Stoa’s CFTC Comment on Compute Derivatives

October 22, 2026
VIA ELECTRONIC MAIL

Mr. Christopher Kirkpatrick
Secretary of the Commission
Commodity Futures Trading Commission
Three Lafayette Centre
1155 21st Street NW
Washington, DC 20581

Re: Request for Comment on the Listing of Compute Derivatives Contracts
RIN 3038-AF77, 91 FR 54259 (Aug. 21, 2026), Docket No. CFTC-2026-1850

Dear Mr. Kirkpatrick:

Stoa Markets, Inc. (“Stoa”) submits this letter in response to the Commission’s Request for Comment on the Listing of Compute Derivatives Contracts (the “RFC”).¹ We support the Commission’s decision to examine these questions publicly, and we share the view, stated in the AI Action Plan, that the national interest is served by “improving the financial market for compute.”² The comment period closed on October 20, 2026. We respectfully request that the Commission place this letter in the comment file and consider it at its discretion.³

The compute futures submitted to the Commission, and those announced, reference the price of renting compute.⁴ Several commenters present the hedging of GPU-backed loans and GPU fleets as a principal use of such contracts. The asset behind both is the hardware itself, and its price is set in the secondary market for servers, not in the rental market. Whether a rental contract hedges it depends on how hardware prices move with rents, and we have not identified in the comment file a measurement of that relationship based on observed hardware prices. Stoa operates a marketplace for that hardware. This letter explains what hardware prices can and cannot show, and offers aggregated price data to staff.

I. Stoa’s interest

Stoa operates an institutional request-for-quote marketplace on which GPU servers and racks are quoted and sold for physical delivery. Stoa’s founders have traded interest rate derivatives and built trading and pricing infrastructure for energy commodities.

Stoa’s interest is commercial. Its business includes operating the marketplace described above and developing hardware price data, valuation analytics and price benchmarks. Stoa does not operate a designated contract market (“DCM”), a swap execution facility or a derivatives clearing organization, does not trade derivatives, and is not registered with the Commission in any capacity.

We do not ask the Commission to adopt new rules, to delay the contracts now under review, or to prefer any index.

II. Summary of positions

1. The GPU-hour is a service, consumed as it is produced, that cannot be stored. The accelerator, with the server and network around it, is a durable good: it can be stored, carries title, trades in a secondary market, and is graded by condition. The contracts proposed to date reference the compute hour. Owners and lessors hold the device, and lenders lend against it.
2. A contract that settles to a rental index does not reference the value of the asset that earns the rent. Several commenters present rental contracts as a way to value or hedge GPU collateral, and others doubt that they can. Observed hardware prices, matched to the rental indices, would provide the most direct test of that use, and we have not identified such a test in the comment file.
3. The hardware market can inform several of the RFC’s cash-market questions. Dealers already price differences in condition, configuration and warranty; hardware sales leave records that can be reconciled with the goods; and a firm quote in a request-for-quote process commits its maker to a stated price and quantity while it stands. We describe the limitations of each below.
4. The terms and risk disclosures of a rental-settled contract should state that it references the price of compute hours, not the value of the hardware (Question 3(c)). Where a submission presents a rental contract as a hedge of GPU collateral, the Commission should look for evidence of how the reference index has moved with hardware prices, or a statement that none exists.
5. We agree with the RFC that compute is a commodity, and with commenters who would review contracts on it under the existing framework: the Commodity Exchange Act (the “Act”) and its DCM Core Principles, Part 38 with its Appendix C guidance, and Part 40.⁵
6. The accelerator is also a commodity under the Act. A contract referencing its price, whether cash-settled to a hardware benchmark or settled by delivery of graded units, would raise questions already answered by the existing framework (Question 2(i)).
7. We do not object to perpetual contracts in principle. Their funding payments reference the index at every interval, not only at expiry, so the quality of the index matters throughout the life of a position. We take no position on retail access.

III. The hour and the device

The source cited in the RFC describes compute as an installed stock of hardware and the flow of service it produces.⁶ The contracts proposed so far reference the flow, while owners, lessors and lenders are exposed to the stock. The value of the stock is the rent it will earn plus what it will fetch when its contracts end, and that second part is set in the secondary market for servers. Mature markets keep the two prices apart. In shipping, cleared freight futures settle on the Baltic Exchange’s time-charter assessments, while the Exchange separately publishes broker-panel assessments of vessel values.⁷

Several commenters present rental contracts as a way to value or hedge that stock, in whole or in part. Architect treats “the forward value of compute” and “the residual value of the GPU” as separate forecasts. It observes that “[a] lender that takes GPUs as collateral has a long exposure to the residual value of those GPUs,” and says lenders can sell compute futures to reduce the volatility of their portfolios and owners of prior-generation GPUs can “hedge the value of their fleets.”⁸ Polymarket US describes lenders and lessors with residual-value exposure as facing “the same risk as operators.”⁹ DRW says it sees several of the largest lenders slowing new lending because they “have exhausted their internal risk capacity and the exposure is currently difficult to hedge,” and looks to a liquid, cash-settled compute derivative to let lenders “lay off compute price risk.”¹⁰ Other commenters doubt that rental contracts reach the value of the stock. DNA Compute observes that “a 36-month strip reaches the rental part of it and not the rest”; Compute Capital, that “[r]esidual is not rental”; and one commenter’s prototype, built on synthetic loans and resale prices and on observed rents, finds that forward rents explain about 30 percent of value variance on operating loans and about 4 percent on a loan still under construction.¹¹

The question is empirical, and observed hardware prices are the direct way to answer it. A series of server prices matched to the rental indices by GPU model, form factor and date would show how much of a fleet’s value moves with rents. Professors Bandi and Su, whose work the RFC cites, describe a theoretical tail risk in which lenders mark down infrastructure collateral when compute futures fall, and forced sales then depress asset values further.¹² That risk is harder to assess if the only forward price lenders can observe is a rental price.

An accelerator is a durable good. A Blackwell-generation eight-GPU system ships on a pallet and carries a serial number on every GPU. It can be held in a warehouse, and while it is held there its owner pays to finance it, bears its depreciation and uses up its warranty. Commenters who say compute has no inventory and no cost of carry are right about the hour; as Polymarket US puts it, “A GPU-hour not sold is gone.”¹³ The device does have a carry, and the rent is part of it: an owner who holds a server idle gives up the rent it could have earned, as a holder of metal gives up the lease rate. That is the link between the two prices.

The link provides a practical check. A rental rate set against the price of the hardware that earns it implies a gross yield. Posted rates can overstate what owners realize, so the level of that yield is only a rough guide. Its changes are more useful. When a rental index and matched hardware prices diverge, the cause may be a change in financing costs, utilization or the generation outlook, or a problem in the index’s inputs, and the divergence tells a reviewer where to look.

Hardware prices should also mark generation transitions early. NVIDIA now brings new data-center architectures to market on a one-year product cadence,¹⁴ while operators depreciate the equipment over as long as six years,¹⁵ so the value of a unit at the end of a lease depends on how many generations have arrived in between. We expect a new generation to be reflected in hardware prices before it is reflected in rental rates, because a hardware price reflects the remaining life of the asset and an on-demand rental rate reflects current use. If that holds, hardware prices can identify the timing of a transition, which bears on the retirement and fallback terms commenters propose and on DRW’s recommendation that perpetual contracts be evaluated after a reference index has a record through at least one chip-generation transition.¹⁶

Financial Times reporting summarized in the comment file describes discussions of insurance that would pay lenders the shortfall if seized GPUs resell for less than the debt.¹⁷ Reported loan-to-value ratios run from about 60 to 70 percent on GPU-backed facilities to 70 to 80 percent where a backstop supports the loan.¹⁸ A rental-settled derivative does not reference resale value; a reliable hardware reference price would give those positions an observable mark.

IV. The hardware cash market (Question 1(e))

The RFC invites comment on all aspects of the compute markets.¹⁹ We address the characteristics listed in Question 1(e) in order and apply them to the device.

(i) Storability. Hardware is held in inventory by OEMs, distributors, dealers, lessors and operators, and it is resold. A capacity reservation is a contract right whose transfer depends on its terms, while a server held outright can generally be resold, subject to export controls, any end-user undertakings given when it was bought, and any lender’s lien. Storage does not preserve value: the hardware ages technologically while it waits.

(ii) Public transaction record. There is no comprehensive public transaction record for AI hardware. A hardware sale typically generates, beyond the invoice, a packing list with serial numbers, shipping documents and evidence of the transfer of title. These records are private, but they can be reconciled with one another and with the goods. A hardware price can therefore be traced to documents produced in the ordinary course of business, with no new reporting regime. The records have limitations. A serial number identifies a device, not its title, liens, location or export status.

(iii) Producer concentration. Production of the accelerators in question is concentrated, as the comment file notes. Ownership of installed hardware is broad, spanning cloud providers, enterprises, research institutions and funds. Breadth of ownership does not by itself make a price sample representative. Contributor concentration in any hardware series used for settlement should be measured and disclosed, as commenters propose for rental indices.²⁰

(iv) Fungibility. Hardware is heterogeneous in ways that are specified and priced daily. The dimensions are the SKU and memory configuration; the form factor (SXM, PCIe, NVL); the system configuration (an eight-GPU node or a 72-GPU rack); the condition grade; warranty remaining and who backs it; and provenance, including the unit’s export history and any license conditions on where it may run. Dealers routinely quote differentials across these dimensions, although grading conventions differ from dealer to dealer. Grade adjustments of the kind the RFC asks about are market practice for the device; for the hour, no uniform quality grade yet exists across providers.²¹

(v) Price reporting. Several data providers sell or publish hardware prices; their inputs vary and include dealer listings, firm quotes and reported sales. We are not aware of a hardware price series used to settle a derivative, or of a price reporting agency applying an enforced standardized methodology to hardware, as agencies do in energy and metals. Price discovery for hardware occurs in dealer quotes, in competitive requests for quote, and in bilateral sales.

The closest analogues for the device are the physical metals markets, where warrants make graded lots deliverable,²² and markets in long-lived capital equipment such as aircraft, where appraisers distinguish market value, base value and residual value and lenders rely on those appraisals.²³ Those markets rest on standard specifications, inspection and records that tie a price to an identified asset, and those foundations support financing at scale and, in metals, exchange-traded hedging. The AI hardware market is at an early stage of establishing these foundations.

V. Reference prices, quotes and supply

Deliverable supply (Questions 1(g) and 2(h)). Commenters propose estimating deliverable supply of GPU-hours from installed accelerators of the specified type within the index’s scope, net of capacity that owners use themselves or have committed under contract.²⁴ The hardware market observes part of the installed base, as shipments, secondary sales and lease transfers move units between owners, and its records can help test claims about installed counts and identify what one commenter describes as capacity “sold, pledged or counted twice.”²⁵ Those records must be reconciled with ownership, operating status and existing commitments, and de-duplicating by hardware alone can erase legitimate resale of service rights.²⁶ Appendix C already provides the test for committed capacity. Paragraph (b)(1)(i)(A) excludes supply committed under long-term agreements unless the designated contract market demonstrates that it is “consistently and regularly made available to the spot market for shorts to acquire at prevailing economic values,” and for cash-settled contracts paragraph (c)(2) calls for assessment of the underlying cash market’s size and liquidity following the analysis in paragraph (b)(1).²⁷

The input hierarchy (Questions 2(b), 2(c) and 2(g)). FIA asks the Commission to encourage methodologies that distinguish executed transactions, executable bids and offers, and indicative posted rates, and to explain the hierarchy and validation of those inputs;²⁸ Professors Bandi and Su say that where transaction liquidity is insufficient, “firm, binding quotes from a contributor panel” could supplement the inputs.²⁹ For hardware, an accepted quote is an executed transaction, and one that has completed delivery and acceptance is the most complete evidence. A firm quote in a request-for-quote process commits its maker to a stated price and quantity while it stands, and is executable by the party that requested it. A dealer’s published price list is indicative, and a rental provider’s rate card can be executed only when capacity is available at that price, which the card does not show.

Hardware quotes have their own limitations. Quotes for new systems often carry estimated delivery dates subject to manufacturer allocation, and quotes for configured systems must be adjusted for configuration, condition and warranty before they can be compared. A methodology for either underlier should give firm, time-bound commitments priority over posted prices, treat comparable and recent executed transactions as the strongest evidence, and disclose with each determination the indicative share of each type of input on which it rests.³⁰ It should also address sparse data and contributor concentration. IOSCO recognizes that in a low-liquidity market “a confirmed bid or offer might carry more meaning than an outlier transaction,” provided the benchmark remains anchored in an active market and the methodology sets out the hierarchy of inputs.³¹

Intermediary conduct (Question 2(f)). The same hardware is often offered through several intermediaries at once, so that a single lot can appear as several offers at different prices. That is not inherently improper, but a measure of supply or market depth that counts listings or quotes without tracing them to the underlying units would overstate both. The same caution applies to rental capacity resold through more than one platform.³²

VI. Customer protection

Question 3(c). The idiosyncratic risk most relevant to hardware owners and lenders is that the value of the hardware can move independently of the rental rate. The terms of a rental-settled contract, and the risk disclosures that accompany it, should state that the contract references the price of compute hours and not the value of the hardware, and should name residual-value risk as distinct from rental-rate risk. Other commenters support disclosure of this kind. DNA Compute writes that “[l]isting a contract is not a finding that its reference is fit to value collateral or price a loan, and a customer should be told so,” and Li Li would have exchanges publish “the historical relationship between the futures reference price and major categories of commercial compute contracts.”³³ Where a submission presents a rental contract as a hedge of GPU collateral, fleet value or residual value, the relationship that matters is the one between the reference index and hardware prices for the same GPU class.

Question 3(b). Export status is a hardware attribute with a price. United States export controls govern where advanced accelerators may be shipped and to whom they may be transferred,³⁴ and they narrow the set of buyers to whom a given unit may be resold. Where a contract or its reference index specifies geography, it should also specify whether the capacity behind the rate is subject to export-license conditions or end-user restrictions on who may use it. Capacity subject to such conditions and capacity free of them are different goods and can price differently.

VII. The hardware consideration (Question 2(i))

The Commission asked in Question 2(i) for other considerations bearing on a DCM’s compliance with the Core Principles in listing and trading compute derivatives. We submit one such consideration. The accelerator is itself a commodity in the ordinary sense: it can be stored, delivered, graded and financed, and it trades in a secondary market whose records can be reconciled with the goods. The Act’s definition of commodity reaches “all other goods and articles,” and the RFC leaves room for such an underlier when it states that a compute derivative may reference “a different type of compute-related commodity.”³⁵ A contract referencing the price of accelerators, whether cash-settled to a hardware benchmark or settled by delivery of graded units, would raise questions of a kind the Commission has addressed before in the physical metals markets, and none of them requires new rules.

VIII. Requested Commission action

Stoa respectfully requests that the Commission, in reviewing compute contract submissions under Part 40:

1. Look for the terms of rental-settled contracts, and any risk disclosure the exchange publishes with them, to state that the contract references the price of compute hours and not the value of the hardware, and to identify residual-value risk as distinct from rental-rate risk.
2. Where a submission presents a rental-settled contract as a hedge of GPU collateral, fleet value or residual value, look for evidence of how the reference index has moved with hardware prices for the same GPU class, or a statement that no such evidence exists.
3. Consider accelerator hardware as a distinct compute-related commodity and assess any submitted contract referencing its price under the existing Core Principles and Parts 38 and 40, with compliance evaluated against that contract’s terms, underlying market and settlement mechanism.

Stoa would welcome the opportunity to meet with staff of the Division of Market Oversight and to provide aggregated, de-identified hardware price data. Before submitting any confidential information, Stoa would contact the staff identified in the RFC and request confidential treatment under the procedures in 17 CFR 145.9.³⁶ We appreciate the Commission’s consideration of these comments.

Respectfully submitted,

Berat Celik
Chief Executive Officer
Stoa Markets, Inc.
156 2nd Street, Suite 509
San Francisco, CA 94105