Contact Us
Compute futures are coming. The exchange is the easy part. Whoever defines a compute-hour decides whether the market works.
Adapted from Addison Armstrong’s presentation at Yotta 2026.
AI compute is bought years forward, in private contracts, at prices nobody publishes. Lenders finance GPU fleets against those prices. Buyers build budgets on them. Today the forward curve is pure opinion.
A futures market is supposed to fix that. Three venues are competing to build it. The US regulator is still deciding whether the first contracts can list.
The contract is simple by design. One contract covers 730 GPU-hours, which is one chip running for one month. It settles in cash against a published index of rental prices. Contracts list monthly, 36 months out, so the curve reaches into 2029.
No chips change hands. A cloud provider worried about falling rental rates sells. An AI lab worried about rising rates buys. Each locks a price against the index and keeps its real contracts as they are.
CME and Silicon Data had set October 5 as the first trade date for two contracts, on Nvidia H100 and B200 rental prices. On September 21 the CFTC extended its review to November 9, citing novel or complex issues. Its public comment period on compute derivatives closes October 20.
A futures contract is only as good as the index it settles against. In compute, the first question is who defines a compute-hour.
In one recent week, five published prices for one hour of the same chip ran from $2.53 to $7.46. That is a 3x spread. The two numbers the rival exchanges plan to settle against differed by 10.2% on the same day. One index counts only printed transactions. That is the best data available and a narrow slice of the market.
The gap between the index and your own price is basis risk. Your futures position pays on the index. Your business pays or earns something else. A hedge tied to the wrong number is a dirty hedge.
The index also has to hold up against the people who feed it. A participant with a futures position and a price list can move the number without breaking a rule. The CFTC is asking who can influence these prices before it lets the contracts list.
Hedging has a price. One tick is 37.6 basis points of the contract, and a bigger share as rental prices fall. Margin is topped up daily. That is the toll on holding a position.
Then comes liquidity. CFTC economists have tracked US-listed contracts back to 1954, and in any given year about 80% show little or no volume. Exchanges list anyway. A contract survives when natural hedgers on both sides trade it.
Almost every party in this market can be replaced. Exchanges, clearing houses and market makers can be swapped out in months. Two cannot. Hedgers are the demand itself, and an index is close to permanent once contracts cite it.
Weather futures launched in 1999 on something non-storable and undeliverable, and they worked. The settlement data came from weather stations with no stake in the outcome.
Bandwidth is the closer analogue, and it died. Capacity was different on every route, carriers would not make it interchangeable, and Enron was both architect and largest counterparty. Every contract that survived solved its settlement source first.
Hedging a volatile input you cannot store is the daily work of energy risk management. Continuum Partners Energy Advisors does that work in power today. We are watching compute take the same path.
The contract is small. One chip, one month. If it succeeds, hundreds of billions in GPU financing get cheaper. If it stalls, the buildout continues anyway, priced in private.
Addison’s full white paper on compute derivatives is in final review. To get a copy when it’s released, contact our Energy Advisors team.
Whether you need a site that can carry the load or a strategy for managing the energy behind it, the conversation starts the same way.
Rooted in energy.
