Open Models Are Building the Merchant Market for AI Compute

AI compute is turning into a capital market. GPU-backed credit got there first. Open-weight inference is now creating the merchant flow that benchmarks, dealers, and hedgers need.

Lenders were financing entire GPU fleets before any U.S.-regulated dated compute future began trading. Debt amortizes on schedule while rent and collateral value keep moving, turning the missing forward market into a balance-sheet problem. Open-weight models are moving more inference demand into third-party capacity markets, creating the transaction history a benchmark needs. Credit and physical capacity already trade; the missing layer is a term market that can transfer risk and still connect to a working cluster.

The financing shift is easiest to see in the relationship between capital spending and cash generation. PIMCO estimates that Microsoft, Google, Meta, Amazon, and Oracle will spend about $690 billion in 2026 and $870 billion in 2027. That spending would absorb roughly 94% of operating cash flow in both years, up from about 40% in 2023.

These companies still generate enormous amounts of cash. The scale and duration of the buildout are forcing more of it into debt markets. GPU systems and data-center capacity have to be secured well before the associated revenue arrives.

The financing problem is sharper for neoclouds. Most operators lack the corporate credit to borrow unsecured at the scale of a modern AI project, so lenders underwrite the project itself, with the equipment as collateral and the customer contract as the repayment engine.

Take-or-pay commitments make that structure financeable. The customer pays for reserved capacity even when actual use falls below the commitment, converting uncertain utilization into a contractual receivable. Better off-taker credit and broader contract coverage translate into more leverage and cheaper debt.

Look no further than CoreWeave. Committed contracts accounted for 98% of first-quarter 2026 revenue, and its financing moved from expensive private GPU-backed debt in 2023 to investment-grade project financing and a publicly syndicated delayed-draw facility in 2026. Nebius and IREN are following the same basic playbook as large customer commitments become the center of the financing case.

Private credit can underwrite one GPU project against its equipment and customer contract. A futures market asks unrelated firms to accept the same reference grade and settlement rules. That difference explains why GPU-backed lending scaled before a public curve, and why loan covenants are likely to be the first place that curve becomes economically binding.

The GPU count sets the collateral pool, while off-taker quality and contract coverage determine how much leverage the project can actually support. During the contract term, customer cash flow carries the underwriting. Hardware value returns to center stage when the contract rolls or the loan has to be worked out.

Take-or-pay contracts make debt service easier to see while the fleet keeps repricing underneath them. GPUs are awkward collateral because they combine a long physical life with a short economic one. A server can run for years while the next architecture resets how much useful work a dollar of hardware can buy.

The reset first appears in rent and utilization, then reaches renewal pricing and resale value.

For a lender, the risk concentrates when the contract rolls or the credit has to be re-underwritten. A strong off-taker can keep a project current through falling spot rent. Once renewal or enforcement begins, market rent and secondary value return to the credit decision. If they fall faster than principal amortizes, LTV rises and the recovery cushion tightens.

Accounting depreciation settles none of this: a five- or six-year useful life spreads historical cost through the income statement, while the market reprices earning power every day. The gap between those two measures has already turned hyperscaler depreciation policy into a serious market debate.

Long-dated capacity agreements already function as private forwards. They set a price inside one commercial relationship and bundle delivery terms with customer credit. That works well inside a specific buyer-seller relationship. A broader market needs a reference price that can be used by lenders, dealers, and investors who were never part of the original procurement agreement.

The H100 term market shows the transition. One-year capacity traded at a large discount to spot early in the cycle, then converged as spot rent fell and contract pricing recovered. Supply conditions and contract terms both shaped that move. The larger point is that the physical agreement still had to carry procurement and price management at the same time because no standardized hedge existed for the price leg.

Long contracts made the credit market scalable while absorbing much of the near-term rental exposure. Futures volume will come from the risk that reopens when those contracts end, especially the new rental price and the lender's collateral mark.

A bilateral agreement can allocate risk inside one commercial relationship. A capital market starts to emerge when the same exposure needs a public mark and can move across institutions. The more GPU assets sit behind debt, the more often that need will recur.

GPU-backed credit creates recurring demand for marks and hedges. Open weights supply the other ingredient: external transactions. A closed model keeps most procurement inside a lab and its cloud partners. An open model can be served by several independent providers, which turn one model release into competing capacity orders, visible quotes, and realized service data.

Those two forces reinforce each other. The credit market gives price risk financial consequence. The merchant market gives benchmarks something to observe. Either force on its own is insufficient: leverage without external flow leaves lenders marking illiquid contracts, while external flow without balance-sheet pressure may never produce repeat hedging demand.

Once compute enters the capital structure, price moves land on identifiable balance sheets. Operators carry future rent, buyers carry procurement cost and capacity scarcity, and lenders care about the value left beneath the loan.

Installed capacity overstates the opportunity for a compute derivatives market. Much of hyperscaler capacity is consumed internally, so its allocation does not itself generate an arm’s-length market price. Fixed-price contracts allocate most near-term rental risk between the two parties. Buyers can also respond to price differences by shifting workloads across providers or hardware, reducing their exposure before a financial hedge is needed.

The relevant base is the value of compute traded between independent parties whose rental price can reset during the hedge period. This is the addressable hedge notional.

Fixed-price capacity may become hedgeable as renewal approaches, when its future rental rate is once again exposed to market pricing. Exposure that cannot be mapped to the reference contract with acceptable basis risk remains outside the addressable market. In a market still dominated by hyperscaler internal use and bilateral contracts, early adoption is likely to concentrate among firms with large and recurring repricing exposure. Adoption can broaden as more procurement moves into external markets and benchmark liquidity improves. Hedge requirements in credit agreements could add a further source of recurring demand.

Open models matter because they push more demand into this funnel. Closed frontier systems keep most procurement inside a small group of labs and cloud partners, where capacity is secured through private contracts and long-term infrastructure commitments. Open weights give enterprises the option to self-host or buy the same model from independent inference providers. Since January 2026, leading open-weight models have trailed the closed frontier by roughly four months on average. That gap is already small enough for a growing share of production workloads to move across providers.

OpenRouter makes the merchant shift visible. By July 2026, weekly routed token volume was roughly 125 times its January 2025 level, while open-weight models represented about 75% of routed volume, up from 27%. A larger share of inference demand is now clearing through third-party providers, where rent and availability become observable market variables.

The same trend cuts in two directions. More providers can compete for each workload, creating more transactions and more observable prices. That competition pushes token prices lower and forces inference providers to extract more output from every GPU-hour. Lower token prices also make a wider range of workloads economical. Total usage can grow faster than revenue per token falls, leaving aggregate GPU demand higher while provider margins compress.

Compute supply moves quickly as well. Better serving software can release effective capacity without another data-center build. New hardware can command a higher hourly rent while reducing the cost of a completed workload. A major model release can tighten the near end of the curve; an efficiency gain can loosen it. The forward curve has to price both.

Compute derivatives will scale with the share of procurement that leaves corporate boundaries and keeps a floating price.

Compute has yet to develop a credible forward curve. A GPU-hour is perishable, and the amount of useful work it can produce changes with hardware and software efficiency. Future prices therefore depend heavily on expectations and real order flow.

Oil futures remain linked to spot through storage arbitrage. When the forward price rises far enough above spot to cover financing and storage costs, traders can buy physical oil, hold the inventory, and sell it forward. That trade prevents the spread from remaining far above the cost of carry.

A GPU-hour expires when the clock moves on. Capacity available at a low price today cannot be purchased and delivered six months later, even though the hardware remains in the rack. Without a cash-and-carry trade, spot prices exert a much weaker influence on the forward curve.

Prices six or twelve months out therefore react more directly to new information about future capacity and demand. A delayed Blackwell delivery or data center energization can reduce expected supply. Model distillation and serving improvements can release effective capacity across the installed base without adding GPUs. New hardware can further increase the amount of useful work produced by each GPU-hour. The forward curve has to reflect how these changes will interact with actual demand for capacity.

This makes term order flow especially valuable. A public spot quote shows where capacity is offered today. A broker or dealer working six and twelve month RFQs can see where data center capacity is beginning to open up and where buyers are willing to pay a premium for future delivery. That information enters forward quotes before it becomes visible in public spot data.

A useful compute index has to turn dispersed, non-standard quotes into a reference price that commercial and financial contracts can use. Its role becomes more durable as capacity agreements reference it, lenders use it to value collateral, and derivatives settle against it. Index providers and dealers closest to executed trades and forward orders will have the best information for maintaining that benchmark.

Upstream, the unit is a GPU-hour. Downstream, the same capacity is sold as API access or completed work.

Those two prices meet in the operating spread on each GPU-hour:

An inference provider is essentially a manufacturer with software inside the factory. Serving efficiency sets output per GPU-hour, while routing and customer pricing determine how much of that output turns into margin. A GPU hedge stabilizes one input, while the rest of the spread still depends on execution.

GPU-hours already require a detailed grade. Token services add another layer of model and service variation.

The GPU contract can specify accelerator SKU, region, provider tier, tenor, interconnect, and settlement index. A token-priced inference unit is produced and consumed in the same request, leaving no inventory that the buyer can resell. The market has a posted ask from the service provider and no natural bid from a secondary owner. Cross-platform arbitrage has no inventory leg that can pull two token quotes together.

Service quality also varies. The same open model can run at different quantization levels, context settings, cache policies, and serving engines. Time to first token, tokens per second, uptime, tool-call performance, and data policy all affect value.

List prices add another problem. Frontier labs can lower public rates, change input-output pricing, give enterprise customers private discounts, or retire a model. An index built from public price cards may end up tracking an administered schedule with little connection to transacted enterprise cost.

Technology deflation makes long-dated token pricing especially fragile. Model architecture and serving engines can reduce the compute required for the same result before a contract reaches maturity. The buyer is budgeting for an outcome, and the model used to produce it can change several times during the term.

The demand-side standard will form around a defined workload. A coding contract can specify the outcome and service envelope while leaving model selection inside the routing layer. The financial unit becomes the cost of completing the job.

Enterprises budget for outcomes, which makes a workload cost cap easier to buy than a portfolio of model-specific futures and provider basis. The router can keep that complexity inside its own book.

GPU-hours map to the production input, while workloads map to economic demand. They will reach capital markets on different schedules.

Compute already has active markets in credit and physical capacity. The next step is to turn scattered transactions into a reference price and then carry that hedge back to a cluster the buyer can actually use.

Even the same GPU model trades at different prices once cluster configuration, location, contract term, and provider credit enter the quote. An institutional benchmark must normalize enough of that variation to support collateral marks and cash settlement while leaving the residual basis visible to commercial users.

Silicon Data, Ornn, Compute Desk, NativX, and SemiAnalysis are all trying to turn fragmented rental markets into contract-ready reference prices, using different combinations of observed quotes, executed trades, and contract data.

A viable compute benchmark will likely take the form of a family of reference grades across hardware generations and service classes. Benchmark power begins when a number enters contract language. Repeated use in credit documents and settlement rules raises switching costs and makes the reference harder to displace.

Exchanges provide standardization, central clearing, and distribution. For institutions already active in listed derivatives, compute futures can fit into existing FCM, collateral, and clearing workflows.

CME and ICE have announced U.S. compute futures with fixed maturities, both pending regulatory approval. Architect has entered the compute derivatives market through its offshore venue, AX, and is separately pursuing a U.S. exchange. Pluto is also pursuing a similar onshore route.

In practice, GPU generation, region, and cluster configuration all change the commercial exposure. Contracts designed to reduce that basis risk can split liquidity across too many instruments. The market must choose a reference grade before trading reveals where liquidity will concentrate.

Beyond contract design, prices further along the curve depend on future chip deliveries, powered capacity, and serving efficiency. Term RFQ desks see those effects in physical orders before they appear in public spot prices.

For now, much of the commercial exposure remains internalized or fixed through bilateral contracts, limiting the natural hedge flow available to listed markets. Early exchange volume may therefore come mainly from proprietary and relative value trading.

A standardized GPU future can cover the common price move. Real procurement still requires a bilateral capacity agreement for the remaining specifications.

The dealer sits between a seller seeking a floor under future rent and a buyer seeking a cap on future cost. Close matches can run back to back. When timing or specifications diverge, the dealer may hold term capacity, use a nearby SKU or region as a proxy, or net the exposure across its book. Residual basis and liquidity risk remain on the balance sheet. Listed futures would offer another way to lay off risk, though they are not yet the foundation of the book.

The economics sit in the basis book. The client quote reflects region and service terms, along with the capital required to carry delivery risk, counterparty exposure, and hedge mismatch.

FalconX’s H100 transaction with Robert Leshner and Wintermute’s early H100 forward provide early evidence that bespoke compute exposure can be documented and settled. Dealer returns will depend on how much residual basis remains on the balance sheet and how long it takes to offset.

A hedge settled in cash can cover the benchmark move while the buyer still has to source a cluster that meets the specification.

ComputeConnect proposes a direct link between this physical market and U.S. listed derivatives. Its announced EFP workflow is designed to pair a futures position with separately negotiated GPU capacity. Hardware, location, service terms, and basis remain bilateral, while the futures leg is booked back to the exchange.

An EFP network still needs providers willing to deliver and enough listed liquidity to close the futures leg. Acceptance testing and default remedies determine whether the last mile ends in a working cluster.

The hardware story continues downstream. An inference provider secures GPU capacity before customer demand is fully known, spends engineering time on the serving stack, and receives revenue one request at a time. The selling price can reset much faster than the capacity commitment.

That gap has produced several financial rights around inference. One set of products finances the equipment, while another pulls service revenue forward. Networks also use production incentives to assemble supply, and model-deployment assets try to allocate the residual margin after serving costs.

4.5.1 Why These Assets Are Appearing

Inference has a working-capital gap because GPU capacity and serving work must be funded before the first request is billed. Large labs can carry that gap on their own balance sheets; smaller providers lean more heavily on customer prepayments or outside capital.

The cloud market already uses early versions of these instruments. OpenAI's Scale Tier lets enterprises pre-purchase model-specific input and output throughput for a minimum 30-day term, with defined service commitments, while keeping the entitlement inside the customer's account.

Onchain systems can separate that right from the supplier's account. A service right can sit in a wallet and move between holders when the contract allows it. The same right can fund an agent's budget or serve as collateral. Programmatic payment will become common across centralized systems as well, so the durable distinction comes from transferability and the ability to finance the right itself.

4.5.2 Four Claims on the Inference Stack

• GPU Credit Starts With a Real Payer

GPU credit has scaled first because the borrower already exists and pays dollar interest. The equipment and customer contract support a familiar legal claim. The legal structure still resembles private credit: a project SPV owns the GPUs, the lender signs a loan and security agreement and files a UCC-1 lien, and the data center has to cooperate with enforcement. Onchain records can distribute loan interests and cash flow; default recovery still depends on equipment control and the court system.

Traditional capital will move into this category first. Senior loans with standard documentation and demonstrated recovery can migrate into bank syndicates or private-credit portfolios at a lower cost of capital. Onchain platforms have a stronger position in origination and asset monitoring, along with junior exposure and smaller cross-border borrowers where institutional servicing remains expensive.

• Transferable Claims on Future Service

Prepaid rights finance the issuer today in exchange for future API access. The buyer receives a transferable service claim, while the provider takes on a future delivery obligation.

Long duration makes the product difficult because model quality can change quickly and the purchasing power of a service credit depends on the issuer's future catalog and pricing policy. The strongest platforms also have little incentive to sell very long, transferable rights because doing so gives up future repricing and the value of unused credits.

A more durable contract would use a shorter term, a defined workload, and a provider pool. Service levels and substitution rules can make the right easier to value. Transferability then solves a real procurement problem by letting an enterprise or agent sell unused capacity.

• Revenue Quality in Decentralized Inference

Decentralized inference networks need supply before external demand is deep enough to support it.

Commercial maturity depends on customer fees replacing protocol subsidy as the source of node income. Proof of execution establishes that a job ran. Payment from an outside customer establishes that the work had economic demand.

A rising share is meaningful only alongside growth in absolute customer fees. The ratio can also increase mechanically when the token price falls and protocol rewards lose value.

Revenue-linked emissions can tie subsidy decay to customer revenue. Customer fees fund completed work, while protocol rewards cover the shortfall during the cold start. Rewards decline as outside revenue grows. The network becomes commercially durable when customer revenue can support provider supply through a weaker token market without requiring additional subsidy.

• Duration Risk in Model Revenue Claims

A single model asset packages the economics of one deployment into a standalone unit. Residual API revenue can fund token buybacks after compute and operating expenses are paid. The resulting exposure resembles the inference spread. Holder rights still depend on the contract.

Most model token designs rely on buybacks. Holder value therefore depends on platform policy and secondary market liquidity. A contractual revenue share creates a direct claim on cash flow.

A model’s commercial life may be shorter than the financing claim written against it. Traffic can migrate to a better or cheaper model before investors have recovered their capital. Durability improves when the financed unit is a workload that can route among approved models as quality and cost change. The customer relationship can then survive model turnover.

An institutional structure needs an operating entity, audited API revenue, and a contractual cash flow waterfall. Investor value would come from a live inference business and its ability to retain workloads through model turnover.

4.5.3 Routers Can Price Work Before Token Futures Can Price Tokens

Routers sit close to the demand side of inference because they observe requests, model selection, provider performance, and realized customer cost. This transaction record carries more information than public API price cards, which reveal neither negotiated rates nor the quality delivered at those rates.

Part of the cost volatility can be absorbed operationally. A router can shift eligible workloads across models and providers as relative prices change, preserve session affinity when cache matters, batch non-urgent work, and change the serving path. Provider switching only helps when realized session cost falls. The cheapest quoted token can still produce a more expensive conversation once cache loss, retries, and failures are counted. The router has to optimize realized workload cost, not the posted input price.

Finance only enters for the exposure routing cannot remove: fixed model requirements, latency commitments, regional constraints, customer price guarantees, and committed capacity.

As transaction volume grows, the router's realized-cost data can support workload indices. Coding, research, and latency-sensitive chat can be segmented by quality floor, latency band, session profile, and data policy. The benchmark can then track the actual cost of delivering each service class across approved models and providers, including cache, retries, and provider continuity.

A workload cost cap turns that operating capability into a risk management product. An enterprise fixes the budget for a defined volume of work over a specified period. The router manages model and provider selection, session continuity, provider contracts, and capacity procurement, using GPU hedges where operating flexibility no longer covers the exposure. The customer receives budget certainty without managing model-specific contracts, hedge positions, and basis risk. The router buys variable inference and sells a predictable cost of work.

A cost cap moves the router from traffic allocation into a dealer-like role. Its core asset becomes the record of prices actually paid, service quality delivered, and substitutions that worked across models and providers. Public API cards show the posted quote; the router sees what the workload actually cost. Workload pricing standards may form here before they reach index and derivatives markets.

4.5.4 Crypto’s Advantage Narrows as Contracts Standardize

Traditional finance becomes more competitive as a right gains common terms and reliable enforcement. Senior GPU credit already points in that direction. Standardized futures will also concentrate around regulated clearing once commercial depth develops.

Onchain systems retain an advantage where ticket sizes are small or the contract remains bespoke. Machine-held rights add another durable use case because an agent can own a service budget and execute the terms directly. Global stablecoin funding can also reach borrowers or providers that institutional credit desks overlook.

Each asset has its own migration trigger. Senior GPU risk moves once recovery is proven, while production incentives need customer fees that survive declining emissions. Prepaid rights become easier to finance when the contract supports substitution and a real remedy after service failure, and model economics require audited revenue with enforceable holder rights.

Institutional capital enters as the contract becomes familiar. Crypto has the most leverage while the right is still being defined; the moat comes from customer flow, performance data, and the first workable standard built around a live market.

The thesis weakens under three conditions. Merchant procurement may remain too small if hyperscalers keep most demand inside their own balance sheets. Commercial basis may stay too wide if a broad index fails to track the cluster a buyer actually needs. Lenders may also decide that take-or-pay coverage and conservative advance rates are sufficient, leaving hedging outside normal credit documents. Any one of these outcomes would keep compute derivatives as a niche overlay rather than a core market.

Compute price risk is beginning to matter to both financing and procurement.

Debt is moving deeper into AI infrastructure. A financed GPU fleet carries fixed repayment and refinancing obligations against rental revenue that can reprice much faster. As merchant exposure begins to affect DSCR, collateral value, and covenant headroom, lenders have a reason to distinguish hedged revenue from unprotected revenue. Better advance rates or loan pricing could eventually make the hedge part of the financing package.

Open models are expanding the contestable procurement market at the same time. Demand that once remained internal to hyperscalers is becoming direct enterprise procurement exposure as open-model performance improves and deployment costs fall. The same weights can run across multiple providers, allowing each external purchase to add another observation of price, availability, and delivered service quality. Those transactions provide the raw material for a benchmark.

Infrastructure debt creates demand for risk transfer, while portable workloads generate the external transactions needed to price it. Physical specificity will keep much of the early market bilateral and basis heavy, but compute prices are starting to matter simultaneously to procurement, underwriting, and collateral valuation.

Commercial compute exposure usually has a defined horizon: a training run, a capacity renewal, or a GPU loan. Dated futures can match those horizons and, through an EFP, connect the hedge to a physical contract specifying hardware, region, and SLA. For asymmetric exposures, options may fit industrial users better than a large linear position.

Cash timing also shapes product choice. When rent rises, a short futures position can generate a large variation-margin call before the operator receives the higher physical revenue, creating liquidity strain even when the hedge works economically. An option fixes the maximum premium at trade date, while a structured OTC contract can align settlement with the underlying cash flow cycle.

In practice, perpetuals are better suited to front end price discovery and trading liquidity, while dated futures fit commercial hedging, term pricing, and physical procurement.

Value will accrue at different layers as the market develops. In the early stages, fragmented physical supply gives brokers, dealers, and capacity platforms an advantage in placement and basis management. The same order flow reveals the premium attached to a particular region, configuration, credit profile, and delivery date.

Routers occupy a similar position in inference. Their transaction data can support procurement, budget guarantees, and eventually workload benchmarks. Benchmark providers gain leverage as their reference prices enter capacity contracts, credit documents, and settlement rules.

Listed venues provide the standardization, clearing, and distribution needed to scale the market. As liquidity develops across maturities, dealers gain a more efficient way to recycle risk, while portfolio margin makes compute exposure easier for institutions to hold.

At this stage, an attractive business model is a compute merchant bank that finances capacity, sees physical order flow, quotes term risk, and manages the basis outside standard contracts. Exchange liquidity strengthens that model by adding another route for transferring risk.