Power futures face a challenge uncommon in traditional commodity markets: GPU hours cannot be stored, and the actual value can vary significantly due to differences in network, software, location, and reliability among chips of the same model.
As the cost of renting AI computing power gradually becomes a business expense, the U.S. futures market is preparing to launch a new contract based on GPU rental rates. According to Bloomberg, ADM Investor Services, BGC Group, and several brokerage firms are already preparing for this business, but the proposed compute power futures still require regulatory approval from the U.S. Commodity Futures Trading Commission (CFTC).
CME Group, ICE, Nodal Exchange, and Architect Financial Technologies are all vying to launch related products. Regulatory authorities required the completion of the product feedback process in August this year; however, the exchanges have not yet received approval for listing, and brokers’ business preparations do not equate to the contract having commenced trading.
Brokers and market makers position themselves in advance.
ADM Investor Services is considering hiring additional researchers to provide market analysis on computing power for potential clients. BGC established a new department this June dedicated to secondary trading and real-time risk management of computing power and storage capacity; Zach Espinosa, co-head of BGC’s computing power infrastructure market, believes that contract standardization will help attract liquidity.
Previously, foreign media reported that the proprietary trading firm Jump Trading has formed a dedicated team to participate in over-the-counter transactions related to computing power using its own capital and to provide liquidity for upcoming exchange-listed contracts. DRW plans to involve three trading teams, focusing on price indices and derivatives; both companies are moving forward while CME Group’s products are still pending approval.
Potential demand comes from data center operators, cloud service providers, and AI developers. Compute providers may wish to lock in future rental income in advance, while buyers aim to mitigate the impact of rising compute prices for training and inference tasks; financial trading institutions are also exploring the price differential between compute rentals and costs of electricity and natural gas.
The rental rates for H100 and B200 will serve as the contract benchmark.
The Financial Times previously reported that CME Group and index provider Silicon Data have selected the hourly rental prices of NVIDIA's H100 and B200 chips as the reference for the proposed contract, aiming to establish a forward price curve of up to 36 months. Different models will correspond to separate contracts to avoid treating computing power with significantly different performance levels as the same commodity.
The Financial Times, citing Silicon Data, reported that at the time of writing, the hourly rental rate for the B200 was $5.86 and for the H100 was $2.77. The article also noted that during the chip supply shortage in early 2024, the hourly rental rate for the H100 briefly reached $8, but had fallen below $2 by the end of 2025; these figures represent historical quotes from different time points and should not be considered current uniform market prices.
Market standards still face challenges. Even with identical chip models, network configurations, software, geographic locations, and reliability of server clusters can affect actual available performance; rental price indices calculated by different data providers are not entirely consistent. Futures prices may deviate from companies’ actual procurement costs, meaning contracts may not fully hedge against cost fluctuations.
The forward price curve still requires market validation.
Computing power also has the characteristic of being non-storable. Unused GPU hours cannot be stored like oil inventories for sale next month, making it difficult to replicate traditional arbitrage mechanisms between spot storage and futures. Don Wilson, founder of DRW, believes that reliable transaction data and price indices are prerequisites for the development of tradable derivatives in this market.
Dave Olsen, President and Chief Investment Officer of Jump, expects financial trading institutions to enter the market before companies that actually purchase computing power. Companies will need to observe whether the cash-settled price accurately reflects their own procurement costs before deciding to hedge on a large scale. Traders have already deployed teams and capital, but the contract launch date and final liquidity have yet to be confirmed.