Author: OKX Ventures

Computing power is transforming from IT costs into capital assets, while the uncertainties of future rental and residual value are turning it into a new financial risk exposure.

Introduction: The Capitalization of Computing Power and Risk Exposure

AI infrastructure is entering a capital-intensive expansion phase, and the economic attributes of computing power are changing accordingly. In the past, companies viewed computing power more as on-demand IT costs; now, GPUs, data center capacity, and long-term procurement contracts are increasingly appearing on balance sheets, and computing power is beginning to exhibit characteristics of capital assets: high initial investment, a payback period of years, and future income and equipment value continuously influenced by supply-demand dynamics and technological iterations.

GPU procurement and data center construction are becoming increasingly reliant on debt financing. Today, GPU credit can already be financed through equipment collateral and long-term customer contracts, but the market still lacks publicly available, standardized GPU-hour forward prices and hedging instruments. Debt needs to be repaid according to a set rhythm, and how much rent this batch of GPUs can generate or what they will be worth in a few years depends on the market at that time. The higher the leverage, the greater the impact of this price uncertainty on debt repayment and refinancing.

1. Background: The Capital Structure of the Computing Power Market

1.1 Computing Power Expansion Begins to Enter the Leverage Phase

One of the most noticeable changes in AI infrastructure over the past few years is that capital expenditures have begun to outpace internal cash flow expansion. From 2020 to 2023, the AI-related CapEx of the Top 5 cloud providers accounted for an estimated 20%–30% of operating cash flow, and by 2025 it will approach 94%. By 2026, confirmed CapEx for the Top 5 cloud providers will exceed $700 billion.

These companies still possess strong cash generation capabilities, but as infrastructure construction expands at scales of hundreds of billions of dollars over the long term, debt capital will naturally play a larger role. The capital asset of computing power is starting to form a separate financing logic.

GPUs, servers, and data centers require upfront investment before revenue is realized, and operators like Neocloud generally do not have the mature corporate credit of hyperscalers. Therefore, lenders gradually shift their financing judgment to the projects themselves: Is future cash flow sufficiently stable? Can it cover principal and interest payments, and how much asset value can be recovered in case of default?

Take-or-Pay long-term contracts are the key to the establishment of this structure. Even if customers do not fully use the booked capacity, they must pay the agreed amount according to the contract. For operators, this locks in a portion of future income in advance; for lenders, it converts the highly uncertain GPU utilization rate into a more predictable contract cash flow.

CoreWeave is the most typical case of this model (with over 98% of its revenue coming from Take-or-Pay clause contracts). As the proportion of long-term contracts with investment-grade customers increases, its financing increasingly relies on the creditworthiness of purchasers and contract coverage. Operators like Nebius and IREN, which have long-term contracts with large customers, have also begun to adopt similar structures.

This changes the risk ranking of GPU credit. Lenders first determine whether the contract cash flow can cover the debt, and finally assess how much value can be recovered from the GPUs after a default. The number of GPUs determines the scale of collateral, while long-term orders further dictate the financing conditions that this batch of assets can obtain.

1.2 After Cash Flow is Locked, What Remains is Asset Price Risk

Take-or-Pay improves the visibility of cash flow for debt service, yet the economic value of the GPUs themselves will continue to change. A few years later, how much income this batch of equipment can still generate and how much it will be worth is still a risk that remains on the balance sheet.

GPUs possess both the attributes of production equipment and technology products. They can continuously generate rent but are subject to rapid technological iteration cycles. After the launch of a new generation of chips, the impact on the previous generation of equipment will gradually reflect in rent, utilization rates, renewal prices, and second-hand market values.

Therefore, the accounting depreciation and economic depreciation of GPUs are often difficult to reconcile. Accounting standards can amortize the cost of equipment over fixed years, while the market is continuously re-evaluating how much competitive computing power a GPU can produce. For lenders, the core variables ultimately come down to the cash flow within the remaining loan term and how much value can be recovered in case of default.(Concerns have emerged around the depreciation policies of data center assets held by large cloud providers,drawing criticism)

This also explains why current GPU credit can temporarily remain stable during rent declines. As long as purchasing customers continue to fulfill their obligations, fluctuations in spot prices may not immediately impact current debt repayment; risks will be concentrated more at contract renewal, refinancing, and default resolution. By the time these points are reached, the market rent and residual value of GPUs will redefine the safety cushion for loans.

If rent and equipment value decline faster than the principal repayment of loans, LTV will rise again, and the safety cushion of the loan will narrow. Long-term contracts can push risk back but cannot lock in the economic value of the equipment throughout its entire lifecycle.

The current gap in the computing power credit structure is thus clear: contracts lock in some customer payments, while the market price of GPUs remains volatile.

Currently, the market mainly relies on long-term capacity contracts to lock in prices in advance. Early one-year contracts for the H100 exhibited clear discounts compared to spot prices, but as spot prices fell and forward contract prices rose, the price difference has significantly narrowed. Supply improvements and changes in contract terms can affect this outcome, but it also reflects a more fundamental issue: in the absence of standardized forwards, futures, and swap tools, industry participants can only have long-term physical contracts perform dual functions of procurement and price management.

Long-term contracts can facilitate risk allocation between individual trading parties; when the market needs to continuously form publicly determined prices and transfer risks across institutions, further standardized financial instruments will be required. As more GPU assets are supported by debt capital, this demand will become increasingly apparent.

2. What Does the Computing Power Market Truly Need to Trade?

Once computing power enters the capital structure, price fluctuations begin to have clear bearers. Neocloud holds the risk of declining rent, AI companies hold the risk of rising procurement costs, and creditors hold the residual value and refinancing risk. The derivatives market often forms around risks that cannot be digested by these balance sheets.

2.1 Three Types of Participants and Risk Exposures

Fluctuations in computing power prices ultimately trickle down to three types of balance sheets. Neocloud locks in GPU, data center, and financing costs at the outset, but future rents will still need to be reassessed; when long-term coverage is insufficient, a decline in rent will quickly compress EBITDA and DSCR, thereby motivating the sale of forwards to lock in revenue early. AI labs and inference platforms are on the other side, where rising GPU prices directly erode gross margins, and capacity risks often coincide with price risks during markets of scarcity; thus, they need to lock in both costs and available capacity simultaneously. Lenders are more concerned with how much safety cushion remains between the loan balance and the economic value of the GPUs, lacking public indices and forward curves, making LTV, refinancing, and covenants difficult to manage dynamically. As hedging requirements gradually enter loan agreements, the demand for computing power derivatives will also evolve from proactive corporate risk management to further integration within the financing system.

2.2 Potential Market Size: Which Risk Exposures Will Enter the Derivatives Market

If we estimate the Total Addressable Market (TAM) for computing power derivatives directly by global GPU shipment volumes, data center CapEx, or AI infrastructure sizes, the results are likely to be overestimated by an order of magnitude.

What can truly enter the derivatives market is that portion of risk still exposed to market prices and not absorbed by other means.

A significant portion of the risks associated with GPUs built and used internally by hyperscalers remains on their own balance sheets; the price risks of capacity already contracted at fixed rates over multiple years have also been distributed in advance through bilateral contracts. They carry economic risks but may not necessarily need to transact in the market daily.

Thus, from global demand for computing power to the actual derivatives TAM, it must undergo several layers of filtration:

This is also why the future size of the computing power derivatives market is likely to have a greater relationship with the share of merchant compute rather than simply growing linearly with global GPU installations.

In the past, many advanced models operated within a few large laboratories and hyperscalers, with computing power procurement also highly internalized. As the performance of open-source models improves, more companies can deploy them independently or outsource workloads to third-party Neocloud and inference platforms. The computing power demand originally confined to a few large balance sheets will gradually transform into purchasing orders in the open market. This also explains why the development of open-source models is worthy of attention:

First, more computing power is beginning to form real transactional prices. Without sufficient external trades, it is difficult to form an index, and the establishment of forward curves is even more elusive.

Second, demand shocks can more swiftly transmit to the spot market. After the release of a popular model, if numerous companies simultaneously ramp up deployments, the new demand will directly impact third-party GPU capacities, leading to rapid changes in price and availability. The more marginal demand the open market handles, the more likely rental fluctuations will truly become a P&L risk that companies need to manage.

Thus, the real significance of open-source models for the financialization of computing power lies in their potential to increase the ratio of market-based procurement. Only through enough external transactions can computing power gradually establish a price worthy of indexation, term structuring, and financialization. The variables ultimately determining the scale of the computing power derivatives market include: 1) how much computing power enters the open market, 2) how much price remains variable, and 3) how much of that risk cannot continue to remain on the original balance sheets.

DeepSeek V4 was released, and H100 rent rose approximately 7.5% in two weeks; Kimi K3 and GLM 5.2 also saw similar strengthening trends before and after rollout, as reflected in H100 and H200 rental prices.

3. Products and Pricing: How to Price Future Computing Power?

The difficulty with computing power forwards lies in the fact that the market has not yet established a sufficiently credible term pricing curve.

GPU-hour cannot be stored. An idle hour of H100 today permanently disappears after the time window, and cannot be pre-purchased during times of low spot prices for delivery six months later. The spot-forward constraints established through inventory arbitrage in traditional commodities are inherently much weaker in the computing power market.

This means that what price GPUs should transact at in six months is more reliant on chip deliveries, data center and power supply, model efficiency, and actual order strength. Hardware delays can quickly tighten future capacity, while software optimizations may release significant amounts of effective computing power without increasing the number of GPUs. A computing power forward curve simultaneously encompasses the market's judgments on hardware supply, energy constraints, and algorithmic advancements.

3.1 Without Inventory Arbitrage, Forward Prices Rely More on Expectations and Order Flows

The crude oil market has a relatively stable storage arbitrage constraint on its term structure. When forward prices are significantly above spot prices plus financing and storage costs, traders can buy spots, hold inventory, and sell forwards, making it hard for price spreads to deviate significantly from holding costs in the long run.

Computing power lacks the cash and carry arbitrage constraint, thus the constraints of spot prices on the forward curves are weaker.

Future prices are therefore highly sensitive to marginal information. Delays in delivery from Blackwell and power supply delays from data centers reduce expected supply; model distillation, inference optimization, or new generations of chips raising computational efficiency will increase effective computing power across the network. Many changes could rewrite supply and demand for forwards within months.

This elevates the value of real order flows. Public spot quotes can only reflect what has happened today, while long-term capacity orders can expose future supply and demand earlier. A broker or dealer that continuously matches Neocloud and AI Labs can see which data centers are beginning to experience spare capacity, and which buyers are willing to pay a premium for GPUs six months down the line. This information will eventually feed directly into forward pricing.

Thus, the value of a computing power index lies in compressing highly decentralized non-standard quotes into a set of price benchmarks that can be referenced by contracts, credit, and derivatives alike. If a specific index can continuously feed into long-term capacity contracts, loan valuations, OTC settlements, and futures deliveries, it will gradually become a pricing anchor commonly used across the entire market. Whether index providers and dealers can establish barriers largely depends on how close they are to real transactions and future orders.

3.2 The Financialization Rhythm of GPU-hours and AI Tokens Will Not Synchronize

In the AI value chain, there are actually two layers of prices present. Upstream sells GPU capacity, the midstream inference platforms convert GPU-hours into model calls, and downstream sells AI Tokens, APIs, or specific workloads.

For inference platforms, operational profits depend on both ends simultaneously. Rising GPU rents will increase input costs, while falling AI Token prices will suppress income. If prices on both ends form tradeable prices, inference platforms can theoretically lock in a compute to intelligence spread, similar to how refiners manage crack spreads in the energy market.

However, GPU-hours and AI Tokens currently differ greatly in their degrees of standardization.

While GPUs also have differences in models, regions, interconnection methods, cluster sizes, and SLAs, they can at least gradually define standard specifications around certain generations of chips, delivery timelines, and regions. An H100 or H200 still maintains relatively clear boundaries in its technological properties.

In contrast, Tokens lack a stable economic unit. Simply settling by token quantity makes it difficult to incorporate different models and qualities of service into the same financial contract. The same million tokens can come from models with completely different capabilities and correspond to entirely different latencies, throughput, and stability. Major clients' actual procurement prices often fall below public API quotes, making the listed prices themselves challenging to serve as a financial settlement benchmark.

Additionally, technological deflation further amplifies the difficulty of pricing AI Token terms. The ongoing improvements in model architectures, inference frameworks, and hardware efficiencies can substantially decrease the computational costs necessary for tasks of equivalent quality within a short time. Once terms are extended, the market must predict not only Token supply and demand but also how many Tokens and GPUs will be required for a unit of "intelligence capability" in the future.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。