Between October 2025 and April 2026, a market that did not exist became real. Six exchanges, index providers, and auction designers are building regulated or crypto‑native venues for derivatives on GPU compute:

• Architect Financial Technologies — founded by former FTX US president Brett Harrison; regulated perpetual futures on GPU rental and RAM prices, launch imminent pending regulatory approval. Partnered with Ornn as price‑index provider.
• Ornn Compute Exchange — $5.7M October 2025 seed round; positioning as the first regulated derivatives exchange for compute. Publishes its own GPU price index, added to the Bloomberg Terminal on 2026‑04‑02. Vertically integrated: index and exchange.
• Helix — crypto‑native perpetual‑futures DEX. Launched the first‑ever NVIDIA H100 perpetual futures market.
• Compute Exchange — backed by DRW; institutional derivatives venue in development. Projects a potential $5T compute‑derivatives market at maturity.
• OneChronos × Paul Milgrom — combinatorial auction market for GPU compute, designed by the Nobel‑laureate auction theorist.
• SF Compute — cluster‑time auctions and forward sales. $300M valuation, $40M Series A late 2025.

Every one of these venues has the same structural dependency: a reference price to settle cash contracts against or to fund perpetual positions against. The index layer is the missing piece of market infrastructure for compute, and it is where value has historically accrued in commodity derivatives.

Commodity futures markets converged on a pattern over 150 years. CBOT wheat futures settle against USDA spot reports. Brent crude settles against Platts' Dated Brent assessment. Fed Funds futures settle against the effective federal funds rate published by the New York Fed. Every serious commodity derivatives market is settled against a benchmark that is:

• Independent of the exchange. A venue that settles its own contracts against its own index has an unhedgeable conflict of interest. Institutional counterparties demand third‑party settlement.
• Auditable. Counterparties with capital at risk must be able to verify the settlement price against reproducible data and methodology. Opaque indices survive only where counterparties accept provider risk.
• Low‑friction. Exchanges route around licensing and royalty costs when possible. A fee‑free index with an open methodology has structural adoption advantages over a paywalled one.

The current GPU compute index landscape fails at least two of these criteria. Ornn's index is proprietary, fee‑gated on the Bloomberg Terminal, and operated by an entity that is also launching the exchange that would settle against it. Survey‑based indices (SemiAnalysis, Silicon Data) are respected but not daily‑published, not reproducible from public inputs, and not operated as exchange‑ready settlement references. The open‑index slot is unfilled.

H.GPU is the daily statistical release of listed GPU rental rates published at 00:30 UTC by gpumarkets.dev. Twelve rental venues, thirteen chip‑tenor series across spot, on‑demand, and reserved tenors. The estimator is a trimmed median with ±3.0 median‑absolute‑deviation outlier rejection. Every property that matters for settlement is in place by construction:

• Data under Creative Commons Attribution 4.0. Any exchange can settle contracts against H.GPU without licensing cost or royalty. Attribution preserves the reference brand without fee friction.
• Estimator under Apache 2.0. Open source at github.com/gpu-markets/gpumarkets. Any counterparty can self‑host the estimator for zero‑trust verification of any published fix. No settlement ever has to be "trust us."
• Independent. Not operated by any exchange, venue, or neocloud. No revenue from settlement. No vertical integration. No conflict.
• Auditable methodology. Every outlier rejection is logged publicly on /news. Methodology changes ship through /changelog. The exact observation set on any given fix day is reproducible from the public data repository.
• Daily cadence, reliable timing. 00:30 UTC fix, compatible with standard end‑of‑day cash‑settlement conventions used in commodity markets.

Contract expires on a calendar date; settlement price is the H.GPU fix published at 00:30 UTC on the expiry date for the underlying chip‑tenor series. The exchange clears cash P&L at that price and the contract terminates — identical in mechanics to how oil futures on NYMEX settle against Platts Dated Brent on the contract's last trading day.

Architect Financial Technologies' planned perpetual futures on GPU rental and RAM prices map directly onto this pattern, as does Compute Exchange's institutional roadmap. The specific series a contract settles against is a market‑maker choice: an H100.SXM.R1Y contract settles against the H200 reserved one‑year series; a B200.SXM.SPOT contract settles against the B200 spot fix. The underlying is the series identifier; the settlement price is that day's published value.

A perpetual futures contract has no expiry date; instead, long and short positions exchange a funding payment each period, sized to pull the perp's mark price toward the underlying spot reference. Helix's NVIDIA H100 perpetual futures market and any future crypto‑native perp DEX can consume H.GPU as the underlying reference without on‑chain streaming infrastructure: a single daily published value is simpler to ingest than a tick‑by‑tick oracle.

Mark price on perp positions ticks against the day's H.GPU fix; the funding rate for the next period is computed from the mark‑versus‑index spread. The index is read‑only; the venue handles its own book and risk. Cost to the venue: one HTTP GET per day to gpumarkets.dev/data.

The spread between H.GPU (the listed‑side fix published here) and a transacted‑side contract index (SemiAnalysis's H100 1‑year rental index is the reference example) is itself a tradeable signal. As of mid‑April 2026, that spread has detached by more than at any point in the market's short history — listed B200 SXM spot down 28% over six months, contract‑side H100 1‑year up roughly 38% over the same window, on‑demand capacity described as "sold out across all GPU types."

A venue can list a basis contract whose payoff is H.GPU − SemiAnalysis.1Y (or the equivalent against Ornn's OCPI once a legal framework exists for cross‑index settlement). Traders take views on whether price discovery returns to the listed surface, stays in contracts, or widens further. The B200 decomposition note (Apr 11) lays out the three hypotheses — supply, exhaustion, absorption — each of which implies a different path for the spread.

GPU Markets is not a commercial data vendor. There is no license to sign, no royalty, no rate card. The data and estimator are published under Creative Commons Attribution 4.0 and Apache 2.0 respectively, and will remain so. The only asks for integration as a settlement oracle or reference price are attribution on public materials and a reciprocal link to gpumarkets.dev.

Derivatives venues, prediction markets, structured‑product issuers, or anyone consuming H.GPU as a reference are invited to coordinate directly with the maintainer: email john@gpumarkets.dev. Long‑form specifications for the estimator and the venue pipeline live at /methodology; complementary indices on the transacted side are catalogued at /references.

This is commentary on market infrastructure rather than price analysis. It carries no view on the correct mark for any chip‑tenor series at any fix date. Venue, valuation, and regulatory‑status references above are directional summaries of publicly reported facts as of 2026‑04‑20; a researcher integrating H.GPU into a specific venue's settlement pipeline should verify each against the primary filings.

Not investment advice.

← Back to Research