, so the contract is where that risk gets allocated. Specific protections worth negotiating:
- Price-adjustment caps: a clause limiting how the provider can pass through refinancing-driven cost increases, or defining the conditions under which a rate can be repriced mid-term.
- Capacity guarantee independent of collateral events: language ensuring committed capacity survives a provider refinancing or collateral revaluation, rather than being the first thing shed under margin pressure.
- Exit and migration rights on credit events: the right to terminate or migrate without penalty if the provider enters restructuring, defaults on its debt, or pledges the buyer’s contract as collateral.
- Disclosure triggers: a requirement that the provider notify the buyer of material changes to its debt structure or collateral composition, so the buyer is not learning about a contract pledge from the press.
These clauses are not standard in commodity cloud contracts because commodity cloud providers are not carrying the same gearing against the asset you are renting. Here, they are.
Frequently Asked Questions
Does the GPU debt loop apply to cloud providers that own their own silicon, like AWS Trainium or Google TPUs?
The loop is specific to providers that finance Nvidia GPUs. Because Nvidia holds more than 80% of the AI GPU market and its chips are the actual collateral, a provider training on Trainium or TPUs is outside this chain; its debt is secured by broader cash flows or data-center assets rather than a single supplier’s resale curve.
How does Nebius’s financial profile differ from CoreWeave’s today?
Nebius reported $529.8 million in 2025 revenue with a $446.7 million net loss against $12.45 billion in total assets. CoreWeave, by contrast, posted $2.08 billion in Q1 2026 revenue and a backlog near $100 billion, making Nebius a smaller, still-loss-making expansion story with far less disclosed operating detail.
What should a finance team model beyond the quoted H100 hourly rate?
They should add a credit-risk line item to the rate model. CoreWeave’s balance sheet shows a debt-to-equity ratio near 13.8:1, so financing charges are a large component of the hourly price, not a marginal overhead. Stress-test against an H100 collateral markdown and a refinancing at a wider spread, then negotiate price-adjustment caps and exit rights tied to those events.
What happens if a provider pledges customer contracts as collateral and then faces a margin call?
A margin call can force the provider to defend the contracts it can pledge and drop or reprice the ones it cannot. CoreWeave’s February 2026 push for $8.5 billion in financing reportedly used its Meta contracts as collateral, which means a tenant’s commitment can become security for the provider’s next debt round before the tenant is asked for consent.
When would a multi-year GPU commit stop making sense under this structure?
It becomes hard to justify when the collateral behind your provider is likely to be revalued during the term. Nvidia’s January 2026 $2 billion CoreWeave investment and its March 11, 2026 $2 billion Nebius investment both signal faster generation turnover, which raises the chance that H100 fleets used as collateral lose resale value before your contract ends.
