Compute Derivatives at the CFTC

• The DCM and Commission "should be able to reconstruct the lineage of each settlement value, identify the methodology version that governed it, and determine what market population the value actually represents."
• "Define the underlier as a bounded economic identity: contract terms should specify the dimensions that materially determine equivalence rather than treating ‘compute’ as a unitary commodity."

Signed by Stephen Weber, founder, author of the Palimpsestus Open Spec. The letter "does not recommend prohibiting compute derivatives" and "does not recommend approval or disapproval of NYMEX Submission No. 26-370," the CME H100 and B200 filing it reviews.

• "Require DCMs seeking to list compute derivatives to demonstrate that settlement indices rely on verified, arm's-length execution data rather than unverified supplier rate cards or self-reported surveys."
• Quality adjustments "based on realized workload throughputs specifically actual execution yield (Tokens/Sec per watt) and live interconnect efficiency ratios (NVLink/PCIe CRC error penalties)," instead of "theoretical nameplate chip ratings or uncalibrated energy proxies."

Signed by Alfred G. Francis III. The letter argues that settlement should move to "deterministic, hardware-level execution metrics (‘bare-metal telemetry’)" and proposes a "Deliverable Compute Supply (DCS) framework" that deducts encumbered capacity.

• "A cost-to-move test. As a documented stress test, the DCM shows what it would cost to move the settlement price through each index input and compares that cost with the gain available to the largest aggregated position."
• "Only arm's-length, net and complete data. Deals under common control or with circular payments are excluded, other related-party deals are disclosed and reviewed, reported prices are net of all credits and rebates, and reporters submit every covered deal."

Filed in a personal capacity, from Switzerland. The letter addresses indexes "built from executed transactions, invoices or billing records" and discloses no position in compute derivatives and no commercial relationship with any compute exchange, index provider or capacity provider.

• "We respectfully suggest that any settlement reference publish, for each input, at least the offer type, the exact hardware variant, the minimum quantity, the billing basis, the region and the time of observation."
• On Question 2(b), protections "that are cheap to specify: count each provider once; use a median or a trimmed statistic rather than a minimum; require evidence that capacity was available at the posted price; and publish every constituent observation with its timestamp, so that a change to a posted rate during an observation window is visible to everyone at the time it happens."

FastGPU is a public price-comparison service for rented GPU compute. It has collected posted prices since July 31, 2026 and releases them under CC BY 4.0. It discloses that it joins provider referral programs, sells paid data access and offers a service that starts GPU instances at listed providers.

• "Treat short-tenor principal forwards as their own category. Do not fold them into provider-administered reservations."
• "Require two-oracle settlement."

The comment is unsigned. It addresses short-tenor GPU forwards in which the seller "stands as principal, takes the buyer's price risk, and must deliver the capacity," and cites the extended review of the CME H100 and B200 compute futures filing.

• Evidence supporting a compute benchmark "should be evaluated not only at initial listing, but through a defined process for material-change revalidation."
• "Where practical, benchmark governance should distinguish the evidence state of contributing observations and identify material dependencies on data that cannot be independently observed or verified."

Signed by Neil P. Osnato. Describes the firm as "an independent advisory and analytics firm focused on Infrastructure Assumption Verification and Decision Assurance" and states that it "takes no position on whether any particular compute derivatives contract should be approved or prohibited." Its first comment is 0013, below.

• "Publish every input of every settlement value with its source, read time and value, so the contract market, the Commission and the public can recompute it."
• "Never blend posted rates with marketplace prices in one series"

GPU Economy publishes posted GPU rental rates read hourly from 71 providers and a Cloud GPU Price Index, the median posted on-demand H100 SXM rate. It discloses a settlement license offered since October 1, 2026, which no one has signed and which has no fee set, referral programs with five providers (three of which post rates in the index), and possible fees on reserved-capacity requests it passes on. It holds no positions in compute or compute derivatives.

• "Recognize an energy-normalized unit of qualifying compute delivery as a viable denomination for a broad compute benchmark."
• "Permit completed bilateral sales to contribute alongside venue executions when reporting obligations, verification controls and supervisory access are adequate."

NativX administers COIL, a benchmark priced in U.S. dollars per Normalized Compute Unit, and operates the constituent-token spot market that feeds it. It states that ICE and NativX have announced plans for COIL-based futures, and that it has recently launched with no executed volume to report.

• The fungibility question "is better posed as “at what level is the unit defined, and who defines it” than as “what grade adjustment is required”."
• On Question 2(g), the governance a compute index would need, including "an administrator that is independent of the contributors and does not itself supply compute capacity" and "a stated hierarchy of inputs that prefers observed transaction data over posted or administered rates, and that discloses the proportion of each in every published value."

An independent legal and tax adviser, a Registered Foreign Lawyer in the UK and qualified in Portugal and Brazil. States that she holds no position, acts for no one and offers no data. Offers electricity markets as a comparator and the EU Benchmarks Regulation (2016/1011) as a comparative instrument, and recommends neither.

• "The Commission should therefore not condition listing on the maturation or standardization of the compute cash market."
• The Commission "should confirm that an index computed from price inputs submitted directly by multiple independent contributors through a published, outlier-resistant aggregation methodology can satisfy Appendix C's requirements for a “reliable, acceptable, publicly available and timely” settlement price."

Douro Labs is "the core contributor to the Pyth Network" and operates Pyth Pro, a median-based reference-price service built from first-party publisher contributions. States that it "does not operate a compute venue, does not sell compute capacity and does not trade compute," and that it "is prepared to support a compute reference index on this basis." Cites the CME and Silicon Data H100 and B200 rental-index futures filing and the ICE and Ornn announcement, and takes "no position on any compute index or contract now before the Commission."

• Central recommendation: "verifiability before fungibility."
• "Begin with narrowly defined, fixed-date contracts and clearly specified hardware or workload profiles; broaden only when transaction data and conversion evidence support it."

States a commercial interest: it is developing a proposed "Digital Compute Unit (DCU) framework" and "intends to provide commercial services related to DCU." Ranks evidence in five tiers, with posted or advertised rates as "context only unless the administrator demonstrates that rates are actionable and representative." Lists Ornn’s OCPI (ICE), the Silicon Data rental-index futures filed by NYMEX, ICE with NATIVX, and Nodal Exchange with Compute Desk, and endorses none.

• Recognize that "a compute future listed in isolation from a documented, reportable OTC book will not produce the robust market the Chairman described."
• "Evaluate DCM filings against whether the settlement series is a transaction print with governance the Commission can inspect, not a posted offer the supplier administers."

Argues that listed futures "inherit liquidity from the OTC market they sit next to. They do not invent it." Its briefing names the underliers: the Silicon Data H100 and B200 rental indexes for the CME contracts and Ornn’s OCPI for ICE, and says an OTC confirmation names one reference price, "never an average."

• "Advance DCM/SEF Listings for Cash-Settled Rolling GPU Benchmarks: Provide clear guidance affirming that transparent, volume-weighted median benchmark methodologies meet CEA Core Principle 3 standards."
• "Formally Recognize Compute Basis Mechanics: Endorse exchange contract architectures that accommodate hardware, reliability, and regional variations through standardized basis differentials."

Calls compute "a $500B+ bilateral OTC market" already. Proposes settlement indexes "derived exclusively from programmatic, auditable API billing records and cleared bilateral trade prints, explicitly excluding non-binding, advertised list prices," with no provider above 20% of index weight, and tenors out to 12 months.

• Takes no position on any contract: it "does not recommend approval or rejection of a particular compute derivative, index, provider, venue, or perpetual design."
• Keep source, provenance, benchmark-input eligibility, validation, execution and settlement separately attributable: "Observable information is not automatically qualified settlement information."

An information-governance comment. It states that it holds no compute-market data and answers each question within that limit.

• The summary records only the topic: "Discussed observations regarding compute markets generally, including with respect to surrounding pricing methodologies."

Meeting of September 24, 2026, between CFTC staff and OneChronos, with its policy advisers Delta Strategy Group. An August remark from OneChronos appears in the statements section below.

• "Distinguish representation from evidenced availability": represented capacity should not count as deliverable without evidence of its hardware, location, energization, reservation status and dependencies.
• Define "a defined state" of the underlier as well as the commodity itself.

Warns that mixing on-demand, reserved and nominal capacity "may produce a price series without establishing economic equivalence."

• Publish compute-specific guidance under Appendix C "setting out what a compute settlement reference price must demonstrate", before first-of-kind contracts are listed.
• "Provider-administered posted rates are not admissible as a primary settlement input, and where used at all are weight-capped and disclosed, with executed-transaction and live-market-quote primacy in the input hierarchy."

An index provider; discloses that it "would stand to benefit if the Commission established criteria for compute settlement reference prices". Its central claim is that "the sellers set the settlement price": posted rates are "a publication, not a transaction" and the firms that administer them are the natural short side. Answers Q2(g) "No, not today, and not from any provider, ourselves included."

• A central reporting authority for bilateral cash-market transactions, with "current and historical price data, publicly available"; "the independence of the reporting authority from the entities that it measures will be key, going forward".
• "Transparency and oversight of Compute Futures index providers, their index calculation methodologies and transparency of settlement pricing."

Concludes that "the present underlying market is not ready to support listed derivatives", on concentration grounds: ten operators holding two thirds of one GPU generation, cited to a third-party report. Describes the announced exchange products and their index administrator. Carries a no-AI disclosure.

• The Commission should commit, "in a follow-up notice, to heightened review of self-certified compute filings until the open questions in Section II.2 are answered".
• "Condition self-certification eligibility on public disclosure of the number and market share of rate-contributing providers feeding any settlement index."

States: "None of this is a case for leaving compute derivatives unlisted indefinitely." The concern it names is sequencing.

• Define "a precisely and narrowly defined hardware generation, configuration, and location as the reference underlier", with tenors short enough to avoid a hardware-generation transition or a contract that "transparently rolls and re-defines the underlying".
• "Require both minimum contributor counts and a cap on a single contributor’s share of index volume", and "ensure that quotes during observation windows are genuinely executable."

Argues that two features of compute are durable: the hardware "depreciates on a functional basis on a roughly annual cycle", and a compute-hour "is not a homogeneous unit". Treats list-price markets and bilateral reserved capacity "almost as two different commodities". Sees perpetual futures as better suited to hardware turnover, with the same reference-price standards as a prerequisite.

• Contract specifications and risk disclosures should distinguish the reference-price exposure covered, the differences between the index and common commercial compute contracts, that "cash settlement does not ensure access to physical capacity", and that "an effective price hedge does not ensure adequate project cash flow or debt-service capacity".
• Risk materials for commercial hedgers should explain "the potential timing mismatch between derivative margin obligations and cash flows from the underlying compute business".

Supports listing. The argument is that a hedge transfers price risk only: "the hedgeability of compute should not be equated with the bankability of the underlying project." No data, no index named. Carries an AI-assistance disclosure.

• Do not treat cash-market opacity "as an inherent property of compute. It is a property of a settlement architecture."
• "A contract should not settle to a price the Commission cannot observe or verify"; this "counsels a sourcing requirement, not abstention from listing".

Data is a 17-day census of machine-priced inference endpoints paid over the x402 protocol. The filer states the census covers the API layer, "not evidence about hourly accelerator pricing", and gives its own scale as 226 settlements for $3.12.

• Permit and facilitate compute derivatives on regulated U.S. venues; "encourage competition among venues and methodologies".
• Do not delay "until a perfect benchmark exists".

Their second letter. An open letter posted on LinkedIn on August 20, before the docket opened, asked for "an independent reference layer that can sit beneath competing venues". The filed letter replaces that with competition among methodologies.

• "Clear provisioning on the index calculation methodology, transparency with respect to the underlying market data collection process and source."
• "Clear separation of index providers business functions and the interests of its investors, of whom many are trading firms."

• Surveillance should distinguish "when a derivatives price is primarily measuring scarcity from when that financial signal is also materially participating in the production, allocation, or reduction of scarcity".
• Observe concentration across "physical capacity + benchmark observations + reservation behavior" together rather than within each layer alone.

States that it takes no position for or against listing, and that "large-language-model assistance was used to organize and draft this comment".

• The Commission "should refrain from the listing of compute derivatives until a standardized, independently governed benchmark is in place and mandatory, verifiable reporting from power-dependent compute providers is established".
• An index "must adjust for the operational factors that determine compute’s economic value; simple posted-rate averaging is structurally inadequate": hardware age and efficiency, data egress costs, regional power constraints.

Notes that "compute", "compute capacity" and "compute services" are used interchangeably in the RFC and asks the Commission to say which it means.