CME Compute Futures are a planned family of financially settled futures contracts tied to Silicon Data benchmarks for hourly rentals of Nvidia H100 and B200 graphics processors. New York Mercantile Exchange, Inc. (NYMEX), a CME Group exchange, announced two contracts: Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures. CME has set October 5, 2026 as the target first trade date, but both its product page and the exchange notice make that date conditional on regulatory review.[1][2]

The contracts are intended to turn changes in standardized GPU rental rates into a cash-settled financial exposure. They do not reserve, deliver, or guarantee access to physical computing capacity. As of August 15, 2026, they had not begun trading, so no trading volume, open interest, bid-ask spread, or adoption record existed for the products.[1][2]

Field	H100 contract	B200 contract
Official title	Silicon Data H100 Rental Index Futures	Silicon Data B200 Rental Index Futures
Exchange	NYMEX	NYMEX
Product code	`GPU1`	`GPU2`
Rulebook chapter	1045	1047
Contract size	730 GPU-hours	730 GPU-hours
Quotation	US dollars and cents per GPU-hour	US dollars and cents per GPU-hour
Minimum price movement	$0.01 per GPU-hour	$0.01 per GPU-hour
Value of one tick	$7.30	$7.30
Settlement type	Financial	Financial
First listed month	October 2026, planned	October 2026, planned

Announcement and status

CME Group and Silicon Data announced their partnership on May 12, 2026. The initial release said the companies planned to introduce compute futures later in the year, pending regulatory review, and that the products would use Silicon Data's daily benchmarks for on-demand GPU rental rates. It did not yet identify the individual contracts or their sizes.[3]

NYMEX supplied those details in Special Executive Report 9785 on August 11. The notice states that the exchange intends to list the contracts effective Sunday, October 4, for trade date Monday, October 5, 2026, pending all relevant US Commodity Futures Trading Commission review periods. CME's dedicated product page gives the same October 5 date and labels it pending regulatory review.[1][2]

The distinction between an announcement and a launch is material. Before the target date and completion of the stated review periods, the products are prospective. CME and Silicon Data describe possible benefits for AI builders, cloud providers, infrastructure owners, and financial institutions, but those descriptions are intended uses. They are not evidence that a liquid market or an effective hedge already exists.[2][3]

Contract design

Both contracts represent 730 GPU-hours and are quoted in dollars per GPU-hour. A quoted price of P therefore corresponds to a reference contract value of 730 x P. The minimum price change is one cent per GPU-hour, producing the $7.30 tick value published by NYMEX.[1]

Trading is planned on CME Globex, with trades also eligible for submission for clearing through CME ClearPort. The contracts use a first-in, first-out matching algorithm. NYMEX plans to list 36 consecutive monthly contracts, and trading in an expiring month is scheduled to end on the last business day of that month. The block-trade threshold is five contracts with a 15-minute reporting window.[1]

NYMEX classifies the settlement type as financial. The CFTC uses financial settlement and cash settlement as equivalent terms: an expiring position is settled by a cash amount based on an index or price rather than by delivery of the underlying commodity.[8] A holder therefore cannot use the contract to obtain an H100 or B200 instance, a cluster reservation, a service-level agreement, or a place in a data center. The instrument addresses an index-linked price exposure only.

The two-page exchange notice does not publish the final-settlement formula, the observation window used at expiration, initial or maintenance margin, or a contingency rule for unavailable index data. It also does not identify an index ticker. Those details should not be inferred from examples produced by vendors or trading platforms.[1]

Reference indices

Silicon Data publishes daily rental-price benchmarks for several accelerators. Its public methodology overview says it draws observations from cloud providers, colocation markets, brokered cluster sales, and private rental platforms. It says the data are standardized for machine specifications, rental terms, platform performance, cluster scale, interconnect, and geography, then processed with outlier removal and quality-control checks.[4][5][6]

For NVIDIA H100, Silicon Data publishes a neo-cloud reading under ticker SDH100RT and a separate hyperscaler reading. Both are described as standardized hourly rental rates. For NVIDIA B200, it publishes a single standardized series under ticker SDB200RT while provider-segment coverage develops.[4][5]

The contract names establish that one product references a Silicon Data H100 Rental Index and the other a Silicon Data B200 Rental Index. The available SER does not, however, specify which of the two public H100 readings controls settlement or confirm that the displayed tickers are the legal settlement series. A precise mapping requires the governing contract chapter or settlement methodology, not an assumption based on a marketing page.[1][4][5]

Silicon Data's methodology is also not static. A December 2025 index notice describes a restatement that removed a divisor adjustment and changed its pricing model and provider weights. The company estimated that change would reduce the H100 index by 4 to 6 percent. A provider expansion effective in April 2026 had an estimated negative effect of 3 to 7 percent without restating prior values.[7] These disclosed changes illustrate why index governance, provider coverage, and revision policies matter to a cash-settled contract.

Hardware scope

The products do not treat all computing resources as interchangeable. H100 is a Hopper-architecture data-center GPU, while B200 is part of the Blackwell generation.[13][14] Separate contracts allow the prices of the two rental markets to move independently. They do not imply that one GPU-hour on each device provides the same performance, memory capacity, energy use, networking capability, or value for a particular workload.

That distinction also limits the meaning of the generic phrase "compute futures." The announced CME products cover rental-price indices for two specific AI infrastructure components. They do not reference CPUs, custom AI accelerators, inference tokens, electricity, complete servers, or a performance-normalized unit of computing.[1][4][5]

Intended hedging use

CME presents the products as a possible way for buyers of compute to manage rising rental costs and for capacity providers to manage falling rental revenue.[2][3] In a simplified case, a buyer whose future H100 bill moves closely with the designated H100 index could take a position intended to gain value when the index rises. A provider with index-linked rental revenue could use the opposite exposure. The actual hedge ratio would depend on the size and behavior of the cash exposure.

This mechanism cannot lock in physical availability. A cash payout may offset part of a price increase, but it cannot supply GPUs if a region or provider has no capacity. It also cannot make a rented instance conform to the index's standardized configuration. The CFTC calls the risk that the relationship between a cash price and its futures hedge changes unexpectedly basis risk.[8]

Potential sources of basis risk include provider type, region, GPU form factor, host configuration, memory, interconnect, cluster scale, support terms, rental duration, and availability guarantees. Silicon Data says it normalizes several of those variables, but a benchmark cannot be assumed to match every cloud computing invoice or provider revenue stream.[4][6][8]

An August 2026 working paper by Federico M. Bandi and Yinan Su highlights a second difference from many physical commodities. Compute service is non-storable: a GPU-hour left unused today cannot be carried into a future month. The authors argue that ordinary cash-and-carry pricing does not directly connect today's spot rental rate to a later compute-futures price. They also distinguish a financial contract from a term rental, which includes actual access to capacity. Their empirical analysis uses synthetic forward curves because traded compute-futures returns did not yet exist, and its findings are preliminary rather than measurements of the CME products.[10]

Market context and open questions

CME called the project a "first-in-class" market in its May announcement. That wording is the company's positioning, not proof that it will be the first active venue.[3] Intercontinental Exchange and Ornn announced a separate suite of dollar-denominated, cash-settled GPU compute futures on May 19, subject to regulatory approval.[11] ICE and NATIVX announced another planned family based on an energy-normalized COIL Index on July 1, subject to completion of regulatory processes.[12] Axios also reported a proposed OneChronos compute market awaiting federal approval.[9]

These initiatives use different index concepts and cannot be treated as interchangeable. Their existence shows that compute derivatives are a developing category, while their conditional status makes any ranking by launch order premature.[9][11][12]

Several questions will remain open unless and until the CME products begin trading and fuller rules are available:

Which exact index series and calculation window determine final settlement?
How will corrections, methodology changes, or unavailable observations be handled?
What margins and position-accountability rules will apply at launch?
Will trading remain active across several monthly maturities?
How closely will the futures track the rental exposures of actual buyers and providers?

Until those points have public answers and observed market data, claims that the contracts create reliable price discovery, deep liquidity, cheaper financing, or consistently effective hedges remain projections rather than established outcomes.[1][2][7][9]

References

CME Group. SER-9785: Initial Listing of Two (2) Compute Futures Contracts - Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures. August 11, 2026. https://www.cmegroup.com/content/dam/cmegroup/notices/ser/2026/08/ser-9785.pdf
CME Group. Compute Futures. Accessed August 15, 2026. https://www.cmegroup.com/markets/energy/power/compute-futures.html
CME Group. CME Group and Silicon Data Partner to Launch First Compute Futures. May 12, 2026. https://www.cmegroup.com/media-room/press-releases/2026/5/12/cme_group_and_silicondatapartnertolaunchfirstcomputefutures.html
Silicon Data. H100 Rental Price Index. Accessed August 15, 2026. https://www.silicondata.com/products/silicon-index/h100
Silicon Data. B200 Rental Price Index. Accessed August 15, 2026. https://www.silicondata.com/products/silicon-index/b200
Silicon Data. GPU Rental Price Indices. Accessed August 15, 2026. https://www.silicondata.com/products/silicon-index
Silicon Data. GPU Index Announcements. Accessed August 15, 2026. https://docs.silicondata.com/products/gpu-index-announcements
US Commodity Futures Trading Commission. Futures Glossary. Accessed August 15, 2026. https://www.cftc.gov/LearnAndProtect/AdvisoriesAndArticles/CFTCGlossary/index.htm
Scott Rosenberg. The push to create a futures market for AI compute. Axios, August 12, 2026. https://www.axios.com/2026/08/12/ai-futures-compute-cme
Federico M. Bandi and Yinan Su. (Early) AI Compute Asset Pricing. Working paper, revised August 5, 2026. https://arxiv.org/abs/2607.12156
Intercontinental Exchange. ICE and Ornn to Launch GPU Compute Futures Contracts. May 19, 2026. https://ir.theice.com/press/news-details/2026/ICE-and-Ornn-to-Launch-GPU-Compute-Futures-Contracts/default.aspx
Intercontinental Exchange. ICE and NATIVX to Launch Energy-Normalized Compute Futures Contracts. July 1, 2026. https://ir.theice.com/press/news-details/2026/ICE-and-NATIVX-to-Launch-Energy-Normalized-Compute-Futures-Contracts/default.aspx
Nvidia. H100 GPU. Accessed August 15, 2026. https://www.nvidia.com/en-us/data-center/h100/
Nvidia. Supported GPUs. NVIDIA Multi-Instance GPU User Guide, accessed August 15, 2026. https://docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-gpus.html

Editing: CME Compute Futures