Research · Compute · Index methodology · 1 August 2026
Four providers are positioning to become the settlement benchmark for GPU compute derivatives, and they represent four fundamentally different answers to the same question: what is the price of an H100-hour? This catalogue documents, component by component, what each of Silicon Data, Ornn, Kalshi, and Compute Desk actually publishes, how each price is calculated as far as public record allows, who feeds the data, and — because none of it currently meets settlement-benchmark standards — the specific regulatory mechanics that will force disclosure over the next two to three quarters.
This is the transparency baseline underneath the practice's compute work: the IOSCO assessment and implied forward curves, the EFP analysis, and the lender hedge-program design all hit the same wall — undisclosed methodologies — from different directions. What follows fixes the record of what is knowable as of August 1, 2026.
Four providers are positioning to become the settlement benchmark for GPU compute derivatives, and they represent four fundamentally different answers to the same question — what is the price of an H100-hour?
| Silicon Data | Ornn | Kalshi "Compute Forward Curves" | Compute Desk | |
|---|---|---|---|---|
| Core input | Observed/posted rental rates (list prices + private-platform quotes), ~150k records/day | Executed transactions only ("printed trades," invoice-verified) | Kalshi's own prediction-market trading (market-implied) | Undisclosed; GPU rental price indices tied to its ComputeConnect physical-delivery network |
| Utilization in the calc? | No — pure $/GPU-hr price benchmark | No — pure transacted $/GPU-hr | No — implied forward price only | No indication — rental price indices |
| Weighting | Proprietary; weights by source reliability, geography, "participation levels" (aggregation statistic undisclosed) | Volume-weighted average (GPU-hrs vs. dollars undisclosed) | Order-flow determined (market prices, no administrator weighting) | Undisclosed |
| Geography | Global (40–50 countries), normalized inside the model; no published regional weights | US flagship; claims city-level granularity (LMP analogy); headline aggregation undisclosed | None published; lists contracts by chip, expects market to converge on a benchmark | Location handled via basis tables (by SKU, memory config, location) for physical delivery |
| Component coverage | H100, H200, A100, B200, MI300X; neocloud vs. hyperscaler split | H100 SXM, H200, B200, A100 SXM4, RTX 5090; on-demand + term | B200, H200, A100 curves (H100 notably excluded) | H100, H200, B200, B300 |
| Public methodology | No rulebook; 4-step process description + change log | No (an Oct 2025 methodology paper was taken down); API docs only | Conceptual blog only; curve-stitching math undisclosed | None |
| IOSCO / audit / governance | None found | None found | None found (but underlying contracts are CFTC-certified DCM specs) | None found |
| Exchange path | CME Group futures (announced May 12, 2026; H2 2026, pending regulatory review) | ICE futures (May 19, 2026, pending); Kalshi settlement; Architect (Bermuda) perps live; own DCM ambitions | Already listed — event contracts live since Mar 2026 on Kalshi's DCM | Architect's American Innovation Exchange (US DCM, pending) — futures + ComputeConnect EFP (announced Jul 8, 2026) |
Three important structural findings:
Nobody uses utilization as a price input. All three are price-of-rental benchmarks (USD per GPU-hour), not utilization or revenue indices. Utilization shows up only indirectly — e.g., Silicon Data reportedly weights data sources partly by "data center participation levels in rental markets." The price-vs-utilization gap is one of the most cited criticisms of the entire construct, since operator economics (and lender underwriting) hinge on utilization, which no scraped or contributed price feed observes — the gap our measured GPU power work quantifies from the energy side, and which drives the tracking-error check in our lender hedge-program design.
The list-price vs. transacted-price divide is THE methodological fault line. Bernstein frames the CME/ICE race exactly this way: Silicon Data (CME) aggregates observable posted rates; Ornn (ICE) prints only negotiated transactions. Each approach has a defect — list prices overstate what large buyers actually pay (private discounts, bundling), while a transaction-only index built from ~10 contributors is thin and concentration-exposed.
No provider currently meets settlement-benchmark governance standards. No published rulebook, no IOSCO Principles compliance statement, no independent audit, no oversight committee, no contributor code of conduct — for any of the three. That's not a search gap; it's the current state — and it is the same conclusion our IOSCO principle-by-principle assessment reached from the governance side. CFTC listing mechanics (below) are the forcing function that should change this over the next two to three quarters.
Daily indices, USD per GPU-hour, published once per business day (the "RT" in tickers overstates frequency — capture is continuous, publication is daily):
| Index | Ticker | Segment | Live since |
|---|---|---|---|
| H100 Neocloud | SDH100RT | Neocloud | May 2025 (history from Sep 1, 2024); $2.77 on 7/30/26 |
| H100 Hyperscaler | portal-only | Hyperscaler | Dec 2025; $7.18 |
| A100 Neocloud / Hyperscaler | SDA100RT / portal-only | — | Dec 2025; $1.64 / $3.71 |
| B200 Neocloud | SDB200RT | Neocloud | Dec 4, 2025; $5.66 |
| H200 Neocloud | — | Neocloud | Jul 15, 2026; $3.10 |
| MI300X (AMD) | — | — | Mid-2026; $2.61 |
No GB200/NVL72 index exists yet. Also publishes a no-arbitrage GPU forward curve derived from term-structure rental data (1–36 month leases), an LLM token expenditure index, and a GDDR6 RAM index. The Dec 2025 neocloud/hyperscaler split is the key segmentation: hyperscaler H100 prices ~2.3–2.6x neocloud.
Global coverage (40–50 countries). Geography is a normalization input and a weighting factor, not a published regional weighting scheme — whether weights reflect capacity, dollar consumption, or listing counts is undisclosed. No public in/out region list; no public regional sub-indices (though the firm publishes regional analyses ad hoc, e.g., US East $5.76 vs. US West $6.80 for H100 in Mar 2025).
What's public: a 4-step process description, the normalization variable list, the lease taxonomy, and — importantly — a public announcements log with pre-announced methodology changes, effective dates, and quantified impact estimates (index-industry best practice). What's not: aggregation statistic, weights, outlier rules, provider list, reference configuration, restatement policy, governance.
The change log itself is revealing about maturity: - Dec 2025: first major revision, restated the full history back to Sep 2024 — SDH100RT moved −4 to −6%, but SDA100RT moved +35 to +40%. A near-40% restatement of a benchmark's entire history is disqualifying for anything already settling contracts; doing it before listing was the right sequencing, but it shows how young the methodology is. - Mar 2026: provider addition, −3 to −7% impact on SDH100RT. - Jun 2026: B200 provider expansion (up to −6%) and H200 launch.
Provider changes moving the index mid-single-digits per event is direct evidence of composition sensitivity — the index level is partly a function of who's in the panel, which is exactly what CFTC review will probe.
The Ornn Compute Price Index (OCPI) family: OCPI-H100 SXM (flagship; Bloomberg ORNNH100), OCPI-H200, OCPI-B200, OCPI-A100 SXM4, OCPI-RTX5090. On-demand and term rates tracked distinctly. Also token price indices (realized $/M-token for OpenAI/Anthropic), DRAM/flash spot indices (underlying Architect's RAM perps), and forward curves. Founded 2025 (Menlo Park; ex-SIG/Optiver/Google/MIT team); $5.7M seed Oct 2025, then $33M led by a16z (June 2026). API publishes hourly updates with a 24-hour rolling average and a daily settled close.
Flagship is a US H100 index. Ornn markets city-level granularity — the Northern Virginia vs. Amsterdam framing, analogized to electricity LMPs — and the API exposes region parameters, but how regions roll up into headline indices is undisclosed.
This is where Ornn's model differs most: it is a contributed-feed + own-marketplace model, and it does have named partnerships: - Hydra Host (Oct 2025): neocloud/brokerage with 30,000+ GPUs under management across 50+ locations, contributing real-time infrastructure data. - InfraSight Software cited as a partner; Hyperbolic Labs' CEO endorses in the ICE release (data relationship inferred). - "10+ data partners" total, mostly unnamed, under contract with invoice-level verification; plus prints from Ornn's own compute marketplace. Claims 400+ data center operators, investors, and AI companies on the platform; authorized for up to $8B notional in swaps under the CFTC de minimis exemption; FalconX executed the first OTC compute forward on OCPI H100 (May 2026). - No lender/lessor partnerships found by name — but the invoice-parsing pipeline is precisely the artifact a lender syndicate would want, and Ornn markets "capacity finance" and residual-value products aimed at that audience. Hyperscaler contribution: no evidence.
Ornn is the strange case: it moved backwards before it will move forwards. A public methodology paper ("Ornn's US H100 Compute Price Index Methodology," Oct 11, 2025) was published and later taken down — the URL now redirects to the homepage and no archive copy survives. Current public disclosure is API docs plus marketing-level claims in exchange press releases (transaction-only, VWAP, normalization, Asian settlement). Withheld: contributor list, normalization formulas, outlier rules, region weights, volume definition. No IOSCO statement, audit, or oversight committee. The stated goal (CEO Kush Bavaria) is an "anti-manipulative, transparent, and verifiable" index users can replicate — replication is impossible today. The ICE listing (and Ornn's own DCM license pursuit) is what will force the methodology back into the open, this time in rulebook form.
A note on identification: there is no index-administrator company named Compute Forward. The phrase belongs to Kalshi's product — "Compute Forward Curves," announced July 14, 2026. (A separate, apparently stealth "ComputeForward.com" forward GPU-cluster marketplace exists — B300 through A100, 1–48 month terms, tiered seller verification — but it publishes no prices and has no discoverable corporate footprint.)
Kalshi doesn't assess a price; it derives implied forwards from its own regulated prediction-market order flow: - Underlying: CFTC-certified event contracts on GPU rental prices (live since March 2026), structured as "Will B200 compute be above $X by date Y?" across strikes and weekly/monthly/quarterly expirations. - The curve: an algorithm stitches those contract prices into implied forward curves for B200, H200, and A100 — H100 conspicuously excluded. Curves are explicitly non-tradeable references for structuring OTC deals; hedging happens in the underlying contracts or via block trades (mechanism proven April 2026). - Settlement anchor: Ornn. Kalshi's GPU contracts cash-settle to OCPI — so Kalshi's market-implied curve ultimately keys off Ornn's transaction prints. Ornn is upstream of both ICE and Kalshi — the settlement concentration our implied-forward-curve work flagged as the complex's single largest governance exposure. - Utilization: not an input. Geography: none — Kalshi's CRO Udesh Jha (16-year CME veteran) openly concedes non-fungibility ("H100s in Virginia perform differently than A100s in Frankfurt") and argues the market will converge on a representative benchmark the way crude converged on WTI/Brent, rather than an administrator solving heterogeneity by normalization.
The curve-construction math (bootstrapping/interpolation, weighting across strikes) is undisclosed; there's a conceptual blog, not a spec. But Kalshi's transparency profile is structurally different: the underlying contract terms are public CFTC-certified DCM specifications, and the "methodology" is largely just the market. Kalshi's pitch — "unlike other forward curves, Kalshi's are backed by the market" — is a direct shot at Silicon Data's model-derived no-arbitrage curve. The trade-off: prediction-market prices embed risk premia and thin-liquidity noise, and Jha's argument that event-contract structure delivers "robust price discovery at much lower levels of total open interest" is unproven at institutional hedging scale.
Because Kalshi's instruments are already listed, there is no pending regulatory milestone forcing further methodology disclosure — unlike CME/ICE, whose partners face benchmark-grade scrutiny before launch.
The cash-settled index race in Sections 1–3 has a physical counterpart: Compute Desk (legal name: The Compute Index, Inc.), whose indices anchor the first US exchange-for-physical network for GPU compute — the mechanism we analyzed in depth in Paper becomes racks.
Announced July 8, 2026: the "U.S. financial industry's first compute exchange-for-physical network." Mechanics: - Qualifying compute futures positions on Architect's forthcoming American Innovation Exchange (a US DCM, via Architect's DCM acquisition announced May 28, 2026; pending regulatory review) can be converted into physical GPU capacity. - An open protocol lets capacity providers respond to delivery requests; the physical leg clears via Compute Desk's Compute Clear platform; the futures leg books to the exchange. - Reference prices: Compute Desk's own GPU rental price indices for H100, H200, B200, and B300 — notably not Ornn's, even though Architect separately lists Ornn-referenced perps on its Bermuda AX venue. Architect now runs two index partners across two venues (Ornn offshore, Compute Desk onshore). - The standardization payload: basis tables by GPU SKU, memory configuration, and location — the first public attempt to define location/config differentials for physical GPU delivery, i.e., the beginnings of a grading system for the "an H100-hour is not an H100-hour" problem. - Status: announced only — no named capacity providers, no executed first EFP found, everything subject to regulatory review. Compute Desk's index methodology: undisclosed, same as everyone else.
Why the EFP matters for the benchmark war: every index in Sections 1–3 is cash-settled against a measurement of a market that mostly trades privately. An EFP mechanism is the classic commodity-market answer — it stitches the paper price to physical delivery, letting basis arbitrage discipline the index. If ComputeConnect works, convergence pressure comes from deliverability rather than from methodology disclosure alone; historically (grains, metals, power), a credible physical-delivery bridge is what turned contested assessments into real benchmarks. It also makes Compute Desk a fourth index family with DCM-listing-driven disclosure obligations of its own.
Is price a function of lease rate, or is utilization a factor? Lease rate only, in all three cases. Silicon Data = quoted/posted lease rates (on-demand headline); Ornn = executed lease transactions; Kalshi = market-implied expectations of future lease rates. No provider adjusts price for utilization. Utilization enters only as (a) an indirect source-weighting consideration at Silicon Data and (b) a well-known criticism: rental price indices can stay firm while fleet utilization (the thing that actually determines operator cash flow and collateral value) deteriorates — the two series can and do decouple. If a lender-driven utilization benchmark emerges, it will likely come from the invoice/telemetry side (Ornn's pipeline, or InfraSight-style software) rather than from price scraping.
Is it geographically weighted, by utilization or dollar consumption? No provider publishes a geographic weighting scheme. Silicon Data treats geography as a normalization variable and a (secret) weighting factor across 40–50 countries; Ornn is US-flagship with city-level ambitions and undisclosed roll-up; Kalshi punts entirely to market convergence. Nobody has published whether regional weights reflect capacity, consumption dollars, or observation counts. This is a genuine open design question the CFTC filings should partially answer, because Appendix C requires demonstrating the index reflects "market values and conditions" in the relevant cash market.
What feeds each group's transactional data? Lender/lessor partnerships? - Silicon Data: no transactional feed per se — 150k daily observed price records from 50–100 platforms, contributor list proprietary, no named partnerships. The Compute Exchange leadership overlap is the latent transaction channel. - Ornn: the only true transactional feed — 10+ contracted contributors (Hydra Host named; InfraSight partner) plus own-marketplace prints, invoice-verified. No named lender partnerships yet, though its capacity-finance ambitions target exactly that. - Kalshi: its own order flow, with Ornn's transaction index at settlement. - Compute Desk: methodology and sourcing entirely undisclosed; its indices reference the ComputeConnect physical-delivery network, which could eventually make it the only index fed by EFP-linked physical prints. - Nobody has announced a data partnership with a GPU lender or lessor (the CoreWeave-style secured-debt underwriters sitting on the richest term-contract dataset in the market: $14B+ of GPU-backed debt is underwritten on bilateral take-or-pay cash flows invisible to every index). That dataset is the biggest unclaimed prize in this race.
The structural criticism all three face: the observable spot/on-demand market these indices measure is a thin residual. ~77% of CoreWeave's revenue is two customers on long-term take-or-pay; roughly three-quarters of supply sits with a handful of hyperscalers; large deals are private, bundled (compute + storage + networking + power), and discounted. An H100-hour is also not a standard unit — interconnect, cluster scale, SLA, interruptibility, and region create ~7x price dispersion for the "same" SKU in the same week, and even delivered performance on identical hardware varies up to ~38%. Every methodology above is an attempt to manufacture a standard grade where the physical market hasn't yet agreed on one.
Index transparency and replicability has surfaced as a limitation in nearly every piece of our compute work — the IOSCO assessment could grade only what is published, the lender validation program had to be designed to work without administrator cooperation, and every settlement analysis carries the caveat that no calculation is public. This piece is the baseline: what is actually knowable about each methodology as of August 1, 2026, so that when the CFTC filings land, what changed — and what still has not been disclosed — is measurable against a fixed record.
This is the crux: today's opacity is not an equilibrium. Two mechanisms force disclosure, on a knowable timeline.
CME and ICE will list via Part 40 self-certification against DCM Core Principle 3 (contracts "not readily susceptible to manipulation"). Appendix C to Part 38 spells out what a cash-settled contract's underlying index must demonstrate: - The settlement price must be "reliable, acceptable, publicly available, and reported in a timely manner," and shown to be "a reliable indicator of market values" — i.e., the DCM must document cash-market depth, transaction volumes, and participant counts. For GPU compute, that means quantifying exactly the thin-spot-market problem above, on the record. - Where a third party computes the index, the DCM must verify the provider "utilizes business practices that minimize the opportunity or incentive to manipulate" — Appendix C's own examples: lock-downs, employee derivatives-trading prohibitions, public dissemination of source names — and enter an information-sharing agreement with the provider. - Neither the CME/Silicon Data nor ICE/Ornn certification had been filed as of Aug 1, 2026 ("pending regulatory review" in both releases). The filings themselves will be the single largest transparency event to date in this market — they will put provider counts, data-basis (transacted vs. quoted), and anti-manipulation practices into the public record. - Cautionary precedent: the 2017 bitcoin futures self-certifications on thin spot indices went through but drew lasting criticism — the CFTC has institutional memory here, and GPU compute's cash market is arguably thinner and more concentrated than 2017 bitcoin.
Every analogous nascent commodity benchmark followed the same sequence once exchange settlement use arrived, and it's a reliable template for what Silicon Data and Ornn must ship:
| Precedent | What exchange adoption forced |
|---|---|
| US gas/power indices (post-Enron) | FERC 2003 policy: contributor codes of conduct, error-correction policy, per-print volume/transaction-count disclosure, confidence tiers |
| Baltic Exchange (freight/FFAs) | Public versioned "Guide to Market Benchmarks," panellist obligations and audit, public IOSCO compliance statement |
| Fastmarkets lithium/cobalt → CME | Annual independent IOSCO assurance reviews (BDO), separate benchmark-administration legal entity, published methodologies with formal consultations |
| UxC uranium → NYMEX | Public rulebook chapter naming the price indicator, documented administrator arrangement in a thin bilateral market |
Generalized checklist — none of which any GPU index provider has published yet: (1) public versioned methodology/rulebook, (2) contributor code of conduct and named-source or count disclosure, (3) independent annual IOSCO assurance audit, (4) oversight committee / separate administration entity, (5) restatement and error-correction policy, (6) exchange information-sharing agreements.