Please see relevant disclosures here.
What follows is a primer for the rapidly developing compute derivatives market. I have received a lot of inbound inquiries from people looking to understand this market better. This primer synthesizes many of those conversations.
While this primer mentions a number of the principal players in the compute derivatives market, I make no endorsement of any of them; “do your own research” remains underrated advice. This is for educational use and is not investment, legal, tax, or accounting advice.
It’s important to note that this is a very nascent market, and some of what appears below might not come to pass, or might change materially as the market matures.
Introduction
This primer explains the emerging market for compute derivatives: financial contracts whose value references the price of renting compute capacity, especially GPU capacity used for AI training, inference, rendering, simulation, and high-performance computing.
The simplest version is a cash-settled futures contract on a standardized GPU rental price index, such as a U.S. H100 GPU-hour index. A buyer of the futures contract benefits if the reference price rises. A seller benefits if it falls. No GPUs need to be physically delivered; the contract can settle financially against an index.
This is no longer purely hypothetical. On May 12, 2026, CME Group and Silicon Data announced plans to launch a compute futures market later in 2026, pending regulatory review. On May 19, 2026, ICE and Ornn announced plans to launch U.S. dollar-denominated, cash-settled GPU compute futures based on Ornn’s Compute Price Index. And on May 28th, 2026, Architect announced plans to launch compute futures on a new US-based exchange.
1. The Basic Idea
Compute has become a material input cost for AI companies. A company that expects to rent thousands or millions of GPU-hours is exposed to rental price changes, capacity shortages, regional availability, hardware generation shifts, and contract terms.
A compute derivative turns part of that operational exposure into a tradable financial exposure. The underlying might be:
USD per H100 GPU-hour
USD per H200 GPU-hour
USD per B200 GPU-hour
USD per cluster-hour for a standard multi-GPU configuration
USD per standardized inference token
A broader basket of accelerator compute prices
The most likely early products are cash-settled futures and swaps referencing GPU rental price benchmarks. Physically delivered contracts are possible but harder, because actual compute depends on location, networking, uptime, security, software, minimum reservation windows, queue time, and data movement.
2. Why Now
AI infrastructure has moved from a technical procurement item to a financial planning problem. Several things now exist at the same time:
Large and recurring GPU demand from AI labs, model API companies, enterprises, research institutions, rendering businesses, simulation users, and autonomous systems companies.
Material price dispersion across cloud providers, neoclouds, and marketplaces.
Public, programmatic GPU rental markets, such as Vast.ai, where prices are set by supply and demand across a large listed inventory.
Cloud reservation products, such as AWS EC2 Capacity Blocks for ML, that already treat GPU capacity as something buyers schedule, reserve, and budget for in advance.
Emerging index providers and exchange partners attempting to create benchmark prices for compute.
NVIDIA’s first-quarter fiscal 2027 results, reported on May 20, 2026, showed record revenue of $81.6 billion, up 85% year over year, illustrating the scale of AI infrastructure demand. This does not prove that GPU rental prices will rise, but it explains why buyers, sellers, financiers, and traders now care about compute as a priceable input.
3. What This Market Would Solve
Price Risk
AI companies often quote product prices, sign customer contracts, or plan model launches before knowing their final compute cost. If GPU rental prices rise, margins compress. Futures let the buyer hedge that input risk.
Example: an AI application company expects to need 2 million H100 GPU-hours over the next six months. It buys H100 futures. If the index rises, gains on the futures help offset higher rental costs. If the index falls, it loses on the futures but benefits from cheaper physical compute.
Revenue Risk
GPU owners, neoclouds, and datacenter operators are naturally exposed to falling rental prices. If they finance a fleet assuming a certain utilization and rental rate, a decline in GPU-hour prices can impair revenue.
Example: a neocloud expects to sell H100 capacity for the next year. It sells H100 futures. If rental prices fall, the short futures position gains and offsets lower physical revenue. If prices rise, the hedge loses, but the provider earns more from physical capacity.
Financing
Better benchmark pricing can help lenders underwrite GPU fleets. If a lender can observe forward prices and hedge some revenue risk, it may be more comfortable financing new datacenter capacity.
Budgeting And Procurement
Compute buyers can separate two decisions:
Where to run workloads physically
How much price risk to carry financially
That separation is common in energy, freight, metals, rates, and currencies. A company may buy the physical input from one provider while hedging price exposure on an exchange.
Price Discovery
A liquid futures curve would show the market’s view of future GPU rental prices by month or quarter. That could inform procurement, cloud pricing, debt financing, depreciation assumptions, and equity market analysis of AI infrastructure companies.
4. Natural Longs And Natural Shorts
In this context, the cleanest way to describe exposure is physical exposure first, then the likely futures hedge.
The natural shorts are the obvious first hedgers: they feel GPU price spikes immediately. But the natural longs may be the more important source of sell-side hedging flow if they are financing fleets with debt and need to stabilize cash flows.
5. Speculators
Speculators take price risk without needing physical compute. They matter because hedgers rarely arrive in perfectly matched pairs.
Likely speculator types include:
Commodity and macro funds trading AI infrastructure supply-demand.
Equity investors hedging or expressing views related to AI capex, neoclouds, chipmakers, or datacenter companies.
Relative-value traders comparing H100 vs H200 vs B200 curves.
Calendar-spread traders betting on near-term scarcity versus future supply.
Volatility traders using options once listed.
Data-driven traders arbitraging exchange prices against live cloud quotes, marketplace prices, capacity reservations, and bilateral forward deals.
Speculation is not just casino activity in a functional derivatives market. It can absorb risk that natural hedgers want to shed. The danger is excessive speculation without a strong underlying cash market, which can create noisy price discovery and poor hedging performance.
6. Liquidity Providers
Liquidity providers quote bids and offers, hold inventory, and help participants enter or exit trades without moving the market too much.
In compute derivatives, liquidity providers would need to understand:
GPU generation performance differences.
Cloud pricing models.
Spot, on-demand, reserved, and committed-use pricing.
Regional and availability-zone differences.
Cluster topology and interconnect quality.
Index methodology.
Correlations to power, datacenter capacity, chip supply, public equities, and AI model demand.
Their hedges may include other compute futures, OTC swaps, direct provider capacity agreements, cloud reservations, power derivatives, datacenter equities, chipmaker equities, or simply diversified market-making inventory. Early liquidity will likely be expensive because the basis risk is real and the data history is short.
7. Exchanges And Clearing
An exchange standardizes the product: contract size, listed months, tick size, last trading day, settlement method, margin, reporting, position limits, and trade surveillance.
The CFTC’s futures education materials describe a futures contract as an agreement to buy or sell a commodity at a future date, with price and amount fixed when the agreement is made. The CFTC also notes that many contracts allow cash settlement rather than delivery. In the U.S., public futures and options trading generally must occur on regulated exchanges unless an exemption applies.
Clearing matters because the clearing house becomes the buyer to every seller and the seller to every buyer. That reduces bilateral counterparty risk but introduces margin requirements. Futures traders post margin, positions are marked to market daily, and losses can require additional variation margin.
For compute futures, regulated exchange infrastructure is especially important because the underlying market is fragmented and operationally complex. The exchange does not solve all basis risk, but it can make contract performance, credit risk, and settlement mechanics more standardized.
8. Index Providers
The index provider may be the most important actor in the market. A compute futures contract is only as useful as the benchmark it settles against.
A strong compute index should answer:
Is it based on actual transactions, executable quotes, scraped offers, surveys, or a blend?
Does it include only available inventory or also unavailable listed prices?
Does it normalize H100 SXM versus H100 PCIe?
Does it adjust for region, power cost, uptime, reliability, networking, CPU, RAM, storage, and egress?
Does it measure single-GPU rental, multi-GPU node rental, or large co-located cluster capacity?
Does it include spot, on-demand, reserved, marketplace, and private-contract prices?
What happens during outages, stale quotes, or provider API failures?
Can a provider manipulate settlement by printing small trades or changing listed quotes?
Is the methodology public, versioned, audited, and resistant to conflicts of interest?
The most important distinction is transaction-based versus quote-based pricing. A quote can be visible but not truly executable at scale. A transaction proves someone actually traded, but transaction data may be narrower and controlled by a small set of venues.
9. Contract Design Choices
A simple first contract might look like this:
This is only illustrative. Actual contracts may use different multipliers, final settlement windows, region definitions, hardware classes, and margin models.
The core tradeoff is standardization versus hedge precision. A highly standardized contract attracts liquidity but may not match a user’s exact workload. A highly customized contract hedges better but fragments liquidity.
10. Product Roadmap
If the first futures contracts work, the market could expand into:
H100, H200, B200, B300, RTX 5090, and accelerator basket futures.
Calendar spreads between near-term and long-term compute.
Basis spreads between regions or provider classes.
Options on compute futures.
Compute price caps and collars for AI buyers.
OTC swaps for large custom hedges.
Physically settled capacity forwards for large clusters.
Inference token futures or swaps.
Power-adjusted or carbon-adjusted compute indices.
Volatility indices based on compute futures options.
Inference may eventually matter more than training. A GPU-hour benchmark is easier to define, but an inference-token benchmark may map more directly to AI application economics.
11. Why The Market Might Work
The bullish case is straightforward:
The underlying cash market is large and growing.
Buyers and sellers have recurring, opposite exposures.
Compute prices are volatile enough to create hedging demand.
Existing procurement tools do not fully solve financial price risk.
Futures can cash settle, avoiding physical delivery complexity.
Major exchange announcements validate institutional interest.
Benchmark data improves financing and valuation even before futures become liquid.
The analogy is not perfect, but compute increasingly behaves like a production input: a company needs a certain amount, at a certain time, at a tolerable cost, to produce revenue.
12. Why The Market Might Fail
The failure case is just as important.
Basis Risk May Be Too High
An H100 GPU-hour is not always an H100 GPU-hour. Region, interconnect, uptime, interruptibility, storage, CPU, RAM, security, and cluster scale all change the real economic value. A startup needing 512 co-located GPUs next Tuesday may not be hedged by a broad H100 index.
The Cash Market Is Fragmented
Hyperscalers, neoclouds, marketplaces, private bilateral deals, reserved capacity, and capacity blocks all price differently. Many important transactions are private.
Availability May Matter More Than Price
The buyer’s problem may not be “$3.50 versus $4.00 per hour.” It may be “can I actually get 2,000 GPUs in the right region for three weeks?” A price index only partly captures availability.
Index Trust Is Hard
If the settlement index is based on quotes that are not executable, thin transactions, conflicted contributors, or easily influenced prints, hedgers will not trust it.
Natural Hedgers May Not Trade
Large cloud providers may prefer bilateral contracts and may not want to reveal or hedge economics publicly. AI labs may prefer strategic cloud commitments. Smaller startups may lack the credit, governance, or margin capacity to trade futures.
Technology Cycles Are Fast
GPU generations turn over quickly. A liquid H100 contract may be less useful when workloads migrate to H200, B200, custom ASICs, TPUs, or inference-optimized hardware. Too many hardware-specific contracts could split liquidity.
Existing Products May Be Good Enough
Reserved instances, committed-use discounts, cloud credits, enterprise cloud contracts, AWS Capacity Blocks, and direct neocloud reservations already give buyers some cost and capacity certainty. A derivative must solve what those products do not.
Liquidity Chicken-And-Egg
Hedgers want tight spreads and depth before trading. Market makers want hedger flow before quoting tight spreads. New futures contracts often take years to mature, and many never do.
Regulatory And Operational Complexity
Exchange listing, clearing, customer eligibility, margin, tax, accounting treatment, and internal risk approvals all slow adoption. A derivative that is elegant in theory can still be too hard to operationalize.
13. How To Evaluate A Compute Derivatives Market
For an index:
What exact physical exposure does it represent?
Is the methodology public and stable?
Are prices transaction-based, quote-based, or survey-based?
Are contributors independent from traders?
How much history exists?
How does it handle outliers, unavailable inventory, and stale data?
Can a hedger replicate or verify the index?
For a futures contract:
What is the contract multiplier?
What is the settlement window?
How large is initial and maintenance margin?
What months are listed?
What are position limits and block trade rules?
Is it cash-settled or physically settled?
Is it centrally cleared?
Who are the likely market makers?
For a hedger:
How closely does the index match actual compute spend or revenue?
What hedge ratio should be used?
What happens if physical prices and futures prices diverge?
Can the company handle margin calls?
Who approves the hedge program internally?
Is the goal cost certainty, budget protection, revenue stabilization, or speculation?
14. Market Map
Physical compute suppliers include hyperscalers, neoclouds, GPU marketplaces, datacenter operators, and specialized inference providers. Public examples include AWS, Google Cloud, Microsoft Azure, Lambda, CoreWeave, Crusoe, Nebius, RunPod, Vast.ai, and others.
Procurement substitutes include on-demand instances, spot instances, committed-use discounts, reserved instances, enterprise cloud contracts, short-term capacity reservations, and AWS EC2 Capacity Blocks for ML.
Financial market infrastructure is emerging. CME Group and Silicon Data have announced planned compute futures. ICE and Ornn have announced planned GPU compute futures based on Ornn’s Compute Price Index. Other index and data providers are also trying to measure GPU rental markets.
15. The One-Sentence Version
Compute derivatives try to do for AI infrastructure what energy futures do for power and fuel: turn a volatile, operational input into a transparent, hedgeable financial risk.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:

Interesting read! I can definitely a GPU futures market opening up in the near future
Along the same lines, this post from the great Corey Quinn: https://artificialconfidence.com/p/artificial-confidence-xai-the-neocloud