Compute futures, also searched as GPU futures, are contracts on the price of renting GPU time. Two exchange groups have announced them. Neither had started trading when this page was written on 3 November 2026. This post explains what they are, what sits behind them and where they fall short.
What a future is, in plain words
A future is an agreement to settle a price on a set date. You agree today on a price for something you will use or sell later. If the market price moves, one side gains and the other loses.
Most financial futures settle in cash. No goods change hands. At the end of the period, the side that guessed right collects the difference. Silicon Data's own guide to its contract puts it this way: "No GPU changes hands; money does."
So a compute future does not deliver a GPU. It pays out on a price.
Why compute might need one
GPU time is a cost that swings. A buyer who rents a cluster for a year does not know what the market rate will be in month nine. A provider that has bought GPUs on credit does not know what they will earn. A lender behind that provider shares the same risk.
A future lets each side fix a price in advance, or take the other side of the risk. It also gives the market a public price. That matters because most GPU deals close in private. The list price is not the price, and the real price stays hidden. A traded contract puts a number in view every day.
The indices behind the contracts
A future needs a number to settle against. Here, that number is an index. Two have been named, and both are the companies' own descriptions of their methods.
- Silicon Data. It describes its indices as the first daily benchmarks for rental prices of on-demand GPU compute. CME's contracts settle against them.
- Ornn. Its Compute Price Index, or OCPI, is described in ICE's announcement as the first compute index built only from printed transactions. ICE's contracts settle against it.
The two methods differ. One looks at rental prices. The other says it uses prints of actual trades. Neither number tells you what a buyer of your size, term and region pays.
The announcements, with dates
CME Group and Silicon Data announced compute futures on 12 May 2026, pending regulatory review. On 19 May 2026 ICE and Ornn announced a suite of GPU compute futures. They are in US dollars and cash-settled, and "subject to regulatory approval". ICE said they may cover the H100, H200, B200 and RTX 5090 over time.
On 11 August 2026 CME set out two contracts, on Silicon Data's H100 and B200 rental indices, with a first trade date of 5 October 2026. Silicon Data's guide describes them as GPU1 and GPU2. One contract covers 730 GPU-hours, which is one GPU for one month. One cent of price movement is worth $7.30 per contract. The contract settles against the index averaged over the month, not against one day.
The October date did not hold. On 21 October 2026 Runtime Wire reported that regulators had not cleared the contracts, and that CME's calendar listed 9 November 2026 as the first trading day, pending review. As of 3 November, the contracts are due to start, not trading. ICE's product page for its H100 contract describes a monthly, cash-settled contract, with up to 24 consecutive months listed.
Who would use them
- Buyers. A company that rents GPUs fears rising rates. It buys contracts and gains if the index goes up.
- Providers. A provider fears falling rates. It sells contracts and gains if the index goes down.
- Lenders and investors. A lender to a provider cares what the GPUs will earn. Contracts give them a way to test, and perhaps cover, that risk. This is the logic of the product, not a statement from the exchanges.
Both exchanges point at these users. ICE's announcement quotes Ornn's chief executive on reaching institutional buyers and operators, and Hyperbolic Labs' chief executive on neocloud providers and AI labs managing compute exposure.
The limits
- Index quality. A future is only as good as its index. Two indices already exist, built differently. A buyer must ask what is in the number and how it is built.
- Basis risk. The index tracks a broad market. Your contract may track something else. Silicon Data's guide says a hedge protects against the market cycle, not against local contract renegotiations.
- Mismatch. The guide notes that GPU1 settles against an index of independent GPU clouds, so it may suit hyperscaler workloads poorly.
- Cash, not delivery. A contract pays money. It will not get you a cluster, in a hall, with power. Buying capacity still means buying capacity.
- Liquidity. A contract is useful only if others trade it. Silicon Data's guide warns that early spreads between buy and sell prices can be wide. Runtime Wire names enough real buyers and sellers as one of three conditions for success, with clearance and a trusted index.
- Cash needs. The guide notes that positions are marked to market each day, so a hedger needs cash ready when prices move against it.
Compute futures may help. They are new, unproven and not yet trading. Treat any claim about them as a forecast.
Sources
- Silicon Data, CME Group and Silicon Data partner to launch first compute futures, 12 May 2026
- ICE, ICE and Ornn to launch GPU compute futures contracts, 19 May 2026
- ICE, Ornn Compute Price Index H100 Future product page, read 3 November 2026
- CME Group, special executive report on Silicon Data rental index futures, 11 August 2026
- Silicon Data, a practitioner's guide to hedging with compute futures, read 3 November 2026
- Runtime Wire on the 9 November 2026 start, 21 October 2026
- Casitel, The list price is not the price, 3 November 2026