What are GPU futures?

Cash-settled derivative contracts tied to published GPU rental price indices, letting buyers and sellers lock in future compute costs without exchanging hardware. CME's 2026 contracts reference the Silicon Data H100 Rental Index; ICE's contracts, with index partner Ornn, are planned to reference the OCPI series covering H100, H200, B200, and RTX 5090.

Why did CME and ICE launch GPU futures in 2026?

Because GPU rental prices became volatile enough to need hedging and standardized enough to be indexed. H100 1-year contract rates swung from $1.70/hr (Oct 2025) to $2.35/hr (Mar 2026), and Blackwell spot prices surged 48% between February and April 2026. Two major exchanges moving within days of each other signaled institutional conviction that compute is now a commodity.

What is the Silicon Data H100 index?

SDH100RT is a daily benchmark of like-for-like H100 rental rates in USD per GPU-hour, drawn from observations across neocloud providers, hyperscalers, colocation markets, and private rental platforms. It sat near $2.53/GPU-hour in mid-2026 and is the reference index for CME's compute futures.

How does a public GPU price index help compute buyers?

Three ways: negotiation leverage (quotes can be benchmarked against a published index, where hyperscaler medians run ~3.3x non-hyperscaler medians), budget certainty (futures lock in training-run costs months ahead), and financing (lenders can underwrite GPU fleets against a hedgeable revenue curve rather than guesses).

What is Compute Exchange?

A GPU marketplace running auctions for reserved capacity across providers, whose published results became de facto price discovery: H100 reserved listings spanned $1.07-10.14/hr (median $5.52), with marketplace medians at $1.70 vs. $6.11 at hyperscalers. In July 2026 it added a secondary market for used H100s/A100s, pricing used units at $6,000-22,000 versus $25,000-40,000 new.

GPU Futures Arrive: CME and ICE Make Compute a Tradable Commodity

CME (with Silicon Data) and ICE (with Ornn) launched GPU compute futures in 2026. A public H100 forward curve changes how AI compute is bought and hedged.

Within days of each other in May 2026, the world’s two largest derivatives exchange groups moved on the same asset class: CME Group announced compute futures referencing the Silicon Data H100 Rental Index, and ICE — owner of the NYSE — announced cash-settled GPU compute futures with index partner Ornn. Compute, the input everyone in AI budgets for and no one could hedge, is becoming a listed commodity. The mechanics matter less than the consequence: for the first time, the price of an H100-hour will have a public forward curve.

What was announced

	CME Group	ICE
Index partner	Silicon Data	Ornn
Reference index	H100 Rental Index (SDH100RT), published daily in $/GPU-hour	OCPI series
Coverage	H100 (like-for-like rate across neoclouds, hyperscalers, colo, private platforms)	H100, H200, B200, RTX 5090; more GPU types to follow
Settlement	Cash	Cash
Announced	12 May 2026	May 2026

Per the ICE press release and coverage of CME’s launch.

Cash settlement is the pragmatic choice: nobody delivers a rack of GPUs against an expiring contract. Instead, contracts settle against the index — which makes the index itself the critical infrastructure. Silicon Data’s SDH100RT publishes a standardized H100 hourly rate daily (~$2.53/GPU-hour in mid-2026), constructed from observations across provider types.

Why now: volatility met standardization

Futures markets emerge where two conditions coincide — price volatility worth hedging, and a fungible underlying worth indexing. Both arrived in 2025-26:

Volatility: H100 1-year contract rates fell to $1.70/hr in October 2025, then rose almost 40% to $2.35/hr by March 2026, per SemiAnalysis’s rental index; spot jumped 10% in a single four-week window at year-end. Blackwell spot rental prices surged 48% between February and April 2026. These are commodity-grade swings on billion-dollar budget lines.
Standardization: an H100-hour is now a substantially fungible unit. Marketplaces made that visible — Compute Exchange’s March 2026 auction data showed reserved H100 listings from $1.07 to $10.14/hr (median $5.52), with a marketplace median of $1.70 against a hyperscaler median of $6.11. In July 2026 it extended price discovery to hardware itself, opening a secondary market where used H100s trade at $6,000-22,000 versus $25,000-40,000 new.

The 3.3x spread between hyperscaler and non-hyperscaler medians for identical silicon is precisely the kind of opacity that public indices erode.

What a public H100 curve changes for buyers

Procurement gains a benchmark. Every GPU quote — cloud contract, neocloud reservation, colocation-plus-owned-fleet — can now be marked against a published index, the way freight is marked against the Baltic indices. Expect index-linked pricing clauses (“SDH100RT minus X%”) in larger compute contracts, and expect the widest-spread sellers to face the most repricing pressure.

Budgets become hedgeable. An AI lab planning a Q2 2027 training run can lock compute costs today; a neocloud carrying merchant exposure on 20,000 GPUs can sell forward and sleep. This is the grain-elevator logic of 1877 applied to FLOPS, and it converts compute from a spot-market gamble into a plannable input.

Financing gets cheaper. The binding constraint on neocloud growth has been lender skepticism about GPU residual values and rental-rate risk. A liquid forward curve lets operators hedge the revenue line, which collapses the risk premium on debt secured by GPU fleets — likely the single largest real-economy effect. Cheaper neocloud capital ultimately means more supply and lower rental prices; facility-side economics are tracked in our colocation index and SEA catalog.

Speculators arrive too. Financial participants will trade the curve without ever renting a GPU. That adds liquidity and occasionally adds noise; commodity history suggests the liquidity is worth the noise.

Caveats

Index construction is the weak point: rental “price” bundles interconnect, storage, support, and contract term, and like-for-like normalization is judgment-laden. Fast hardware cycles are the second risk — an H100 contract matters less each quarter that Blackwell and Rubin displace the installed base, which is why ICE’s multi-GPU OCPI coverage (H100 through RTX 5090, with more to follow) is the more future-proof design. And early open interest may be thin; benchmarks earn trust over years, not press releases.

The bigger picture

Power got markets, bandwidth got markets, freight got markets — each time, opacity premiums shrank and capacity planning improved. Compute is following the same path at unusual speed: from allocation-by-relationship in 2023, to marketplaces and daily indices in 2025, to listed futures in 2026. For anyone budgeting AI infrastructure, the practical takeaway is immediate: stop accepting quotes in a vacuum. The reference prices exist — we publish the operational ones across GPU rentals and colocation, with market series in stats — and the forward curve is now, literally, on exchange.