Five Indices, One Price

It is easy to mistake a benchmark for a calculation. The calculation is the least of it. What makes a series referenceable is a set of institutional properties that have nothing to do with arithmetic: a disclosed methodology, a governed panel, a restatement policy, a published construction that a third party could reproduce, and an owner whose interests do not run against the number.

The regulator states the last of those unusually plainly, in a document worth knowing by name. Appendix C to Part 38 of the CFTC’s regulations is the Commission’s guidance on how a designated contract market — a registered futures exchange — demonstrates compliance with the core principles it must satisfy to list a contract. Adopted in 2012, it is not a rule that binds an index provider directly; it binds the exchange, and it tells the exchange what to check before settling a contract on somebody else’s number. That indirect posture is why it matters here: an administrator has no CFTC registration to lose, but an exchange that cannot answer Appendix C cannot list. The relevant paragraph, (c)(3)(i), says that where a third party calculates the cash settlement price, the exchange should verify that the provider “utilizes business practices that minimize the opportunity or incentive to manipulate” — and offers, as examples of such practices, “lock-downs, prohibitions against derivatives trading by employees, or public dissemination of the names of sources and the price quotes they provide.”

Read that sentence closely, because the conjunction is the whole test. Opportunity is addressable by policy: information barriers, employee trading bans, clean rooms. Incentive is not. An administrator whose owner trades the products that settle on its series carries the incentive by construction, and no compliance manual removes it. That is the standard the compute complex is now being measured against, and it is a higher bar than the market has been holding itself to.

One clarification before the survey, because it changes what counts as a gap. “Published” in benchmark administration does not mean free. It means the methodology is disclosed and the construction is reproducible; the assessments themselves are almost always sold. IOSCO’s Principles for Financial Benchmarks require an administrator to “Publish or Make Available” its methodology in enough detail for stakeholders “to understand how the Benchmark is derived and to assess its representativeness” — and the defined term resolves to posting a document on the administrator’s website. The word Subscribers appears in the Principles themselves. Appendix C likewise attaches “publicly available” to the underlying cash price series, not to the index product, and routes methodology disclosure to the exchange and the Commission rather than to the world. Platts and Argus are the working model: methodology guides free to anyone, assessments behind a subscription.

So a paywall is not a deficiency, and this piece does not treat one as such. What follows distinguishes between a series that exists and is sold, a series that exists in a form short of settlement grade, and a series that does not exist at all — because only the last is a gap that no commercial decision can close.

The second institutional property is the one that decides whether a series can exist at all: somebody has to contribute the data, and contribution is a commercial act. The question every benchmark answers, before it answers anything about weighting or windows, is why a data-holder hands over something valuable. There are only about six workable answers.

Benchmark administration is a mature craft with a small number of working forms. Each is really a different answer to why anyone contributes, and each suits a different kind of underlying. Compute has instances of four of them. The two it lacks are the two the missing series require.

The pattern is worth pausing on. The four models compute has adopted are the four that need no contributor to agree to anything. Scraping needs no permission. Deriving a forward from a term structure needs no permission. An assessment needs reporters, but reporters submit because the assessment already moves their own contracts — a bootstrapping problem in a market where nothing references the assessment yet. Only the transaction-print aggregator has solved contribution properly, by paying in hedge access rather than cash.

The two missing models are the two that require an administrator to negotiate with a data-holder who has something to lose. That is not an accident of youth. It is the market taking the path that does not require institution-building.

This is more coverage than the market had a year ago, and the differences between the families are real: Ornn builds from executed transactions and says so, Silicon Data spans posted and transacted rates across a wide provider set, SemiAnalysis publishes the most granular tenor structure in the market, NATIVX normalises by energy and has put a second, open-source construction beside its licensed one, Compute Desk sits closest to a settlement workflow and is now the only family administered by a regulated third party. Anyone claiming these are interchangeable has not read them.

But step back and the shape is uniform. Every one of them measures the price of an on-demand GPU-hour, and the on-demand market is the residual. That includes the newly regulated family: what GX Benchmarks calculates for Compute Desk, to whatever standard, is still the rent. The capacity that matters commercially is sold two to five years forward at a fixed price, and every participant with a real exposure prices it that way. Five providers are competing to publish the flat price of the thin part of the market. A sixth would be noise — and in a sense the sixth has already arrived and priced itself accordingly: FSKU reproduces the leg from public rates under an Apache license, for free. When the marginal construction of a leg is an open-source repo, that leg is measured.

The interesting layer is beneath the flat price. Four series would each carry a real exposure, and not one of them exists in a form a contract could reference. But they fail in different ways, and the differences matter more than the headline: one is flatly absent, two exist in forms that fall short of settlement grade in a specific and stateable way, and one exists in three incompatible constructions at once. Being imprecise here is how a market talks itself into believing it has infrastructure it does not have — or, just as bad, into ignoring what it does.

The token series is the one this practice has written about most, because it is the leg that actually moves. Rent has been close to static — weekly changes under a percent, with the hyperscaler series unchanged on 39 of 50 trading days — while token prices reprice within hours across a hundred-odd endpoints. We built the crack spread from a live endpoint book to show what that series would look like: reference token revenue per GPU-hour less the rent, and the market heat rate underneath it. Two thirds of the active book clustered within a few percent of one another at a heat rate a saturated H100 covers three times over, and almost nobody cleared at low load.

What that exercise also showed is the limit. Dispersion across active endpoints was five-fold for one declared good on one screen, and the residual after search costs is unverified grade: precision is self-declared at some endpoints and simply unknown at others, and serving configuration moves published benchmarks by double-digit points. A token series without an attested grade carries that dispersion as noise. Which is the general lesson of this section: each missing series fails for a different reason, and only one of the four fails for want of arithmetic.

A data gap is a solvable problem: someone builds a collection apparatus and fills it. The structural deficiency in this market is that every administrator has sat inside a commercial position that the Appendix C test — the exchange’s obligation to check its index provider’s incentives — is aimed at, and the market has treated that as normal. In August it acquired its first structural exception, which is examined below on its own terms.

To be explicit about what is and is not being said: none of the arrangements below is improper, none is hidden, and several are disclosed by the firms themselves. Vertical integration is how young markets get built — somebody has to fund the index, and the people who understand compute pricing are the people trading compute. The point is narrower and it is the Commission’s point, not ours: incentive is a separate test from opportunity, and it is not satisfied by conduct policy.

Two August answers, one incentive test

The uniform picture above acquired its first two genuine exceptions this month, and they are worth separating carefully because they answer different halves of the problem in different currencies.

The first bought independence institutionally. On August 13, Compute Desk and General Index announced that GX Benchmarks Ltd — a wholly-owned subsidiary of General Index, the London algorithmic-benchmark administrator regulated by the FCA — now calculates, publishes and maintains regulatory oversight of Compute Desk’s daily H100, Hopper and Blackwell benchmarks, built to IOSCO Principles and EU/UK Benchmark Regulation standards. This is the arrangement Appendix C actually contemplates: a third-party index provider whose incentives the exchange can examine and find clean. General Index does not trade compute, sell capacity, or hold positions in anything that settles on the number; its commercial incentive runs toward the credibility of the administration franchise it operates across oil, gas and refined products. For the administrator-level incentive test, that is not a compliance manual — it is an owner whose interests genuinely do not run against the number, and it is the first time a compute benchmark has been administered by a firm with no commercial position in compute at all.

What it is not is a closed question, because the test does not disappear — it moves down a layer. The transaction data underneath the GX calculation is Compute Desk’s, and Compute Desk still operates Compute Clear beside it. The Commission has already written the question that survives the fix, in 1(f): whether “the parties best positioned to influence the reference price are the same parties that supply capacity or contribute transactions” — a question asked of the cash market, which no administration arrangement can answer on its own. The honest summary is that the arrangement converts an administrator-incentive problem into a contributor-concentration problem, which is a strictly better problem to have and the same one Platts and Argus manage daily. It also, per §03, administers the same leg as everyone else: the rent.

The second bought reproducibility technically. NATIVX has published FSKU, an Apache-licensed open-source platform that ingests posted rates across AWS, Azure, GCP, CoreWeave, RunPod and Lambda, normalises them into $/GPU-hour, and derives a model-implied forward curve — with the entire construction, seed data included, released as code. Its most interesting features are the ones the commercial administrators do not publish: a provenance ledger naming every source and the normalisation applied to it, a sensitivity matrix showing how six different aggregation methods move the spot index, and a source-ablation view showing the exact price impact of excluding any single provider from the calculation. That last one is, almost verbatim, the attack surface question 2(b) asks about — published as a stress test anyone can run. Open-sourcing does not solve contribution; FSKU is the no-permission model in its purest form, posted rates in and a modelled forward out, and a modelled forward is an assumption set, not a price. But it establishes a floor the paid products now sit above: for one construction in this market, any outsider can rerun the number, and reproducibility — the §01 property usually taken on faith — is checkable rather than claimed.

Set against that, the single most striking absence in the whole survey. Appendix C names three example safeguards an exchange might look for, and one of them is “public dissemination of the names of sources and the price quotes they provide.” Not one compute administrator names a contributor. Silicon Data invites data partners without publishing a roster; Ornn shows unlabelled logos; Compute Desk names none, and the GX methodology describes its inputs as live rental quotes and completed transactions across major GPU cloud marketplaces without naming them; SemiAnalysis describes its pool as “100+ market participants” and its inputs by category. A reader cannot tell, for any published compute index, who is in it — the regulated one included.

The nearest thing to an exception proves the rule. FSKU’s provenance ledger names every source, with direct URLs — because every source is a public price sheet that costs nothing to name. The roster that matters, of parties whose private prints move a settlement value, remains unpublished everywhere.

That is not a small thing to be missing. It is the disclosure that lets an outsider judge concentration, and concentration is exactly what question 2(b) is about.

The contributor mutual — model E — is the structure that unlocks the data compute lacks, and the insurance industry ran the full experiment over five decades. It is worth studying properly, because the ending is usually told as a success story and the middle is not.

The problem was credibility, not generosity. An individual insurer cannot file a rate on its own loss experience if that experience is too thin to be statistically credible. Pooling solves a problem the contributor cannot solve alone, which is the only durable reason anyone contributes anything. The Insurance Services Office was formed in 1971 out of a consolidation of existing state, regional and national rating bureaus, and what members received in exchange for loss data was concrete: standardised policy forms, and advisory prospective loss costs filed with state regulators on their behalf — more than 3,000 filings a year.

Independence came before commercialisation, not after. This is the part usually skipped. In 1995 the member insurers ceded control to a board with a majority of independent, non-insurer directors. Only then, in a vote of November 1996 effective January 1, 1997, did members convert ISO to a for-profit corporation — with roughly 85% of shares held by member insurers as restricted Class B stock, and seven of eleven board seats attached to non-insurer Class A holders. The 2009 listing as Verisk Analytics came twelve years later and was entirely secondary: about $1.9 billion raised, none of it for the company, existing insurer owners selling down. Verisk today runs at roughly $3.1 billion of revenue, effectively all of it insurance data and analytics.

Two other precedents narrow the design. Markit showed that equity-for-data works fast: founded in 2003 on dealer-contributed credit pricing, with the contributing banks holding roughly 70% of the equity in the early years, falling below a third after the 2014 listing. The Baltic Exchange shows the other edge — panellists were historically informal and unpaid, compensated by fee waiver and standing rather than cash, and the attempt to formalise panellist agreements drew an organised objection from brokers in 2016. Contribution arrangements are contested even when everyone agrees they are necessary.

One design rule falls out of all three. Contribution is priced in kind before it is priced in cash, because paying cash for prints creates precisely the incentive Appendix C asks the exchange to minimise. Panel weight, depth access, governance seats and equity that vests on delivery and accuracy — never on level — are the currencies that survive the test.

On August 19 the CFTC issued a request for comment on the listing of compute derivatives, published at 91 FR 54259, docket CFTC-2026-1850, with comments closing October 20. It is generally read as a threat to the October listing. Read instead as a product specification, it describes with unusual precision the series this market does not have.

Notice what the Commission is not primarily asking about. There is comparatively little on calculation method and a great deal on disclosure, concentration, contributor identity and utilization. That is a regulator asking for a cash-market exhibit and a meter, not for a better average. The August developments sharpen the point rather than blunting it: a regulated administrator and an open-sourced calculation are both answers about the average, and the Commission’s questions are about the market underneath it. It is the same conclusion the crack-spread construction reached from the other direction: the binding constraint on this complex is not the arithmetic of an index but the observability of the market underneath it.

Not a prediction, and not a proposal for anyone in particular. These are the observable conditions under which a compute benchmark becomes referenceable, stated so that a reader can check them off as they happen.

• A published contributor roster. The first administrator to name its sources and their quotes satisfies the one Appendix C safeguard nobody currently meets, and the disclosure is cheap for a provider with genuine depth and expensive for one without. That asymmetry is the point. Neither August development produced one: FSKU names only public price sheets, and the GX arrangement names its input categories, not its counterparties.
• An attested meter with a certification standard. Telemetry exists at every operator; what does not exist is a third party that certifies the agent, publishes the methodology, and can be relied on by a lender against the operator. This is the only flat absence on the list — there is no utilization series to subscribe to, which means the gap cannot be closed by anyone opening a paywall. Until it is closed, no volume-contingent product is writable.
• A term series built from disclosed prints. Tenor curves exist; disclosed prints behind them do not. The renewal is where the credit exposure lives, and it is the field currently absent from every published construction.
• A token series with mix held constant and grade attested. The leg that floats needs a fixed basket and a declared, verified precision. Without the first it falls when users switch models; without the second it carries five-fold dispersion as noise.
• A residual series that reconciles the three that exist. A model-implied curve derived from rental forwards, a band built from observed secondary transactions, and an indicative quote range are three different objects, and they will not agree. Until one is daily, transaction-based and governed, residual-value swap strikes will keep being negotiated bilaterally rather than struck against a reference — which is the observable test of whether a residual benchmark exists.
• Governance independence established before commercial scale, not after. The insurance precedent is explicit about the order: independent board control in 1995, conversion in 1997, listing in 2009. Reversing that order is what the antitrust record is about. The Compute Desk–General Index arrangement is the first instance of the separation in compute — achieved by hiring an existing regulated administrator rather than building governance — and it checks this box at the administrator layer while leaving the data-provider layer, where the Commission’s 1(f) question lives, open.

None of these requires a new calculation. All of them require somebody to negotiate with a data-holder who has something to lose — which is the part of benchmark administration the compute market has, with one August exception, not started. That exception is instructive about the rest of the list: Compute Desk had something to lose — control of its own number — and traded it to a regulated administrator for credibility. The open-source route gave up nothing, because posted rates were never anyone’s to withhold. The four series in §04 sit behind data-holders with far more to lose than either, which is why neither August move produced any of them.

• Regulation. CFTC request for comment on the listing of compute derivatives, 91 FR 54259, August 21, 2026, RIN 3038-AF77, docket CFTC-2026-1850, comments close October 20. Appendix C to 17 CFR part 38, paragraph (c)(3)(i), on third-party index providers. FERC’s 2003 price-index policy statement, Docket PL03-3-000, on contributor codes of conduct.
• Provider methodologies. Silicon Data forward curve and LLM token expenditure index; Ornn Data / OCPI; Compute Desk, with the GX administration arrangement per General Index’s announcement of August 17, 2026 (GX Benchmarks Ltd; FCA reg. 933348); SemiAnalysis GPU index methodology; NATIVX COIL per the ICE announcement of July 1, 2026, and FSKU per the nativ-x/fsku repository (v0.9.0, Apache-2.0, retrieved August 28, 2026).
• Administrator positions. Silicon Data seed, co-led by DRW and Jump and Series A of August 11, 2026; Axios on Ornn’s de minimis swap activity; ICE and Ornn, May 19, 2026.
• The insurance precedent. ISO member vote to convert, November 1996 (including the 1995 transfer of control to a majority-independent board); In re Insurance Antitrust Litigation, 938 F.2d 919 (9th Cir. 1991); 723 F. Supp. 464 (N.D. Cal. 1989) on the 1986 forms; Hartford Fire Ins. Co. v. California, 509 U.S. 764 (1993); Verisk FY2025 Form 10-K. Baltic Exchange panellists.
• Adjacent market facts. BGC Compute Infrastructure Markets, June 18, 2026; American Compute on used-GPU price discovery; Barker warrantied valuations, backed by Munich Re’s aiSure; InfraX and the workload compute unit.
• Kinetic Alpha. The Compute Crack Spread — the graded token series and the market heat rate, built from a live endpoint book. The Contract Has a Date. It Needs a Dealer. — who holds a floating exposure, and why the physical market prices fixed. How the Indices Calculate Price — the component-level catalogue. Silicon Data’s Product Suite — the eleven-product audit and the utilization blind spot. The Token Leg Has a Price Screen — grade construction and router microstructure.