AI computing capacity is moving closer to the financing model used for physical infrastructure, with GPUs, data centers and long-term customer contracts increasingly serving as collateral for debt. That transition gives cloud operators more ways to fund rapid expansion, but it also ties their balance sheets to a market that lacks broadly accepted forward prices, liquid hedging instruments and transparent benchmarks for GPU rental rates.
The risk is becoming more visible as capital expenditure rises. AI-related spending by the five largest cloud providers climbed from roughly 20% to 30% of operating cash flow between 2020 and 2023 to nearly 94% in 2025, according to figures cited in the source material. Confirmed 2026 capital expenditure for those companies exceeds $700 billion.
Much of that spending covers assets with an unusual combination of characteristics: GPUs can produce recurring income through cloud rental, yet they can lose value rapidly when a newer chip generation improves performance or power efficiency. A lender may therefore be funding equipment whose expected earnings, resale value and useful economic life can all change before the loan matures.
Contracts support borrowing, but do not remove hardware risk
Long-term take-or-pay agreements have become central to this financing structure. Under a take-or-pay arrangement, a customer reserves computing capacity and must pay for it whether or not it fully uses the allocation. The contract turns uncertain future utilization into a more predictable revenue stream, allowing lenders to focus first on whether committed payments cover debt service.
CoreWeave has been cited as one example, with more than 98% of its revenue tied to take-or-pay contracts. Nebius and IREN have also used longer-duration customer agreements as part of financing arrangements for AI infrastructure.
For creditors, these contracts can offer more immediate protection than trying to estimate the liquidation value of specialized equipment. A GPU cluster supporting a contracted customer may generate reliable cash flow even if spot rental rates decline. In a default, though, lenders would have to consider the value of the hardware itself, and that value could be far less stable.
The difference between contracted and market pricing becomes particularly consequential at renewal, refinancing or default. If newer chips lower the market price for comparable compute, an operator may face weaker renewal rates while still carrying debt sized around earlier assumptions. If equipment values fall faster than principal is repaid, loan-to-value ratios rise and the lender’s collateral cushion narrows.
Economic depreciation can outrun accounting schedules
GPU hardware is exposed to two forms of depreciation. Accounting depreciation spreads the cost of an asset over a set period. Economic depreciation occurs in real time as the market reassesses what that equipment can earn relative to newer alternatives.
A new chip release can affect older hardware through several channels at once: hourly rental rates can fall, utilization can weaken, customers can demand lower renewal prices, and secondary-market buyers can reduce their bids. The effect may be gradual in a supply-constrained market, but it can accelerate when new capacity enters service or when software becomes efficient enough to reduce the computing needed for a task.
Long-term contracts can delay the impact of those changes rather than eliminate it. If customers continue paying agreed rates, an operator’s near-term cash flow may remain intact despite falling spot prices. The financial pressure reappears when the agreement expires, the company seeks new financing, or a default forces collateral into the market.
That structure resembles other asset-backed lending markets, but GPU capacity has fewer tools for managing the underlying price risk. There is no widely adopted futures curve for GPU-hours comparable to the benchmarks available for commodities, interest rates or foreign exchange. Nor is there a standard swap market allowing a cloud operator to lock in rental income or allowing a lender to hedge changes in collateral values.
Bilateral contracts are filling the gap left by derivatives
In the absence of standardized hedging markets, multi-year capacity agreements are functioning as both procurement contracts and informal price-management tools. Customers gain more certainty over future computing costs, while operators gain contracted revenue that can support borrowing.
Early one-year agreements for Nvidia H100 capacity reportedly traded at a discount to prevailing spot rates. That gap later narrowed as spot prices declined and forward contract prices rose. The pattern illustrates how bilateral contracts can capture expectations about future supply and demand, even without a public forward market.
The limitation is that such agreements produce fragmented pricing. Contract terms vary by chip type, location, power availability, networking configuration, service levels and delivery period. Two H100 rental agreements may have different economic value even when they quote the same nominal hourly price.
A public index could provide a common reference point across those variables, allowing credit agreements and commercial contracts to use more consistent benchmarks. It could also support dynamic loan-to-value calculations as rental rates and residual values change. Without that reference price, firms must rely on private quotes, negotiated contracts and internal assumptions that may become outdated quickly.
GPU-hours are difficult to price like conventional commodities
GPU-hours present a further challenge because unused capacity cannot be stored. A vacant GPU hour disappears once the hour passes, preventing the kind of inventory-based arbitrage that often links spot and futures markets for physical commodities.
That leaves forward prices more sensitive to near-term developments such as delayed hardware deliveries, power-connection setbacks, new data-center capacity, or software optimizations that reduce demand for compute. In such a market, actual customer order flow may offer more useful signals than posted spot quotes, especially when large customers are reserving capacity months or years in advance.
The financialization of AI capacity may also differ sharply from the pricing of AI tokens. GPU capacity can be specified by chip generation, geography, delivery date and performance configuration. Token-based services are harder to standardize because providers vary in model capability, latency, throughput, reliability and enterprise-specific pricing.
Efficiency gains add another complication. If a model can perform a task with materially fewer tokens or less computation, a long-dated token-price contract may no longer reflect the cost of delivering the service. Workload-based units, such as a defined coding, search or inference task, may prove easier to standardize than raw token volumes.
AI infrastructure is therefore becoming a credit market before it has developed the pricing tools usually associated with one. Long-term customer contracts can make debt financing viable, but they cannot establish a continuous market value for hardware whose earning power changes with every new chip cycle.
Explore how blockchain markets price risk with TradFi vs DeFi, informing strategies for GPU-backed AI infrastructure.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
