Compute gets a price. Which one isn't settled yet.

On 5 October, NYMEX lists the first compute futures: two contracts on the cost of renting a GPU. One references Nvidia H100 rental rates, the other B200. CME files them under Energy, alongside power and natural gas, and the product page calls compute "the new currency for the AI economy."

That framing is not marketing. It is a claim about what kind of thing compute is, and the contract design follows from it.

This matters beyond the people who will trade it. Until now there has been no observable forward price for compute at all. Once one exists it gets referenced, first in budgets and valuations and eventually in loan agreements. A benchmark's influence runs well past the people transacting on it, which is why how it is built is worth reading before it starts.

What a compute future actually is

A futures contract is an agreement to settle the difference between a price agreed today and the price that turns out to be true later. Nothing changes hands but cash. These are financially settled, so nobody delivers a GPU.

Each contract covers 730 GPU-hours, which is roughly one month of continuous rental for a single card. Prices are quoted in dollars and cents per GPU-hour, the minimum move is one cent, and because a cent applies across 730 hours, one tick is worth $7.30. Contracts are listed monthly for 36 consecutive months, so the curve runs three years out from day one.

At the H100 rate Silicon Data currently publishes for neo-cloud capacity, $2.57 an hour, a single contract is worth about $1,876. A small unit, which suits a market expected to start thin.

Two details in the specification are worth noting because they shape how the price forms.

The block trade minimum is five contracts, around $9,400 of notional. Block trades are negotiated privately and reported after the fact, here within fifteen minutes. A threshold that low means almost any trade of real size can be done off the screen, so the visible price may represent a small share of activity in the early months.

And exchange fees run $5.50 to trade and $1.35 to settle for a non-member, which against a $1,876 contract is about 0.37% round trip. For a hedging instrument that is a meaningful frictional cost, and it will matter to whether smaller operators actually use it.

There are two H100 prices, and they are 2.8x apart

Silicon Data publishes the indices these contracts reference. It publishes two prices for the H100.

Neo-cloud capacity prints at $2.57 per GPU-hour. Hyperscaler on-demand is $7.17. Same chip, same hour, nearly three times the price. A100 shows the same pattern, $1.57 against $3.75.

There is a straightforward reason. Neo-cloud and marketplace capacity prices below hyperscaler on-demand rates, which is Silicon Data's own explanation, and it reflects genuinely different products: different contract terms, different support, different reliability guarantees. Two prices for two markets is honest reporting, not an inconsistency.

It becomes a live question once one of them settles a futures contract.

The Special Executive Report does not say which index is used. It names the products, the codes, the size, the tick and the settlement type, and it points to rulebook chapters that were not publicly available when this was written. So the single fact that determines who can actually hedge with this instrument is, two weeks out, not something a prospective user can look up.

The consequence is concrete. A lender financing a facility leased to a hyperscaler, hedging with a contract that settles against neo-cloud rates, is hedged against a different market's price. The hedge would still do something. It would not do the thing it appears to do.

What the index discloses, and what it does not

A financially settled contract has no underlying to deliver. The index is the contract. So the methodology carries the weight that a warehouse and a delivery grade carry in a physical commodity.

Silicon Data discloses a good deal. Around 150,000 verified pricing records daily, across 50 to 100 platforms spanning hyperscalers, neoclouds and marketplaces, in 40 to 50 countries, with history back to September 2024. It covers on-demand, spot and reserved capacity on terms from one month to sixty, and normalises heterogeneous machine configurations to a standard unit. That is a serious data operation, and the transparency it brings to a famously opaque market is the point of the exercise.

Four things are not disclosed publicly, and each maps onto a question that has mattered in other benchmarks.

Whether the inputs are executed transactions or advertised rates. This is the question LIBOR failed: its submissions were estimates rather than trades. Silicon Data's material refers to verified pricing records and, in one description, to private transaction data, so there is clearly transaction content. The proportion is not stated.

The contributor list, which is proprietary. That means the index cannot be independently replicated.

The outlier rule and the aggregation method, meaning whether a daily value is a mean, a median or a weighted average, and how extremes are handled.

None of this is unusual. Commercial benchmark providers routinely protect methodology, and publishing a contributor list can itself distort behaviour. But a listed contract settling against a number that outside parties cannot reproduce is a design choice with consequences, and it is reasonable to want the specifics before rather than after.

Who owns the index

Silicon Data was founded in 2024 by Carmen Li, previously a data executive at Bloomberg. It raised $4.7m in seed funding in March 2025 from DRW and Jump Trading Group, then a $30.5m Series A led by the Valor Atreides AI Fund, with participation from CME Group, DRW, Jump, Tectonic, VanEck, F-Prime, Samsung, Further and Wintermute.

So the exchange listing the contract is an investor in the index provider the contract settles against, alongside several proprietary trading firms.

This is worth stating plainly and it is not an accusation. Benchmark infrastructure is frequently built this way, because the firms with the data and the incentive to build it are the firms that trade. CME taking a stake is arguably a signal of commitment rather than a conflict. All of it is disclosed in the companies' own announcements.

The thing a reader should carry while looking at the index is simpler than a conflict: the people who built the price, the venue that lists the contract, and several of the likely early participants are not independent of one another.

The curve mismatch

We wrote earlier this month about GPU residual values, and the argument there was that a card has several prices that behave differently. List prices barely move, and rental rates drift down slowly and fairly predictably. Resale is the chaotic one. That matters because resale is what a lender recovers if a loan goes wrong.

These contracts price the rental rate.

That is the right variable for the people the product is aimed at. An operator buying compute, or a provider selling it, has rental rate exposure and can now hedge it. A real gap is being filled, and capacity planning should get easier for both sides.

But it is a different variable from the one the credit stack rests on. And the mechanism that connects them is lending behaviour: when a borrower can demonstrate a hedge, a lender will generally lend more against the same asset. If the hedge references rental rates while recovery depends on resale values, borrowing capacity expands against a risk that has not actually been transferred.

Each individual decision there is rational. A lender is right that a hedged borrower is safer, and a borrower is right to hedge whatever they can. What the sum produces is more leverage resting on an assumption that the two curves stay close, which is exactly the assumption nobody is currently measuring.

The precedent cuts both ways

The closest analogue is ABX.HE, launched in January 2006 to make subprime exposure tradable without holding the underlying securities.

It did two opposite things. One was to extend the boom, by creating demand for tranches that were otherwise hard to place. And it accelerated the reckoning, because for the first time there was an observable price on subprime risk. The BBB− tranche broke in February 2007, well before equities noticed. Before that index existed, nobody could see the price of that risk at all.

That second effect is the case for compute futures, and it is a strong one. A published forward curve for compute is information the market does not currently have, and information is usually worth more than the discomfort it causes. The 2006 comparison is not a warning that this should not exist. What it shows is that a new benchmark changes behaviour in both directions, and that the direction that shows up first is usually the comfortable one.

What we take from it

We think the arrival of a tradeable compute price is a good development, and the interesting risk is not manipulation. Manipulation is hard here: the contracts settle against an index rather than a deliverable, the largest buyers of compute want prices low rather than high, and benchmark enforcement has become expensive enough to deter it.

The thing worth watching is quieter. It is whether lenders begin referencing this index in advance rates. That does not appear in press releases. It appears in loan agreements, and usually a year later.

What would change our reading is narrow and checkable. If the settlement index turns out to be predominantly transaction-based, with a published contributor list and settlement against an average of the contract month rather than a single day, most of the construction questions dissolve and this becomes straightforwardly good infrastructure. Those are answerable facts, not opinions, and they should be answerable on day one.

Four things to check on 5 October

Which index settles it — neo-cloud, hyperscaler, or a blend. Given a 2.8x gap this is the first question, and it determines who can genuinely hedge.

Transactions or advertised rates in the settlement calculation.

Average or single day at final settlement. An average across the contract month makes an expiry squeeze considerably harder, because each day's value is fixed once published.

Open interest after four weeks. Below a few thousand contracts this is a published quotation rather than a market, and a thin benchmark that gets referenced anyway is a worse outcome than no benchmark at all.

Contract details here come from CME's Special Executive Report SER-9785 and index levels from Silicon Data's published pages, read on 21 September. Two things could not be established from public sources and are stated above as open questions rather than filled in.