Compute futures

Renting an H100 for an hour has become a standardized unit, bought and sold at a price that moves. That is the precondition for a futures market, and one is being built: CME Group and Silicon Data plan to list H100 and B200 rental-index futures on October 5, 2026, subject to regulatory review.

What was announced#

CME Group and Silicon Data announced plans to list futures on GPU rental indices, starting with H100 and B200, on NYMEX. The stated launch date is October 5, 2026, pending regulatory review. The contracts are not live, and nothing on GPUQuant asserts that they are trading: GPUQuant carries no exchange feed that would show it.

Read the CME Group announcement

What one contract represents#

Approximately one month of renting one GPU: the same 730-hour unit GPUQuant quotes as a GPU-month. The contracts track an hourly rental index rather than the price of the hardware, so a contract is a claim on the cost of using a GPU for a period, not on the GPU.

The GPU-month exists for the same reason a barrel or a megawatt-hour does: a contract needs a unit large enough to be worth trading and standard enough that two parties mean the same thing by it. 730 hours is one month of continuous access to one GPU.

What is settled#

These are cash-settled benchmark derivatives. Nobody takes delivery of a GPU. Each contract settles against a Silicon Data rental index: SDH100RT for H100 and SDB200RT for B200. That is why the index definition matters more than any single provider quote, and why Silicon Data indices has its own page.

The two indices are not the same kind of number. The H100 index is a neo-cloud segment reading; the B200 index blends neo-cloud, hyperscaler, colocation and private-market observations into one figure. A comparison that holds for one does not automatically hold for the other.

Why a rental index, not a hardware price#

You cannot deliver an H100 into a futures contract in any useful way. Chips are allocated, not traded on an open spot market, and a physical unit would be unsettleable.

What is tradeable is access: an hour of a GPU, which thousands of providers sell at published prices in a form that is close to identical. Index that hour and you have a settleable reference, which is exactly the structure used for electricity, freight and bandwidth before it.

Who would use them#

• A buyer of compute wanting to fix a cost months ahead.
• A supplier of compute wanting to lock in revenue on capacity it will have.
• A trader taking a view on whether AI compute gets cheaper or scarcer.

List price, index and futures price#

The cloud list price

What a provider currently publishes for one GPU for one hour, on demand, with no commitment. It is observable and verifiable against the provider, and it is what GPUQuant's reference price and neocloud index measure. It is a list price, not a transacted price: large buyers negotiate below it and committed-use discounts sit well below it.

The settlement index

Silicon Data's daily reading of the realized, standardized rental price, modelled from transaction data GPUQuant cannot see. It is the number the contracts settle against, and it is displayed on the market-data page under licence.

The futures price

The price agreed today for a contract month in the future. It embeds expectations about supply, demand and the pace of new hardware, and it can sit far above or below today's cash price. Until the contracts trade, the closest thing to it is Silicon Data's forward curve. GPUQuant publishes no futures price and estimates none.

The two contracts#

Beside each contract, the two GPUQuant figures for the same GPU-hour, kept apart: the reference price from Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure, and the neocloud index from the specialist GPU clouds that publish a quotable rate. Neither is the settlement index.

These are the only two GPUs GPUQuant publishes, because they are the only two with an announced contract. Every provider behind both figures is on the market-data page.

GPUQuant's role#

None in the market itself. GPUQuant publishes the cash market underneath the contracts: what Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure publicly charge for the same GPU, month by month and region by region, and beside it what 16 specialist GPU clouds publish. The two are kept as separate figures and never averaged together.

The contracts are not tradable here. When they list, orders will go to a connected futures broker that keeps the account, margin and clearing, and the H100 and B200 rows on Tradable Compute will carry the market data. GPUQuant is not affiliated with CME Group or Silicon Data, and nothing here is investment advice.