Inference futures are contracts on the future price of AI inference: what it costs to get answers, code and analysis from AI models, usually measured in US dollars per million tokens. They let companies that buy inference, and companies that sell it, fix today the price of inference they will buy or sell later.
They are sometimes called AI token futures. The idea has moved quickly from theory toward practice. Researchers have published proposals for how such contracts could be designed, new venues have launched markets on AI token prices, and a major Chinese futures exchange was reported in 2026 to be designing token futures of its own.
Why inference needs its own market
For most companies using AI, inference is the cost they actually see. They never rent a GPU; they pay per token through an API, and the bill grows with usage. How enterprises manage AI inference costs shows how that cost is becoming a major line on the P&L.
Compute futures do not cover this exposure well. The price of inference depends on the price of compute, but also on how efficiently models and software turn GPU-hours into tokens, on competition between model providers and on how quickly new models arrive. Inference can get cheaper while GPU rents hold steady, or the other way round. A company exposed to the price of tokens needs a hedge on the price of tokens.
What is being priced
This is the hard part. A barrel of crude or an hour on a given GPU is a defined product. A token is not: a million tokens from a small model and a million from the most capable model are very different things, at very different prices, and models improve every few months.
An inference contract therefore needs a reference price that compares like with like over time. One approach is to hold capability constant: to price a defined level of intelligence, so that the index moves when the cost of that capability moves, not when buyers switch to bigger or smaller models. Measuring price at constant capability explains why this matters.
For inference futures, the benchmark is the product. A contract is only as useful as the price it settles against.
How settlement works
Tokens cannot be stored or delivered later, so inference futures are cash-settled. When a contract ends, the difference between the agreed price and the benchmark is paid in cash. The hedger keeps buying inference from whichever providers it chooses, and the futures offset changes in what that inference costs.
As with compute, contracts can be dated, ending on a fixed date, or perpetual, staying open for as long as the exposure lasts.
Who would use them
- Companies building on AI. Software companies, banks, retailers and others that sell AI-powered products at fixed prices are exposed if inference costs rise. Buying futures protects their margins.
- Model providers and inference clouds. Their revenue falls if market prices fall. Selling futures fixes part of it in advance.
- Resellers and platforms. Companies that buy inference wholesale and sell it on carry the gap between the two prices, and can hedge either side.
- Traders and market makers. They provide liquidity and take on risk that natural hedgers want to pass on.
A worked example
With hypothetical numbers: a software company expects to use about 200 billion tokens a month over the next year, at an average of $2 per million tokens. That is about $400,000 a month, or $4.8 million a year, and it has priced its product on that basis.
If the market price of the capability it relies on rises by 25%, its costs rise by about $1.2 million over the year. A hedge sized to that exposure gains about the same amount, and the company's margin holds. If prices fall by 25%, its costs fall and the hedge loses about the same, so it ends close to the cost it planned.
Because the price of a given capability has generally fallen over time, many buyers may prefer options to futures: an option can cap what they pay while letting them keep the benefit if prices fall. Designing a hedge that fits covers how to size and review a hedge.
What a new market has to get right
- The benchmark. It has to compare like with like as models change, be governed independently and be resistant to manipulation.
- Basis risk. A company's own costs depend on the models it uses and its supplier terms. The closer the benchmark tracks those costs, the better the hedge.
- Liquidity. As with any new market, hedgers want depth before they commit, and depth comes from hedgers committing.
- Trust. Contracts that settle against a price need everyone to accept that price. What makes a price index trustworthy sets out the foundations.
Where Carry fits
CICI, the Carry Inference Cost Index, is Carry's flagship benchmark for the price of AI inference. It measures the price of inference at constant capability and is designed as a reference for hedging and settlement. CRTI, the Carry Realized Token Index, measures what the market actually pays per million tokens. Carry Exchange is building markets for AI inference and compute around them.
Frequently asked questions
Are inference futures the same as AI token futures?
Yes. Both terms describe contracts on the future price of AI inference, usually quoted in US dollars per million tokens.
How are inference futures settled?
In cash, against a published benchmark for the price of inference. Tokens cannot be stored or delivered later, so no inference changes hands under the contract.
How do inference futures differ from compute futures?
Compute futures track the cost of renting GPUs by the hour. Inference futures track the price of the AI output itself, per million tokens. The two are linked, but they can move differently as models and software become more efficient.
This article is for information only and is not investment advice. Derivatives involve substantial risk of loss and are not suitable for all investors. The figures in the example are hypothetical, and product availability depends on jurisdiction.