InfrastructureCompute, chips & cloud
CFTC review delays CME and Silicon Data GPU compute futures from October 5 to November 9 or later
CME Group and Silicon Data planned cash-settled futures on Nvidia H100 and B200 rental prices for October 5. A 45-day CFTC review extension pushed first trading to November 9 at the earliest, with regulators citing limited transparency in privately negotiated GPU rental pricing.
From the WSJ · Technology article of October 21, 2026 and cited sources: CME plans to list two contracts on NYMEX: H100 Rental Index Futures and B200 Rental Index Futures. [1] [4] [6]
SOURCES[1] Want to Short the AI Boom? She Wants to Give You a Way to… wsj.com[4] CME Group press release, May 12, 2026 cmegroup.com[6] PR Newswire: October 5 launch prnewswire.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- CME plans to list two contracts on NYMEX: H100 Rental Index Futures and B200 Rental Index Futures. [1] [4] [6]
- Original source
- WSJ · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 21, 2026 article published
- Covered
- Oct 22 edition
- Evidence
- 12 sources across 12 sites
Reach: not measured yet
CME plans to list two contracts on NYMEX: H100 Rental Index Futures and B200 Rental Index Futures. Each is cash-settled, represents one month of hourly GPU rental cost and settles against Silicon Data's daily rental indices. Selling the contracts would pay off if rental prices fall. That design is why they are being framed as a way to bet against the AI boom. [1] [4] [6]
The CFTC called the product novel or complex and used its one allowed 45-day review extension. It pointed to manipulation risk because GPU rental prices are set in private two-party deals. The agency had already opened a consultation on liquidity, manipulation, customer protection and perpetual contract structures. As of October 21 the contracts had still not been cleared, and the CFTC cannot extend the November 9 deadline again unless NYMEX agrees. [2] [3] [7] [9]
The settlement benchmark is hard to check from outside. Silicon Data has published its four-step method, but not its provider weights, adjustment coefficients or raw data. Its claim to cover about 80% of the global rental market comes from the company itself and has no independent audit. One analyst reports that its indices correlate only weakly week to week. Public trackers also diverge sharply on H100 prices, from under $1 an hour to more than $7 an hour, because they cover different providers and contract terms. [8] [9] [12]
Other venues are pushing their own benchmarks: ICE, Nodal Exchange and Architect each back a different GPU price index, and BGC has opened an over-the-counter compute market. For hedgers, the indices leave out availability guarantees, networking and service premiums. That means real contracts may not move in step with the index, and margin calls can strain cash flow. [7] [9]
What is known: a regulated market for GPU rental prices is close, pending CFTC clearance.
Read the full assessment
What it could mean: neoclouds, GPU lenders and AI labs could hedge compute costs or GPU resale values, but only as well as their contracts track an index outsiders cannot reproduce.
Executive brief
A regulated way to bet against AI compute prices is close, but regulators have not approved it yet. CME Group and Silicon Data, the GPU-pricing startup founded by Carmen Li, planned cash-settled futures on Nvidia H100 and B200 rental prices for October 5, 2026. The CFTC used its one allowed 45-day review extension, which moved the date to November 9 or later (BigGo Finance). As of October 21 the market still had not been cleared (Runtime Wire). The open question is whether one company's index can serve as the settlement price for a market whose rental prices are set in private deals.
What changed and event timeline
CME and Silicon Data announce compute futures
The two companies said they would launch the first futures on compute later in 2026, pending regulatory review. Contracts would settle against Silicon Data's daily GPU rental indices (;).
October 5 launch date set
CME named two NYMEX contracts: H100 and B200 Rental Index Futures, each covering one month of rental cost ().
CFTC opens a compute-derivatives consultation
The consultation asks about liquidity, manipulation, customer protection and perpetual contract structures. Chairman Michael Selig framed it as part of the AI race ().
CFTC extends its review by 45 days
The regulator called the product novel or complex. It pointed to GPU rental prices being set in private two-party (bilateral) contracts, which limits transparency and raises manipulation risk ().
WSJ profiles Li as the launch slips
The WSJ presents the contracts as a way to short the AI boom (). The same day, Runtime Wire reported that CME's calendar lists November 9 as first trading, still pending clearance ().
Capabilities and access
- Contracts: Silicon Data H100 Rental Index Futures and B200 Rental Index Futures, listed on NYMEX.
- Shorting: selling these futures would pay off if rental prices fall.
- Underlying indices: Silicon Data also publishes A100 and MI300X indices, plus separate neo-cloud and hyperscaler readings for the H100.
Read the full section
- Contracts: Silicon Data H100 Rental Index Futures and B200 Rental Index Futures, listed on NYMEX. They are cash-settled, and each represents one month of hourly GPU rental cost (PR Newswire).
- Shorting: selling these futures would pay off if rental prices fall. This is the basis of the WSJ "short the AI boom" angle, and it is an inference from the contract design.
- Underlying indices: Silicon Data also publishes A100 and MI300X indices, plus separate neo-cloud and hyperscaler readings for the H100. Its index page lists the neo-cloud H100 (SDH100RT) at about $2.86/hr and the hyperscaler H100 at about $7.17/hr (Silicon Data).
Technical analysis for researchers and developers
According to Silicon Data's methodology (Silicon Data), the index is built in four steps: Implementation details that matter for settlement are not public: the provider weights, the adjustment coefficients and the raw data. Outside researchers therefore cannot reproduce the index. Dave Friedman reports a weekly correlation of only 0.17 between Silicon Data's indices.
Read the full section
According to Silicon Data's methodology (Silicon Data), the index is built in four steps:
- Normalise: convert each rental record to a price per GPU per hour.
- Filter: remove errors in several stages.
- Basis-adjust: correct each record for rental type, geography, CPU platform and GPU variant, so it matches a standard benchmark contract.
- Aggregate: average prices within each provider, then take a weighted average across providers. The index is published once per business day.
Implementation details that matter for settlement are not public: the provider weights, the adjustment coefficients and the raw data. Outside researchers therefore cannot reproduce the index. Dave Friedman reports a weekly correlation of only 0.17 between Silicon Data's indices. He reads this as a sign that the indices measure different economics (Dave Friedman).
Claims and evidence
- "First daily GPU benchmarks" and coverage of about 80% of the global rental market: these are company claims (CME; Silicon Data).
- H100 index down 23%, from $3.06 (September 2024) to $2.36 (June 2025): this figure comes from Introl, a third party, which relies on Silicon Data's own numbers (Introl).
- Regulatory delay: reported by several outlets. No CFTC primary document was reviewed (BigGo; Finance Magnates).
Read the full section
- "First daily GPU benchmarks" and coverage of about 80% of the global rental market: these are company claims (CME; Silicon Data). No independent audit was found.
- H100 index down 23%, from $3.06 (September 2024) to $2.36 (June 2025): this figure comes from Introl, a third party, which relies on Silicon Data's own numbers (Introl).
- Regulatory delay: reported by several outlets. No CFTC primary document was reviewed (BigGo; Finance Magnates).
Context and prior work
- Silicon Data: founded in April 2024 and backed by the trading firm DRW. It has raised about $35.2M, including a $30.5M Series A in August 2026 (Crypto Briefing).
- Competing venues and indexes: ICE, Nodal Exchange and Architect are each promoting different GPU price indexes as settlement benchmarks. BGC opened an over-the-counter compute market in June (Finance Magnates).
Read the full section
- Silicon Data: founded in April 2024 and backed by the trading firm DRW. It has raised about $35.2M, including a $30.5M Series A in August 2026 (Crypto Briefing). It launched a GPU forward-curve product in April 2026 (SiliconANGLE).
- Competing venues and indexes: ICE, Nodal Exchange and Architect are each promoting different GPU price indexes as settlement benchmarks. BGC opened an over-the-counter compute market in June (Finance Magnates).
Limitations, safety and contested findings
- In late July, Gpuinstances listed an H100 at $0.94/hr. The trackers cover different providers and contract terms (Implicator).
- Friedman argues that index contributors who also hold futures positions could have an incentive to move reported prices (Dave Friedman).
- Rental indices leave out availability guarantees, networking and service premiums.
Read the full section
- Trackers disagree widely. In late July, Gpuinstances listed an H100 at $0.94/hr. A Silicon Data hyperscaler reading had a $7.40/hr lower bound, and GetDeploying showed a $3.25 median. The trackers cover different providers and contract terms (Implicator).
- Manipulation risk. Friedman argues that index contributors who also hold futures positions could have an incentive to move reported prices (Dave Friedman).
- Imperfect hedges. Rental indices leave out availability guarantees, networking and service premiums. Margin calls can also strain hedgers' cash flow, as in Metallgesellschaft's 1993 oil-hedging losses (Dave Friedman).
Business and practitioner implications
- Who could use it: neoclouds and GPU lenders could hedge residual-value and utilisation risk, and AI labs could lock in future compute budgets.
- What to watch: the November 9 CFTC deadline. The CFTC cannot extend again unless NYMEX consents (Dave Friedman). Early trading volume is the other signal.
- Benchmark competition: with ICE and others pushing rival indexes, whichever one wins will shape how GPU-backed debt and cloud contracts are priced.
Read the full section
- Who could use it: neoclouds and GPU lenders could hedge residual-value and utilisation risk, and AI labs could lock in future compute budgets. How well the hedge works depends on how closely their actual contracts track the index (basis risk).
- What to watch: the November 9 CFTC deadline. The CFTC cannot extend again unless NYMEX consents (Dave Friedman). Early trading volume is the other signal.
- Benchmark competition: with ICE and others pushing rival indexes, whichever one wins will shape how GPU-backed debt and cloud contracts are priced.
The source trail.
Sources (12)
Want to Short the AI Boom? She Wants to Give You a Way to Bet on ‘Compute’ Prices
Feed excerpt only.
wsj.com