Financialization of AI Computing Power: Open Source Models are Pushing Computing Power to Capital Markets (Part 2)

This article is a deep report produced by OKX Ventures. Due to its length, it is published in two parts: the first part focuses on the assetization of computing power, risk exposure, and derivative pricing logic; the second part will analyze the financial infrastructure of computing power, opportunities for assetization in inference, and industry trends. This is the second part.

IV. Financial Infrastructure of the Computing Power Market: Price, Risk, and Delivery

The financial market for computing power is beginning to piece together several layers of infrastructure. Indices compress the scattered GPU rental prices into a unified benchmark; exchanges organize standardized risks around the benchmark; dealers take on the basis brought by specific data centers, terms, and SLAs; capacity platforms then reconnect financial positions to physical computing power. The one who accumulates enough real orders and contract references first will be closer to the pricing power of the early market.

4.1 Computing Power Index: Who Can Become the Default Reference Price

The transaction prices of the same model GPU are still influenced by factors such as cluster size, interconnection method, region, rental period, prepayment ratio, SLA, and supplier credit. For index providers, the challenge lies in whether these differences can be processed to a point where the market is willing to use them for contract settlement, loan valuation, and derivative delivery.

Currently, three main routes can be observed. Silicon Data and Ornn are closer to independent benchmark providers, mainly collecting market quotes and contract data, then standardizing them to form benchmarks; Compute Desk acts as both a computing power intermediary and settlement, thus can directly see completed spot and forward contracts, including prices, terms, and actual settlement conditions; SemiAnalysis, on the other hand, models TCO from the perspectives of electricity, chip procurement, depreciation, and data center costs, making it more suitable for asset valuation and research, but further from financial settlement benchmarks.

4.2 On-Exchange Trading/Standardized Futures: Traditional and Emerging Compliance Exchanges Entering the Computing Power Market

The core value provided by exchanges is standardization, clearing, and distribution. For Neocloud, AI Lab, and private credit institutions, if hedging can directly connect to existing FCMs, margin, and clearing systems, the cost for institutions will significantly decrease. Traditional exchanges and emerging compliance platforms have thus begun testing computing power products: CME and ICE continue to leverage mature futures infrastructure, while Architect AX/AIX and Pluto/PMEX explore different paths in perpetual contracts and fixed-term futures.

The cold start of computing power futures is much more complex than listing a contract. Computing power contracts can easily become fragmented. Different GPU models, regions, and cluster configurations correspond to different delivery values, making it difficult for liquidity to naturally concentrate on a single contract. The far-end prices are also highly dependent on chip delivery, electricity, and model efficiency, and ordinary financial market makers lack physical order information, making it hard to continuously provide deep quotes over six months or a year.

The industry's hedging positions may also not be large enough in the short term. Hyperscalers can absorb a large amount of risk internally, and leading Neocloud has already covered a significant portion of cash flow with long-term contracts. The early on-exchange market is likely to first accumulate trading and speculative liquidity; whether industry hedging can keep up will depend on whether small and medium operators and AI teams purchasing computing power on demand begin to trade continuously and repeatedly.

4.3 Over-the-Counter Trading, Market Makers: OTC Customization and Basis Management

Before deep formation on-exchange, a large amount of risk is likely to be transferred through OTC first. The specifics of physical computing power, such as region, network, cluster size, SLA, and delivery time, make it difficult to fit into a unified exchange contract.

Dealers thus occupy a very valuable position. On one side, Neocloud hopes to lock in rental prices in advance, while on the other side, AI Lab wants to lock in procurement costs. Dealers can provide customized quotes for both parties and then use standard futures or other positions to manage overall price risk. When buying and selling demands do not align in terms of time and specifications, dealers must first take orders using their own balance sheets.

This type of business does not just earn from the bid-ask spread. Data center, electricity, network conditions, delivery constraints, and customer credit all create basis, and the ability to continuously price these non-standard differences and be willing to bear risks with capital is a capability that is harder for others to replicate. FalconX and Robert Leshner's H100 swap, as well as Wintermute's H100 forwards, can be seen as early attempts by market makers to test computing power forward pricing and risk transfer. Financial intermediaries are beginning to provide customized term quotes for physical computing power while leaving the non-standardized parts in their own books.

Thus, real order flow is very important. The index tells the market the average price, while the inquiries and transactions in the dealers' hands are closer to how much premium needs to be paid for a specific region, configuration, and term; this information directly determines how the basis is priced and where profits are retained.

4.4 Capacity Platforms and Physical Delivery: From Paper Positions to Physical Fulfillment

For AI Lab, locking in financial positions does not guarantee that there will be a set of GPUs available that meet the requirements at the time. This distinction is especially important when computing power is tight. Without the ability for physical fulfillment, futures prices and actual data center rental prices can easily diverge.

Currently, several different levels of attempts have emerged. Spot platforms like Vast.ai, RunPod, and Hyperbolic provide richer physical quotes and available capacity data; SF Compute allows reserved capacity to be resold or repurchased, giving long-term capacity contracts some liquidity; the EFP network attempts to connect standard futures with specific physical contracts, using futures to handle overall prices and offline contracts to manage region, network, and SLA.

Models like SF Compute are particularly noteworthy. In traditional Take-or-Pay contracts, unused capacity is often close to sunk costs; allowing resale gives future capacity a tradable rights attribute. This will increase the liquidity of long-term contracts and provide clearer forward prices in the physical market itself.

EFP addresses another layer of issues. AI Lab can first lock in market beta using standard futures, and when it comes to actual deployment, it can find specific data centers through the fulfillment network and pay the basis corresponding to region, network, and SLA. This way, the financial market does not need to cram all physical differences into a single futures contract while still serving real industry deliveries.

4.5 Inference Capital Markets: Financing, Service Rights, and Revenue Rights in the Inference Economy

Once GPUs are transformed into tokens by inference service providers, financial demand begins to extend from hardware rental to working capital and contract rights for inference businesses. If inference service providers wish to lock in complete gross profit, they theoretically need to manage both GPU input costs and token output prices simultaneously.

The profit for inference service providers is roughly equal to: Token realization price × Token output per GPU-hour × GPU utilization rate − GPU costs − other operating costs.

4.5.1 What are the Demand Points for Inference Economic Assets?

GPUs and capacity need to be procured in advance, while API revenue occurs gradually with customer requests. Overestimating demand can leave idle machines and occupy cash. Large model companies and hyperscalers can absorb this volatility with their balance sheets, but small and medium-sized inference platforms are likely to rely on customer prepayments or external financing.

The time lag between capacity and demand realization has given rise to four types of assets: GPU credit for financing upfront investments; prepaid service rights to recover future income in advance; capacity incentives to subsidize supply before demand forms; and model deployment assets attempting to allocate remaining profits after operating costs.

The traditional cloud market has long used prepaid credits and reserved capacity to address similar supply-demand mismatches. For example, OpenAI's Scale Tier allows enterprises to purchase input and output token throughput units for specific models in advance. This has become a standardized prepaid inference service, but the quota is tied to the enterprise account and cannot be freely transferred.

Crypto adds a new layer of property design on-chain, allowing service quotas originally recorded in supplier accounts to be held by wallets and further transferred and financed. In the future, if agents begin to procure inference services themselves, these rights can also directly enter the budget system of machines. As programmatic payments become a standardized infrastructure capability in the industry, the on-chain position is actually more recognizable at the rights level. Whether a future service can be independently held, transferred, or financed will directly affect its ability to evolve from ordinary credits into financial assets.

Currently, the share of on-chain inference at the execution level is still very small, with crypto-native inference service providers accounting for only about 0.5% to 1% of daily token traffic on OpenRouter over the past three months. This also means that the more realistic opportunities on-chain at this stage may appear in the financial layer, with inference services continuing to be produced by the most efficient platforms while on-chain handles the flow of funds and rights.

4.5.2 Four Types of Assetization Attempts in Inference

Currently, there are four directions for assetization of rights in Inference Capital Markets.

GPU Credit

GPU credit has gained traction primarily due to the presence of external payers and its financial product structure being closely aligned with traditional credit assets, making it easier for institutional capital to comprehend.

A typical GPU loan still follows the traditional asset financing structure: the GPU is held by a Delaware SPV, and the lender establishes a security interest through a Loan and Security Agreement, UCC-1, and data center lien waiver. On-chain records document loan shares, repayments, and profit distributions; in the event of default, debt enforcement still reverts to offline contracts, equipment control, and the court system.

Once the borrower deploys the GPU, downstream customers continuously pay for computing power, and this cash flow supports the loan principal and interest. As long as the purchase agreement continues to be honored, there is a clear source of repayment for the loan.

However, the direction of GPU credit is also the most easily absorbed by traditional capital. As future senior loans come with standard terms, interest rates, LTV, and legal recourse, the capital costs for banks, insurance funds, and private credit will be significantly lower than on-chain funds. Large, investment-grade loans with clear structures are likely to end up in traditional syndicates, ABS, or private credit funds.

The positions that can be retained on-chain for the long term are mainly in areas where the service costs in TradFi are relatively high, such as: loan origination for small and medium-sized operators, cross-border stablecoin funding, loan share distribution, real-time equipment and utilization monitoring, and junior or mezzanine layers that bear initial losses. A more likely future structure is that on-chain protocols will be responsible for discovering borrowers, monitoring assets, and organizing the first batch of capital; after the loan performance stabilizes, the senior portion will be sold to traditional funds, while on-chain continues to hold service fees, subordinate risks, and data relationships.

Prepaid Service Rights

Prepaid service rights address the issuer's current funding needs. The platform sells future API capacity in advance, receiving cash upfront and then continuing to fulfill obligations in the future. As the term extends, the platform risk borne by the buyer becomes increasingly apparent. Will the model still be competitive in a few years? What level will service prices drop to? Can the purchasing power of credits be maintained? All these factors will affect today's valuation.

Thus, these products are more likely to gradually approach standardized capacity contracts. The terms will be shorter, service standards clearer, and models can be replaced according to agreed rules. The clearer the contract, the easier it is for prepaid service rights to form a relatively independent market price.

Additionally, there is a reverse selection issue. Platforms with the best models, strongest customer demand, and most stable pricing power typically lack the incentive to issue long-term, transferable service rights. They prefer to retain customer differential pricing, revenue from unused credits, and the ability to adjust prices at any time. Those most in need of pre-selling long-term service rights are often platforms that require financing and customer acquisition but whose future competitiveness has yet to be proven.

Capacity Incentives

Capacity incentives address the supply organization issues in the early stages of the inference network. Nodes need to configure GPUs and models in advance and also bear operational costs. When demand has not yet stabilized, nodes lack sufficient motivation to go live early; network token emissions can provide the supply side with a revenue guarantee, allowing the network to first gain usable capacity.

Whether this can be established in the long term still depends on whether external customer payments can gradually cover node revenues. The protocol can verify that an inference task has indeed been executed and that the node has used the specified model, but the act of computation itself does not generate economic value. Token emissions can buy a period of supply, but customer demand still requires real orders to prove it.

A more reasonable incentive mechanism may gradually bind to real income. Customers first pay task fees with stablecoins, and after nodes complete inferences and pass verification, they receive income; the protocol then provides additional rewards based on external revenue and gradually reduces subsidies as the network matures. This allows emissions to bear the cold start costs while giving the market a clearer signal: how much of the node's income comes from customers and how much still comes from capital subsidies. In the long run, the real goal of capacity incentives is to transition from purchasing supply with tokens to sustaining supply with customer income; this transition must occur before the network can establish an economic foundation independent of the token market cycle.

Single Model Revenue Rights

Model revenue rights are closest to the previously mentioned crack spread:

Model Deployment Remaining Profit = API Revenue − GPU Costs − Network, Storage, and Engineering Costs − Customer Acquisition and Subsidies

Many Model Tokens primarily offer buybacks and staking incentives to holders, which may not correspond to a legally claimable income right. If there is a lack of a clear contractual relationship between API revenue and Token value, it becomes challenging to form a stable valuation anchor.

The lifespan of models further amplifies this issue. Models update rapidly; an asset that has just completed financing and deployment may already be surpassed by new alternatives. Therefore, a single model is difficult to support as a long-term financial asset.

A more likely development direction is to gradually expand the underlying unit to a workload strategy. For example, a coding inference strategy can continuously switch underlying models, adjusting routing based on quality and cost. This way, holders are engaging in a set of continuously operating inference strategies, with asset lifespan following customer demand and workload.

Furthermore, if such products are to enter the institutional market in the future, they will require clear legal entities, audited API revenues, and a clear cash flow waterfall. Without these foundations, the valuation of Model Tokens will still heavily depend on future usage and the issuer's buyback policies.

4.5.3 Router Provides Demand-Side Risk Management Before Token Futures

The Router is an underestimated layer in the financialization of inference.

What enterprises ultimately care about is how much a task will cost. The specific model used is often just a means to an end in many scenarios. The Router can continuously compare different models and providers, reallocating traffic when prices change. Originally, the underlying services were highly heterogeneous, but after Router scheduling, substitutability will gradually improve.

This will directly affect the scale of demand-side derivatives. Many cost fluctuations can first be absorbed through model switching and software optimization. The remaining exposures that cannot be handled in these ways will require financial instruments to intervene.

The Router also has a significant data advantage: it can see real bills. Public API official prices often fail to represent the actual procurement costs for enterprises; the Router can see how much customers actually paid and what level of service quality corresponds to that price. After accumulating orders, this data has the potential to form a quality-adjusted workload cost index.

Going a step further, the Router can directly provide enterprises with cost ceilings. For instance, enterprises can lock in budgets for future coding workloads over the next three months, and the Router manages model selection and capacity procurement in the background, using GPU derivatives to handle remaining risks when necessary. Customers ultimately gain budget certainty without having to manage model baskets, futures positions, and basis. Such products may align better with enterprise procurement habits than generic Token futures. The Router can keep the underlying complexity internal while providing customers with a more stable and predictable price.

Tokens themselves have not yet formed a unified standard, but the Router can continuously seek alternative supplies around a specific workload. Once trading accumulates to a certain scale, the workload itself may become a new pricing unit.

If this path holds, the Router will gradually extend from a traffic allocation tool to procurement and risk management. Its most valuable assets in the long term will also become clearer: real orders, real transaction prices, and the ability to manage costs and quality across different models. The pricing standards for workloads are likely to first form at this layer before entering the index and derivatives markets.

4.5.4 Crypto's Window Before Standards Are Established

When a right has not yet formed a unified specification, traditional finance incurs high costs in handling it. The contract amounts are small, and the structures are fragmented, allowing for faster experimentation with new property forms on-chain. If agents begin to directly hold and execute service contracts in the future, the value of such infrastructure will further increase.

Once a certain type of asset forms mature contracts, large funds will gradually enter the banking and institutional markets. Senior GPU credit is already showing this direction. After standardized futures mature, exchanges and institutional market makers will also enter. On-chain protocols should accumulate real order flows (customer and transaction data) during this window period, and their contract standards and performance data will gradually form barriers.

The products that have emerged today roughly correspond to different stages of gradually standardizing risk units:

The earliest products still heavily rely on the credit of a single company. As contracts gradually allow for service substitutions, the issuer's risk will decrease. Only when the market can stably define a basket of similarly quality workloads will the demand-side index have a chance to form. Futures and options need to be built on such stable risk units.

Agents may further expand the use cases of on-chain contracts. If software begins to autonomously purchase inference services in the future, a single enterprise API contract will be broken down into numerous real-time procurement decisions executed by machines. At that point, service rights holdings, budget constraints, performance verification, and more will all fall within the design scope of financial infrastructure.

The financialization of inference ultimately aims to transform highly customized AI services into rights that capital markets can understand and price. The pricing power in the early market will depend on who first defines these rights.

5. Conclusion

====

1. Timing/Why Now?

In 2026, traditional financial institutions like CME and ICE began to intensively enter the computing power pricing and derivatives market, with market infrastructure already starting to be built. Compared to the products announced by exchanges, we are more concerned about the pace of changes in two directions.

First, debt is increasingly penetrating AI infrastructure. As GPUs and data centers rely more on external financing, fluctuations in rental prices will directly impact DSCR, LTV, and refinancing conditions. Lenders also have a stronger incentive to write price risks into covenants, requiring operators to lock in a portion of future cash flows in advance. CoreWeave's financing path has already demonstrated how contract coverage affects credit pricing; if similar financing structures continue to spread, private credit and banks are likely to become significant driving forces for computing power hedging.

Second, an increasing demand for computing power is appearing in the public procurement market. Open-source models have lowered the barriers for enterprises to deploy models independently, leading to an increase in workloads undertaken by third-party clouds, Neocloud, and inference platforms. The more marginal demand the public market takes on, the easier it is for GPU rental prices to form continuous price discovery, and more enterprises will be directly exposed to rental price fluctuations.

The market is still in its early stages. GPU iterations are rapid, and the regions, networks, clusters, and contract terms are highly non-standard, so early trading will likely still be concentrated in OTC, with dealers first meeting customer demand and basis. A noteworthy adoption signal is whether industrial clients have begun to hedge continuously and repeatedly.

2. Perpetual Contracts vs. Fixed-Term Futures: Which is Easier to Find PMF?

The risks of industrial clients usually have clear timeframes. Six-month training projects, one-year capacity contracts, and multi-year GPU loans all require a corresponding price for that timeframe. Fixed-term futures are more aligned with corporate risk management in this regard and are easier to connect with physical contracts of specific models, regions, and SLAs through EFP.

Therefore, perpetual contracts are more suitable for near-term price discovery and trading liquidity, while fixed-term futures are better for corporate hedging, term pricing, and physical performance.

3. Which Layer is Most Likely to Capture Commercial Value First?

In the short term, we believe that brokers/dealers who manage physical order flows, as well as midstream platforms that control inference order flows and scheduling, will capture commercial value more quickly. The current market often pays a high premium for "whether there is computing power capacity," and the value derived from securing cards currently far exceeds the hedging costs of optimizing a few basis points. Those who know where there are available cards, who are competing for capacity, and who are willing to take orders with their own balance sheets have the opportunity to earn matching fees, bid-ask spreads, and basis premiums simultaneously.

As market trading gradually standardizes, the value of price indices will increase, and once a benchmark is repeatedly referenced by a large number of contracts, the migration costs and network effects will rise rapidly.

Later, after standard contracts and industrial hedging scales are large enough, exchanges will fully benefit from the economies of scale brought by clearing, margin, and liquidity networks.