Compute futures, explained
The price of renting AI compute is becoming a market. What compute futures are, which indices they would settle on, and how to trade the same prices on Racks, with a cap on what you can lose.
Updated October 16, 2026
What compute futures are
Compute futures are contracts on what it costs to rent a GPU by the hour. They settle in cash against an index of rental prices, so no chip or server time changes hands. A company that rents GPUs can fix what its compute will cost; a trader can take a view on where the price goes.
The first ones follow NVIDIA’s H100 and B200, two of the GPUs most used to train and run AI models. An H100 rents for a few dollars an hour, and that price moves: with new data centers coming online, with demand for training and inference, and as each new generation of chips arrives.
The compute futures announced so far
CME Group plans Silicon Data H100 and B200 Rental Index futures (GPU1 and GPU2), to be listed under NYMEX rules on CME Globex and ClearPort. One contract is 730 GPU-hours, a month of one chip, settled in cash against Silicon Data’s H100 and B200 rental indices, with monthly expiries up to 36 months out. CME announced them for October 5, 2026; they await the CFTC’s approval: on September 21, 2026, the CFTC extended its review until the end of November 9, 2026 (NYMEX submission 26-370, at the CFTC; CME Group, August 11, 2026).
ICE plans OCPI H100 and B200 futures (HPR and BKL) on ICE Futures U.S., on Ornn’s Compute Price Index, pending the CFTC’s request for comment on compute futures, with no listing date set. Each would be a monthly contract settled in cash on the average of the month’s daily prices, each day the average of the hourly OCPI prices; one contract is the number of hours in the month (ICE Futures U.S. exchange notice, September 29, 2026).
Both would be exchange-listed futures: traded through a futures broker, a contract about a month of one GPU’s hours, and a position that can lose more than its margin.
Trading compute on Racks
Racks contracts follow the same kind of price, with a cap. Each GPU has an index of its rental price per hour, and each index counts four published rates: Compute Desk’s US 1-year desk rate, Liquid Compute’s US reference, Silicon Data’s rental index for neo-clouds and Ornn’s OCPI of on-demand rentals. Silicon Data and Ornn publish the indices that CME’s and ICE’s compute futures would settle on.
The price now and its history are on each index’s page: H100 rental price index and B200 rental price index.
Every period starts at 100, and the index shows how far prices have moved since, in percent: 104.21 means rental prices are up 4.21%. Up pays $1 for every point above 100 at settlement, up to $12. Down pays $1 for every point below 100, up to $12. A contract costs between $0 and $12, set by the market, and that is the most a buyer can lose. A writer locks $12 a contract and gets its price, so the most a writer can lose is $12 less that price. Contracts can expire worthless.
Weekly contracts open on Monday at 00:00 UTC and settle the next Monday at 00:00 UTC, on the index’s last value. Monthly contracts settle at 00:00 UTC on the 1st of the next month, on the month’s TWAP: the mean of the index at the end of every hour of the month.
Contracts can expire worthless. Prices are thin at times: large orders move the price, and spreads can widen. Index levels can be late when a price source is late or stops publishing, and a result can be challenged before it is final. Smart contracts can have bugs. More in the docs’ risks.
Everything runs on Robinhood Chain and settles in USDG. You trade from a wallet, at any hour, and every index level is posted on-chain with the hash of the file that lists the prices behind it.
Side by side
| Terms | Racks | CME | ICE |
|---|---|---|---|
| Follows | An H100 and a B200 index of four published rates, Silicon Data’s and Ornn’s among them | Silicon Data’s H100 and B200 rental indices | Ornn’s OCPI H100 and B200 |
| One contract | $1 for each index point, up to $12 | 730 GPU-hours | The hours in the month |
| Settles on | Weekly: the index at settlement. Monthly: the month’s TWAP | The month’s average of Silicon Data’s daily prices | The month’s average of daily prices |
| Most you can lose | A buyer: what they paid, at most $12. A writer: $12 less the price | Can exceed the margin | Can exceed the margin |
| Expiries | Weekly and monthly | Monthly, up to 36 months out | Monthly, up to 24 months out |
| Where | Robinhood Chain, from a wallet, at any hour | CME Globex and ClearPort, through a futures broker | ICE Futures U.S., through a futures broker |
CME’s and ICE’s terms as they announced them. As of October 7, 2026, neither trades yet: both await the CFTC.
Questions
Are Racks contracts futures?
No. They are capped Up and Down contracts on the same kind of price: each pays $1 for every point the index moves your way, up to $12, and costs at most $12. A future has no cap on what it can gain or lose.
Why trade the price of compute?
A company that rents GPUs can offset a rise in what its compute costs, and one that rents them out can offset a fall in what it earns. A trader can take a view on demand for AI through the price of the hardware it runs on.
Where do the Racks index prices come from?
Four published rates: Compute Desk’s US 1-year desk rate, Liquid Compute’s US reference, Silicon Data’s rental index for neo-clouds and Ornn’s OCPI of on-demand rentals. Every index level Racks posts on-chain comes with the hash of the file that lists the prices behind it.
Who can trade on Racks?
Anyone with a wallet on Robinhood Chain. Racks is not available to US persons, in the United States, or in sanctioned countries and regions.