The Commodity Futures Trading Commission issued a Request for Comment on the Listing of Compute Derivatives Contracts on Wednesday, August 19, and it reached the Federal Register on Friday, August 21 at 91 FR 54259 to 54264 as document 2026-17163 under RIN 3038-AF77. Most of the six pages read the way a market-structure consultation normally reads, working through fungibility, settlement indices and manipulation risk. Then Section II.4 arrives under a heading that reads "Perpetual Compute Futures," and asks the public two questions about a contract type that no US designated contract market lists.
Comments close on Tuesday, October 20, 2026, which happens to be the same day the SEC's own crypto securities offering rule closes its file. The instrument in that heading is familiar to anybody who has traded a crypto derivative. The underlier is the part that is new, and it is the part worth understanding before the comment file fills up with submissions from people who rent GPUs for a living.
What a Compute Derivative Actually References
Compute is processing power, the thing large language models consume in enormous quantities, and the Commission defines the likely underlier tightly rather than leaving it abstract. The document states that the commodity underlying a compute futures contract would typically be "access to rented compute capacity from hardware the purchaser of such capacity does not own," and it gives the hourly rental price of an Nvidia B200 as the worked example. It leaves the door open to a different unit, naming access to a stated volume of LLM inference tokens as an alternative.
Rented is the operative word. A compute contract does not reference a chip, a data center, or the equity of the company that owns either one. It references an hour of somebody else's chip, quoted in dollars per GPU-hour, in the same way a power contract references a megawatt-hour rather than a turbine.
A benchmark for that already exists in public. Silicon Data publishes a daily H100 Rental Price Index under the ticker SDH100RT, built from observations across neo-cloud providers, hyperscalers, colocation markets and private rental platforms, standardized for machine specification, rental term and geography before publication. That index read $2.53 per GPU-hour when I pulled the page at 07:50 UTC on Monday, August 24.
The reason a market wants any of this is the same reason airlines hedge jet fuel. Compute is the largest single input cost sitting between an AI agent or a frontier model and the balance sheet paying for it, and the operators renting that capacity out have been building capex plans around demand nobody can forecast cleanly. The CFTC leans on the July 2025 AI Action Plan for the policy framing, one of whose recommended actions is to improve the financial market for compute so startups and academics can get access to it.
Why GPU Time Behaves Like a Commodity and Where It Does Not
An hour of H100 capacity is perishable in a way almost no listed commodity is. Nobody stores it, nobody carries it into next quarter, and an hour that goes unrented is not inventory but a loss. The Commission reaches for exactly that precedent in a footnote, citing its own 2012 electricity work for the proposition that "electric energy must be used or consumed on an as-needed basis" rather than being bought ahead, delivered and stored.
That is the case for treating compute as a commodity. The Commission then makes the case against it with more force than most coverage has registered, writing that compute "may not yet exhibit certain of the characteristics of commodities that typically underlie a commodity derivatives market, including fungibility, standardization, and sufficient liquidity." It describes compute markets as fragmented, with price formation happening mostly in opaque bilateral transactions, and it warns that dominant participants may hold enough pricing power to make an index manipulable.
Reading the question list as a checklist rather than a wish list shows where the real work sits.
|
Test a listed contract has to pass
|
Where compute stands in the CFTC's own reading
|
|
Units fungible across producers
|
Weak. Price varies by provider, region and contract structure
|
|
Standardized specifications and grades
|
Missing. The Commission expects standardization would have to be built
|
|
Publicly observable transaction record
|
Thin. Its preliminary view is that undisclosed bilateral agreements carry the majority of economic value
|
|
Deliverable supply that can be estimated
|
Open. Commenters are asked to supply a methodology under Core Principle 5
|
|
Settlement price hard to push around
|
Open. Providers administer many of the posted rates that feed published price series
|
The manipulation question is the sharpest thing in the document. If a compute index is calculated partly from rates that the capacity providers themselves post, then the parties best positioned to move the settlement price are the same parties supplying the capacity. The Commission asks directly what would stop a provider from adjusting a posted rate or steering capacity onto and away from a venue during an observation window, and that is not a rhetorical question.
The Section That Makes This a Crypto Story
Section II.4 runs to two lettered questions and roughly seventy words, and it is the reason this document matters outside the AI infrastructure trade. The first asks if perpetual compute futures would carry advantages for market participants over "traditional" or "fixed date" futures contracts, and if they would provide commercial risk management features that cannot be met with existing products. The second asks if perpetual compute derivatives would pose any unique risks for market participants or the broader markets, and what additional protections or safeguards the Commission or exchanges should adopt.
A perpetual futures contract has no expiry and no delivery date, and it holds its price near the underlying index through a periodic payment between longs and shorts rather than through convergence at settlement. That payment is the funding rate, and it does the job that expiry does in a dated contract.
Now put that structure against an underlier with no natural delivery date at all. Rented compute is consumed continuously, priced by the hour, and never delivered into a warehouse, so the calendar month a fixed-date contract settles into is an arbitrary boundary rather than a real economic event. A hedger who wants to be short compute cost for three years has to roll a dated contract thirty-six times and eat the basis on every roll. A perpetual removes the roll entirely, and for a genuinely non-storable commodity that is a structural argument rather than a stylistic preference.
The risks travel with it. Funding on a thin market can dominate the economics of a position, an index built from provider-posted rates is a weaker anchor than a deliverable commodity, and a contract with no expiry never forces the reconciliation with physical reality that delivery imposes. Those are the objections the Commission is inviting the industry to answer in writing.
Chairman Michael S. Selig, who has run the agency since being sworn in as its sixteenth chairman on December 22, 2025, framed the exercise in the accompanying CFTC release. "Just as American markets helped establish the gold standard for trading the commodities that powered the industrial economy, we will do the same for the commodity that will power the intelligence economy," he said, calling the request a first step toward clear rules of the road for American compute markets. The voting summary attached to the document records that he voted in the affirmative and that no Commissioner voted against.
And one hard limit governs all of it. A request for comment is a consultation, not a rule, not an approval and not a product. Nothing in the primary text authorizes anybody to list a perpetual compute contract, and October 20 is a deadline for submitting opinions rather than a date on which anything begins trading.
What CME and Silicon Data Have Already Filed For
The regulatory question did not arrive out of nowhere. CME Group and Silicon Data announced on August 11 that they intend to launch two contracts, Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures, targeting October 5 and listed under NYMEX rules, pending regulatory review. Each contract is sized to roughly a month of rent for a single GPU and settles against the Silicon Data indices rather than against any physical delivery of capacity.
Two details are worth holding onto. The first is that the October 5 target is explicitly conditioned on regulatory review, so it is an intention rather than a listing, and the plumbing between an announcement and a live contract is where these timelines usually slip. The second is that both products are fixed-date monthly futures, which is precisely the structure Section II.4 asks the industry to compare a perpetual against.
The sequencing tells you something. Dated products came first and the perpetual question came second, which inverts how the crypto market developed, where perpetuals took the volume long before dated contracts mattered to anybody. Whoever operates the GPU fleets that our data center coverage has been tracking is about to be asked, on the record, which of the two shapes actually hedges their book.
Six Open CFTC Files and Why They Keep Getting Mixed Up
Conflating these is the easiest mistake available on this subject, and plenty of coverage has already made it. The Commission has six separate comment windows open, and they are separate documents with separate scopes and separate deadlines.
|
CFTC comment file
|
Closing date
|
|
Extension of standard futures contracts to 24/7 trading, and perpetual contracts referencing physically delivered or storable energy commodities
|
Wednesday, August 26, 2026
|
|
Joint SEC and CFTC file on portfolio margining
|
Monday, August 31, 2026
|
|
Review of mandatory swap clearing determinations
|
Monday, September 28, 2026
|
|
Duplicative regulation of commodity pool operators and trading advisors
|
Monday, October 5, 2026
|
|
Conflicts of interest and affiliations
|
Monday, October 5, 2026
|
|
Listing of compute derivatives contracts
|
Tuesday, October 20, 2026
|
The first row causes most of the trouble, because two of these files contain the word perpetual. The 24/7 trading filepublished on July 28 asks about round-the-clock trading of standard futures generally, but its perpetual half is scoped only to physically delivered or storable energy commodities. Compute is neither physically delivered nor storable, which is why it needed a document of its own.
The October 20 date also carries two crypto-adjacent files rather than one. The SEC's offering rule at 91 FR 54510 closes the same day, so anybody drafting submissions on both is working to a single calendar.
Frequently Asked Questions
Is a compute futures contract a crypto product?
No. It settles against the rental price of GPU capacity and has nothing to do with any token or blockchain. The overlap with crypto is structural rather than sectoral, because the perpetual contract design the CFTC has put out for comment was developed and stress-tested almost entirely in crypto derivatives markets.
Can US traders buy a perpetual compute contract?
No, and no US designated contract market has one pending. The Commission is at the stage of asking what such a product would need before anybody proposes one, and the two compute products with an announced target date, both from CME and Silicon Data, are fixed-date monthly futures rather than perpetuals.
What would a compute perpetual actually hedge?
Exposure to the hourly cost of renting AI capacity, held open indefinitely. A cloud provider with capacity to sell would use it to lock in revenue and a model developer buying capacity would use it to cap an input cost, with neither side forced to roll a position every month or absorb the basis cost that rolling creates.
Does the October 20 deadline mean something happens on October 20?
Only that the Commission stops accepting submissions. No statutory clock forces an outcome after a request for comment, and the agency can follow one with a proposed rule, with guidance, or with nothing at all.
Bottom Line
The instrument is not the story. A cash-settled Bitcoin perpetual already cleared a US designated contract market, so the contract type itself is settled ground and the argument about it has been had. What changed on August 19 is the underlier, and a federal regulator putting the words perpetual and compute in the same heading is the first formal sign that the design has applications well outside crypto.
Two dates are worth marking. October 5 is the CME and Silicon Data target for the H100 and B200 contracts, conditioned on regulatory review, and a slip there tells you the review is harder than the announcement suggested. October 20 closes the comment file and the submissions become public, which means the argument over what a compute perpetual should look like will be readable in full rather than inferred from press releases.
The useful thing to watch between those two dates is who bothers to file. A comment file dominated by exchanges and data vendors means the product is a market-structure project. A file with hyperscalers, neo-clouds and model labs in it means the demand is real, and demand is what turns a request for comment into a listing.
This article is for informational purposes only and does not constitute financial or investment advice. Cryptocurrency trading involves substantial risk. Always conduct your own research before making trading decisions.
