GPU rental price indexes compared: what each one counts

A GPU rental price index turns many hourly GPU prices into one number, and the published indexes start from different prices. Ornn's index counts prices paid under active on-demand rentals [1]. Silicon Data's settlement indexes combine verified transactions with verified executable listings from neocloud providers and GPU marketplaces [10]. SemiAnalysis surveys more than 100 market participants each month for a one-year H100 contract price [17]. AIMultiple and Price of Compute read the prices providers post [19, 20]. This article compares what each publisher's own documents, viewed on 7 November 2026, say about what its index counts, how it adjusts and weights prices, and who governs it.

Two of these indexes now underlie proposed futures contracts. ICE and Ornn announced futures on Ornn's index on 19 May 2026, and NYMEX filed contracts on Silicon Data's H100 and B200 indexes on 11 August 2026 [6, 24]. In August 2026 the Commodity Futures Trading Commission (CFTC) asked how a compute settlement price should be built and protected [29]. Its questions apply to any index a reader relies on.

What a GPU rental price index counts

Three kinds of price feed the published indexes. A posted rate is the price a provider lists on its pricing page or pricing API. An executed rental price is what a renter pays under a rental that has started. A surveyed contract price is a level that market participants report when asked.

The publishers name the difference themselves. Ornn describes its benchmarks as coming "from executed trades rather than listings or surveys" [2]. SemiAnalysis writes that "most of the GPU rental market is transacted on a long-term basis with contracts of at least 6mths and longer" [17]. Silicon Data writes that "posted prices show what buyers are quoted, and transactions show what capacity actually rents for" [14].

The note on evidence classes in GPU pricing sets out why a listing and a sale stay apart. Posted rates also move slowly: more than half of the spot price series Silicon Data tracks did not change once between 8 and 16 October 2026 [14]. The article on what posted GPU cloud prices tell you covers how posted prices move and what they leave out.

Putting prices on one basis

Ornn and SemiAnalysis state prices in US dollars per GPU-hour, and Silicon Data converts every record to "a standardized per-hour, per-GPU rate" [1, 8, 18]. Each reaches a common basis in its own way.

• Silicon Data maps each record to "an on-demand, U.S.-baseline reference configuration specific to the GPU model". Spot and reserved prices are normalized to their on-demand equivalent, and non-U.S. inputs are adjusted to the U.S. baseline [10].
• Ornn's methodology localizes each index per GPU type and records six regions [1].
• SemiAnalysis's contract index holds one term, a year [17].
• Price of Compute keeps pricing types apart: "Pricing types never mix" [20].
• AIMultiple takes medians per provider, GPU and billing tier [19].

Means, medians and missing data

The calculations differ as much as the inputs.

• Ornn computes "the volume-weighted winsorized mean of executed transaction prices for a given GPU type over a rolling one-hour window". The winsorization level, which caps extreme prices, is set by Ornn and not published [1].
• Silicon Data takes an arithmetic mean per provider, then a weighted sum across providers, with at least 5 providers each business day and no provider above 20% of the weight [10].
• SemiAnalysis's composite uses a weighted mean "with hyperscale clouds weighted more strongly than marketplaces and neoclouds", and a per-step move cap [18].
• AIMultiple and Price of Compute take the median of each provider's prices, then the median across providers [19, 20].

When data stops arriving, Ornn carries its last value forward for a limited number of windows and then suspends the index [1]. Silicon Data can use a provider's last validated price for up to 30 calendar days [10]. Price of Compute's rule is "Missing beats wrong" [20].

Governance as each publisher describes it

Silicon Data describes a Benchmark Oversight Committee with at least one independent member, whose presence a quorum requires. It also names an independent calculation agent and an attestation examination by an independent accounting firm [10].

Ornn's methodology, whose changelog lists only its first publication on 24 July 2026, said on 7 November 2026 that "Ornn is establishing an oversight function, constituted as a standing committee", with three to five members and an independent majority [1]. It adds that Ornn "is developing formal conflicts procedures" [1].

Both disclose commercial ties. Ornn's affiliates transact in physical GPU capacity, and Silicon Data's chief executive also leads a spot GPU procurement marketplace and is excluded from benchmark production [1, 10]. Silicon Data says its series are administered "in alignment with" the IOSCO Principles for Financial Benchmarks, and Ornn says OCPI is "designed with reference to" them [1, 10].

GPU rental indexes and compute futures

ICE Futures U.S. said on 29 September 2026 that its OCPI H100 and B200 futures await "completion of the request for comment period on compute futures" at the CFTC, and that "no listing date or timeline has been set" [7]. Ornn's page still showed the status "Pending CFTC comment period" on 7 November 2026 [3].

NYMEX filed Silicon Data H100 and B200 rental index futures with the CFTC on 11 August 2026, naming trade date 5 October 2026, subject to Commission approval [24]. On 21 September the CFTC extended its review to the end of 9 November 2026, citing "novel or complex issues" [25], and on 4 November NYMEX consented, at the CFTC's request, to a further extension to 8 January 2027 [35]. The CFTC's request for comment, published on 21 August with comments due by 20 October, asks whether it would be appropriate to permit trading in a contract settling to "a price computed from data that the Commission may not be able to observe, verify, or surveil" [29]. The Q3 2026 compute market review follows that timeline through September.

Questions to ask of any GPU price index

IOSCO's 2013 principles for benchmarks and the CFTC's Appendix C for cash-settled contracts give a reader neutral questions [31, 32]:

• What is the unit, and which prices are eligible?
• Where do the inputs sit on IOSCO's hierarchy, which runs from concluded arm's-length transactions to executable bids and offers to expert judgment [31]?
• What minimum data does a value need, and how concentrated can the inputs be?
• What happens when data is missing?
• Who oversees the index, and how are conflicts, changes and complaints handled?
• For settlement use, is the price series "reliable, acceptable, publicly available and timely" [32]?

The full report scores no index against them. It sets each publisher's answers side by side, in the publisher's own words.

Where the Rillor Compute Index fits

The Rillor Compute Index measures what GPU compute costs to rent and to own, from the prices cloud providers publish and the prices of GPU systems, under a published methodology. Its methodology v1.1, dated 7 November 2026, treats prices from official pricing APIs and published pricing pages as listings [34]. It takes each provider's median and then the median of provider medians, and it withholds a print that misses its data rules for publication [34]. Its series are for research and planning use and are not offered for settlement of any contract [34]. Reading GPU prices from cloud pricing APIs shows how posted rates are read and converted.

Read the full report

The full report compares the published methods line by line, with the record counts each document states, the contract terms, the CFTC's questions and the IOSCO and Appendix C questions beside each publisher's own words.