Compute futures are contracts on the future price of computing power, usually the cost of renting a particular type of GPU for an hour. They let the people who rent out GPUs, and the people who rent them, fix today the price of compute they will sell or buy later, without negotiating a private contract with each other.
The idea is not new. Farmers, miners, refiners and airlines have used futures for generations to manage the prices they depend on. What is new is applying it to the hardware that powers AI. In 2026, several large derivatives exchanges in the United States announced or launched compute futures, and the Commodity Futures Trading Commission asked for public comment on how compute derivatives should be overseen.
What is being priced
The usual unit is the GPU-hour: one GPU, rented for one hour. A contract specifies which chip it covers, because each generation does a different amount of work per hour and rents at a different price. It also specifies how much compute one contract represents, for example a month of hours for one GPU, and the period it covers.
How GPU compute is priced explains why two quotes for the same chip can differ: the server around it, the network, the location and what else is included all matter. A futures contract has to define its unit precisely enough that everyone trading it means the same thing.
How settlement works
Most compute futures are designed to be cash-settled. Nobody delivers or receives a GPU. Instead, when the contract ends, the difference between the agreed price and a published reference price is paid in cash. A buyer who agreed to $2.20 an hour, when the reference settles at $2.60, receives the 40-cent difference on every hour the contract covers.
Cash settlement keeps the hedge separate from the supply. A company can rent its GPUs from whichever provider suits it, and use futures only to manage the price. But it puts great weight on the reference price. If the index does not reflect what compute really costs, the contract does not protect anyone.
Dated futures and perpetuals
A dated future covers a particular period and ends on a fixed date. A hedger who needs protection for longer has to roll from one contract to the next as each expires.
A perpetual future has no expiry. Positions stay open for as long as the holder wants, and a periodic funding payment keeps the contract's price close to the reference. Perpetuals suit exposures that run for years, such as a data center's income or a company's ongoing compute budget.
Who uses them
- Data centers and GPU clouds. They sell futures to protect against falling rental rates, so they can show lenders a predictable income.
- AI developers and enterprises. They buy futures to protect against rising rates before a large training run or a product launch.
- Lenders and investors. Their loans and investments depend on future rental income. Hedging, by them or their borrowers, makes that income more predictable.
- Traders and market makers. They take the other side when natural buyers and sellers do not arrive at the same moment, adding liquidity.
A worked example
With hypothetical numbers: a startup plans to rent 64 GPUs for six months, starting next quarter. That is about 280,000 GPU-hours. The futures price for that period is $2.20 an hour, and the startup buys futures covering the same hours.
If rental rates rise to $2.80, the startup pays about $168,000 more for its capacity than planned. Its futures gain the same 60 cents an hour, about $168,000, and its total cost stays close to $2.20 an hour. If rates fall to $1.80, it pays less for capacity and loses on the futures, again ending close to $2.20. Either way, it knows its compute budget in advance.
A compute future lets both sides of the market fix the price of a GPU-hour before they need it.
What a new market has to get right
- The reference price. Compute is sold in many configurations, on many terms. The benchmark has to be representative, resistant to manipulation and governed independently.
- Basis risk. A company's actual compute costs will not match the index exactly. The closer the match, the more useful the hedge.
- Liquidity. Hedgers want a liquid market before they use it, and a market becomes liquid only when hedgers use it. Breaking that circle takes market makers and a benchmark people trust.
- Product cycles. New chips arrive roughly every year. Contracts on one generation lose relevance as the market moves on, and liquidity can split across generations.
From compute to inference
Compute is the input. Many companies never rent a GPU at all; they buy inference through an API and pay per token. For them, the price that matters is the price of inference, which depends on compute but also on how efficiently models and software turn GPU-hours into answers. What are inference futures? looks at that market.
Carry Exchange is building markets for both, alongside the infrastructure behind them, so participants can manage the whole chain of exposure in one place. Managing the price of GPU compute looks at the risk in more depth.
Frequently asked questions
Are compute futures physically delivered?
Most are designed to be cash-settled against a published reference price, so no GPU capacity changes hands. The hedger buys or sells actual compute separately, from any provider.
What is the unit of a compute future?
Usually the price of one GPU-hour for a specific chip, with each contract covering a set number of hours, such as a month of rental for one GPU.
Who sells compute futures?
Typically the owners of GPU capacity, such as data centers and GPU clouds, who want to protect their income against falling rental rates, along with traders and market makers.
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.