CME compute futures:
launch postponed.
The announced H100 and B200 rental-index contracts remain pending. CME’s September 28 notice postponed the launch for regulatory review. Here is the proposed contract design and the risk it would transfer.
The signal
CME’s September 28 Globex notice postpones the launch pending regulatory review and says further updates will appear in Globex notices. The original October 5 trade date is historical, not confirmation of trading. Check the venue roadmap for later changes.
CME Group and Silicon Data originally announced plans to launch two futures contracts on October 5, 2026, subject to regulatory review. One references the hourly rental cost of NVIDIA H100 capacity; the other references the Blackwell B200.
Each contract is designed to represent a month of GPU rental and settle in cash against Silicon Data benchmarks. The result is not physical delivery of a rack. It is a standardized payment tied to how a reference rental market moved.
CME's August 17 Globex notice identifies product codes GPU1 and GPU2 on MDP 3.0 channel 386. The September 28 notice describes New Release customer testing; confirm current testing access directly with CME. Customer testing is separate from production trading; the launch is now postponed pending regulatory review.
Each contract covers 730 GPU-hours. CME lists a $7.30 tick value, Bloomberg codes SDH100RT and SDRB200RT, and 36 monthly contracts.
What gets standardized
The breakthrough is the reference rate. GPU capacity has historically traded through cloud price pages, bilateral agreements, marketplaces and long reservations that are difficult to compare. A daily benchmark creates common language; listed contracts across expiries could support a forward curve and a venue for transferring price risk if trading develops.
For an AI lab, a long futures position could offset part of the damage from rising rental rates. A neo-cloud with capacity coming online could sell futures to reduce exposure to falling prices. An investor could express a view on compute scarcity without operating hardware.
Like energy, compute is an input cost with volatile supply and demand. Unlike a barrel of oil, one GPU hour is not fully fungible: workload performance depends on hardware, topology, region, software and reliability.
Who may care first
AI builders
Training and inference buyers can budget future capacity with a public reference instead of relying only on vendor quotes. The hedge will be strongest when their actual procurement closely tracks the settlement index.
Cloud and infrastructure operators
Providers financing expensive fleets face both utilization risk and price risk. A forward curve can inform underwriting, contract negotiation and the value of committed capacity.
Capital markets
Banks, insurers, lessors and infrastructure funds need defensible assumptions for GPU-backed loans and securitizations. A transparent curve may become an input to those models even before futures liquidity is deep.
The risk that remains
The hardest problem is basis risk: the difference between the index and a participant’s actual economics. An eight-GPU cluster with InfiniBand in one region is not interchangeable with a single GPU instance elsewhere. Reservation length, egress, support, availability and software images can matter as much as the nominal chip.
Futures also add leverage and liquidity risk. A hedge can generate margin calls before the underlying procurement benefit arrives. Early contracts may have wide spreads or concentrated participation. “Tradable” does not automatically mean liquid.
What to watch next
- Regulatory status: confirm the venue’s next dated update.
- Production readiness: New Release customer testing does not establish production trading availability.
- Volume and open interest: price discovery improves when buyers and sellers represent different parts of the ecosystem.
- Index breadth: coverage, normalization and resilience to configuration changes are critical.
- Competing venues: ICE and Ornn have announced transaction-based compute futures, while other market designs are emerging.
The larger change is already visible: compute buyers, sellers and financiers are starting to speak the same market language. A forward price will not make every GPU identical. It may, however, make the risk around those GPUs easier to see.
Primary sources
- CME September 28 notice: launch postponementOPEN ↗
- CME Globex notice: product codes, channel and testing dateOPEN ↗
- CME Group + Silicon Data launch announcementOPEN ↗
- CME Compute Futures product pageOPEN ↗
- Silicon Data index overviewOPEN ↗
- ICE + Ornn announcementOPEN ↗
Related: CL-FCR reference index, GPU rental price mechanics and compute market board.
SiliconTicker has no disclosed financial relationship with the venues or index providers cited. This article is educational and not investment advice.