The Contract Got a Date.

The August 11 release discloses the following, completely. Two cash products: Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures, each representing “a month’s worth of rent” for the referenced NVIDIA GPU, tracking indices “measuring hourly rental GPU costs published by Silicon Data.” Listing October 5, 2026, “listed and subject to the rules of NYMEX,” pending regulatory review, with further detail promised at cmegroup.com/compute. Two quotes: CME’s global head of energy and environmental products, Pete Keavey — “Compute has become the currency of the AI age, and this innovative market will bring transparency to the current and future costs that AI builders and hyperscalers need to hedge as they grow” — and Silicon Data CEO Carmen Li: “Compute futures give the market something it’s never had: a public, tradable reference price for the resource every AI system runs on.”

Three reads on the launch set itself. First, the pair: H100 and B200 — not H200, not A100, not a composite — is the incumbent workhorse and the current-generation flagship, printing $2.73 and $5.59 per GPU-hour respectively on Silicon Data’s dashboard this week. A two-rung ladder makes the grade structure of the complex tradable from day one, and the first quarter’s tape — B200 up 24% in a month while H100 sat flat — suggests those two legs will not move together. Second, the venue placement: NYMEX is CME’s energy exchange, filing compute alongside crude and power rather than equities — a small institutional confirmation of the compute-as-congealed-electricity frame this site has argued from the start. Third, the timing: the announcement came the morning after NVIDIA’s $500 billion financing-platform news, which means the reference price Li describes now has a customer class an order of magnitude larger than the hedgers anyone modeled in May.

An index is an average over dispersion — so before asking how the settlement index is computed, ask what it has to average over. The answer is observable in public prints today, and it is worth laying out as a ladder, because each rung gets a different fate in the market structure that is now assembling:

The tier rung deserves the closest read, because it is the one the futures design quietly takes a position on. Identical H100 silicon prints 2.6× apart depending on who operates it — a differential that is not hardware but SLA, credit, interconnect and enterprise contracting, and that Silicon Data began publishing as separate daily sub-indices only after the May announcements. The two series also behave differently: this week’s prints show the hyperscaler series essentially unmoved (+0.1% and 0.0% on the 7-day for H100 and A100) while the neo-cloud series did the moving (−1.8%, +0.6%) — the pattern of a sticky, contract-driven price sitting on top of a fast spot market. The contracts listing October 5 settle on the fast one. That is a defensible design choice — the neo-cloud market is where transactions concentrate and prices discover — but it has a direct consequence for the hedgers the announcement names: an enterprise hedging a hyperscaler compute bill with these futures is hedging a $7.19 exposure with a $2.73 instrument, and the correlation between the two series is an empirical question the sub-index history is only now long enough to start answering.

Dispersion is also the honest frame for the whole commoditization question, which our compute-complex piece put at the center with an interactive dispersion dashboard: commodities become commodities by sorting their dispersion — crude normalized quality into gravity-and-sulfur adjustments, traded its grade spreads into benchmarks like WTI–Brent, and left location basis as a permanent, priced residual. Every rung of the compute ladder above is visibly mid-sort: performance variance being normalized by a proprietary model, the grade spread becoming a listed market on October 5, the tier spread segmented into sub-indices but outside the settlement sample, and the token axes not yet sorted at all. One useful property of this week’s announcement is that the sorting now happens in public, on a dated schedule.

Silicon Data’s CEO has spent the summer making the case in public, and the record is worth assembling because it previews the design arguments the certification filing will have to settle. On Bloomberg’s Odd Lots (June 15), Li described the market’s procurement shift in terms any commodity trader would recognize: “you see a lot of people shifting from on-demand to reserve, even forward contracts” — and the hedging logic that follows: “producers… need to hedge their revenue volatility by shorting futures or put options.” She put daily H100 cluster volatility “around 20–30” — “a very healthy commodity volatility range” — and, notably for anyone reading the contract announcement two months later, was still hedging on the product question itself: “whether there’s interest in compute futures… it’s very TBD.” From TBD to a NYMEX listing date in eight weeks is its own data point about how fast the demand side moved — or was moved.

Two of her design claims bear directly on settlement. On fungibility — the objection every compute-commodity skeptic reaches for first — she conceded the premise and quantified it: “one chip might not necessarily be equal to another chip,” and Silicon Data’s own benchmarking “proved there’s 38% performance variance for the same chip.” That number is the top rung of the dispersion ladder in section 02, and it cuts both ways. It is the strongest public argument that naive averaging cannot produce a settlement price — and it is precisely why the index’s normalization layer matters so much: “the way we develop our index model is not simple math… it’s not, hey, you have two H100s, do a simple average.” A 38% same-SKU dispersion normalized away by a proprietary model is either the index’s core intellectual property or its core attack surface, depending on whether the normalization is ever published — which is exactly the tension the certification filing will have to resolve. Her earlier public framing (Bloomberg video, May–June; industry talks) rounds out the pitch: compute as a coming global commodity, financial infrastructure “akin to traditional commodity markets,” spot, forward and reserve contracts as the maturity path, and pricing opacity at both hyperscalers and neo-clouds as the problem the index solves.

The futures announcement concerns GPU rental rates. But Silicon Data’s most-watched print this summer has been a different instrument entirely: SDLLMTK, the LLM Token Expenditure Index — a daily, expenditure-weighted average price of a million LLM tokens across 400+ tracked models and a claimed 90%+ of global inference spend, normalized for input/output mix, context window, batching and reliability. Silicon Data itself has been careful about what the number means, noting publicly that it “should really have been named the Token Expenditure Price Index” — a usage-weighted average of what the market is currently paying per million tokens, irrespective of model. The segment split matters as much as the composite: frontier-model inference around $4.20 per million tokens against $0.85 for open-weight serving, a five-to-one spread for capability that our token-index teardown found to be seven-to-one at the venue level for identical open weights — dispersion that any composite necessarily averages over.

The reason it is being watched: the index roughly doubled from its December 2025 inception into a May 2026 peak, and has since given back nearly 20% — including an 11% drop in the past week — and Bloomberg has taken to calling it “the cleanest read” on whether the $700-billion-plus AI capex cycle is earning its keep. The bull reading is mix-shift: cheaper tokens expanding the addressable market, spend migrating to open-weight serving, total token expenditure still growing. The bear reading is pricing power visibly eroding just as the financing layer scales. Both readings route through the same arithmetic we built in the inference spark spread: token revenue against GPU cost, joined by serving efficiency — which is precisely why a token index falling while the H100 and B200 rental indices hold their levels is not a contradiction but a margin signal, and why the two index families are more informative together than either is alone.

On token forwards specifically, the public record should be stated carefully. Silicon Data’s documentation stack now includes a Token Pricebook, a Token Market Pulse daily, a Token Index API and a forward-curve API alongside the GPU products — the full scaffolding of a two-commodity data house, tokens and compute. Li’s public remarks frame forwards as where the compute market is already going. But no token futures or token forward contract has been announced by anyone: the October 5 listing is GPU rental only, and the only listed token-adjacent structure remains the event-market ladders we covered in the cross-venue basis work. The token leg, for now, is an index with a growing audience and no tradable instrument — the same state GPU rental was in eighteen months ago, which is perhaps the most suggestive fact about it.

Because this piece makes methodology claims — and because “the index has no methodology” is the kind of line that circulates — the documentation state deserves one precise paragraph, no more. It is a three-bin inventory, dated August 11 and guaranteed to change. Public: daily index levels with Bloomberg tickers, including the tier sub-indices; a four-stage process description (observation capture across a claimed 80%+ of the rental market, normalization, daily outlier-filtered calculation, distribution); and — genuinely to the provider’s credit — a documentation portal carrying dated, quantified index-change and restatement announcements, with history restated to the indices’ September 2024 start dates. Available behind registration or subscription: the full methodology reference via the client portal, and the index and forward-curve APIs on paid tiers — we have not reviewed the gated document and make no claim about its contents, but any suggestion that the methodology is simply secret is wrong. Absent from the public record: a benchmark-style public rulebook — the IOSCO-aligned methodology statement, governance and contributor framework that settlement benchmarks in mature commodities publish as a matter of course, and that our August 1 audit found nowhere across the four compute index families — along with the contract specifications (multiplier, final-settlement procedure, listed months) and the CFTC certification filing.

The third bin is on a clock: a contract cannot list on a designated contract market without the certification landing first, and Appendix C to Part 38 requires public demonstration that the contract is not readily susceptible to manipulation — cash-market and index-construction disclosure at a level nobody has yet published. An October 5 listing therefore bounds the most substantive disclosure event in this asset class’s history to the next seven weeks. The board in section 07 grades all five venue-index pairs on this same standard.

ICE×Ornn has not answered with a date. The rival May 19 announcement — contracts settling on Ornn’s executed-print, invoice-verified OCPI rather than a quote-based assessment — remains undated as of August 11. Until it resolves, the cross-venue index basis the two announcements set up in May exists only in hybrid form, via the event-market ladders that already settle on Ornn prints. The benchmark-technology contrast — quotes versus prints, composite versus regional weighting — is the deepest structural question in the complex, and it now has a first-mover asymmetry attached: one benchmark family gets a listed market, a certification filing, and seven weeks of institutional attention before the other has a date.

Architect’s American Innovation Exchange advanced by acquisition rather than announcement. Its May 28 disclosure that it had acquired IMX Health, LLC — a designated contract market the CFTC approved in 2024 — means the AI Exchange inherits DCM status and needs product-level filings rather than de novo designation. Since our July coverage of the ComputeConnect exchange-for-physical architecture, the exchange site has published the most granular product map of any compute venue — GPU-hour futures and options on H100, H200, B200 and B300, spot through 24 months, cash-settled and marked daily against an “AI Exchange daily index” — and opened an early-access onboarding funnel. Still absent publicly: a rulebook, a listing date, a named clearing arrangement for the US futures, and any methodology from Compute Desk, whose indices the EFP basis tables reference. (Separately and offshore: Architect’s Bermuda-regulated perpetuals, which settle on Ornn prints, were opened to individual investors and omnibus accounts by the Bermuda Monetary Authority on July 21.)

And the wrapper layer is queued behind all of it. Thirteen compute-futures ETFs across six trusts sit registered with the SEC — filed this spring against contracts that did not yet have a listing date — with none yet effective; Roundhill’s GPUX page still carries the not-yet-effective legend as of this morning. Their amendment filings, when they come, will name benchmarks, fees and — most consequentially — which venue’s contracts they hold: a decision the October 5 date just made considerably easier to make. A market that lists on October 5 with a financing stack announced Monday and a wrapper cohort pre-registered since spring would compress into five months a sequence — futures, then funds, then financing — that took crude oil decades. Whether the sequencing order matters is a question we take up separately.

Two instruments below. The clock tracks the dated and undated milestones between now and October 5. The disclosure board grades every venue-index pair on what is actually in the public record — with the Silicon Data column scored on the three-bin standard from section 05: green requires a public document, amber marks partial or registration-gated availability, red marks confirmed public absence. Click any cell for the evidence.