Renting an Nvidia H100 for an hour costs whatever your supplier says it costs. There is no published price. Two companies buying identical capacity can pay very different rates, and neither of them will know.

That changes on 5 October. CME Group, the exchange that lists crude oil and corn, will start trading two futures contracts tied to the hourly rental price of Nvidia chips, subject to regulatory review.

Each contract represents one month’s rent for a single GPU, CNBC reported. One tracks the H100, the chip most AI systems run on today. The other tracks the Blackwell B200 that succeeds it.

It is a small product with a large implication. The input every AI company depends on is about to get a public reference price, and a forward curve showing what traders think that price will be next year.

What is actually being traded

Nobody takes delivery of a graphics card. Both contracts are cash-settled against indexes published by Silicon Data, a New York firm that has spent two years tracking what GPU capacity actually rents for.

They will be listed on NYMEX, the CME’s energy exchange. That detail is not incidental. The executive quoted in the launch announcement is Pete Keavey, the CME’s global head of energy and environmental products.

He made the comparison explicit. Oil fuelled the 20th century economy and grew from spot trading into a global derivatives market, he said, and these contracts “turn compute into a standardised, tradable commodity”.

The buyers and sellers are easy to picture. A data centre operator owns servers and earns rental income, so it can sell futures to lock in revenue. An AI developer pays those rents, so it can buy futures to cap its costs.

Wall Street built the financing layer first

This is the second piece of machinery to arrive in a fortnight. The first was Nvidia recruiting six of the largest names in finance for a $500bn funding package aimed at the AI buildout.

Silicon Data chief executive Carmen Li drew the distinction herself, speaking to Bloomberg. The Nvidia announcement is a financing layer. This one is a risk management layer. A market of that size, she argued, cannot function without somewhere to hedge and discover prices.

That is a fair description of what the last month has looked like from the outside. Enormous sums have been committed to compute with almost no public information about what compute costs.

Lenders have already been improvising around the gap. Lambda sold a $917m leveraged loan backed by chips whose residual value nobody could independently price. OpenAI went hiring a power-trading lead because electricity, at least, already has a market.

Gavin Baker of Atreides Management, who led the funding round announced the same day, put it in farming terms. Futures markets let farmers finance next season’s seed and equipment rather than guess, he said. “You can’t build against a price you can’t see or lock in.”

The company selling the price just raised $30.5m

Silicon Data announced an initial closing of $30.5m on the same day, led by the Valor Atreides AI Fund. It raised $4.7m in March 2025, so the round is roughly six times the size of its seed 17 months later.

The money funds four things: the pricing benchmarks, an institutional data business, risk infrastructure for derivatives and credit, and a performance product called SiliconMark. Li told Bloomberg the ambition is to be “the independent referee” for the compute stack.

SiliconMark is the one worth understanding, because it exists to fix a problem that could break the futures contracts. Two clusters built from identical chips do not deliver identical output. Networking, topology and configuration change what you actually get.

That is a problem for anyone hedging. If the index tracks a standard hour of H100 time and your cluster underperforms it, the hedge stops matching the exposure. Silicon Data says normalising for performance also opens the door to physical delivery later.

Almost everyone in this market owns a piece of the referee

The investor list deserves reading carefully, and it is public. CME Group itself invested. So did DRW, Jump, Wintermute and Tectonic, all trading firms. So did VanEck, F-Prime, Samsung and Further.

In other words, the exchange holds equity in the company whose index its contracts will settle against, and several likely participants hold equity in it too. Li said as much on Bloomberg, describing those investors as heavily her clients at the same time.

This is not unusual in commodity benchmarks, where the firms that need a price often fund the people who publish it. It is worth stating plainly all the same. A benchmark is only as good as its independence, and the market it serves is currently very small.

Regulators have views on this in other markets. Europe has policed financial benchmarks under dedicated rules since the Libor scandal, precisely because a reference price shapes contracts far beyond the people who set it. Compute pricing has not attracted that attention yet.

What the curve will show before the earnings do

The most useful thing Li said had nothing to do with her product. Asked what would tell her the AI buildout had tipped into oversupply, she named three indicators, and all three become visible once this market exists.

The first is spot prices. A sustained fall means demand is weakening or supply is outrunning it. The second is the shape of the forward curve. It currently slopes upward, which means buyers are paying a premium to lock capacity in for longer.

The third is residual value. If used servers start fetching less on the secondary market, that is the market marking down its expectations of what those machines will earn.

Her own reading of the data is not bearish. A100 and H100 rental prices have risen about 20% since January, she said, and have been broadly flat for the past 20 days. Older chips keep earning because smaller models still run on them.

Depreciation, she argued, is not the same as worthlessness. Ships and aircraft depreciate too, and their residual value is still a function of the cash they can generate.

Why this matters in Europe

European AI companies have spent this year raising against compute they have contracted but not yet earned from. Nscale is preparing a US listing that values it at $51bn, with most of that revenue still ahead of it.

A published forward curve changes how those numbers get argued about. Contracted compute revenue can be checked against what the market expects compute to be worth when the contract runs.

The same applies to the fight over building the capacity at all. More than 500 US jurisdictions now restrict or ban data centres, and the economic case for each one rests on assumptions about future compute prices that nobody could previously see.

The caveats are real

None of this works without liquidity. A futures contract with few participants produces a price that is easy to move and hard to trust, and this one has to bootstrap a market from nothing.

The launch is also still conditional. Both contracts are pending regulatory review, and the CME and Silicon Data first announced the partnership on 12 May, nearly five months before the date they have now set.

And a public price cuts both ways for the people who wanted one. Compute buyers get to stop negotiating blind. Compute sellers, including several companies whose valuations assume rental rates hold, get a number the market can disagree with in real time.

Oil got its benchmark in 1983, more than a century after the first well. Compute is getting one about three years into its boom. That is either a sign of a market maturing unusually fast, or of how much money is now betting on a price nobody could previously check.

Get the TNW newsletter

Get the most important tech news in your inbox each week.