Compute starts trading.
On October 5, the hour gets a public price.
CME Group and Silicon Data intend to list compute futures on October 5, 2026, pending regulatory review — contracts on the hourly rental price of H100 and B200 capacity. ICE has announced its own with Ornn. A third venue followed nine days later.
The detail that tells you what is really happening: they list under NYMEX rules — the same rulebook as crude, natural gas and power. Compute is not being treated as a technology product. It is being treated as a commodity, and specifically as an energy-like one.
Whyte Consolidated Research · 2026-08-17 · 11 min read · On compute becoming tradable
Three venues, four months, one idea.
For years the idea of trading compute was a thought experiment. In the space of four months it became a launch calendar. Two of the largest derivatives exchanges in the world have each partnered with an index provider, a third venue has announced, and the CFTC is engaged.
H100 and B200 Rental Index Futures, listed under NYMEX rules. Each contract represents one month of GPU rental cost, cash-settled against a daily index spanning clouds, brokers and manufacturers.
USD-denominated, cash-settled GPU compute futures referencing OCPI series across H100, H200, B200 and RTX 5090. Asian-style settlement on the arithmetic average of daily index values.
Compute futures on a new US-based exchange, announced days after the ICE partnership.
Carmen Li, Silicon Data's chief executive, put the purpose plainly: compute futures give the market a public, tradable reference price for the resource every AI system runs on. CME's Pete Keavey framed the users as AI builders and hyperscalers managing compute cost and price risk. Both descriptions point at the same absence: until now there was no observable forward price for the single largest input cost in the industry.
Nothing is delivered. That is the breakthrough.
A GPU-hour is not fungible. An H100-hour in Virginia is not interchangeable with an H100-hour in Oregon on different interconnect at a different service tier, and neither is interchangeable with a B200-hour. Fungibility is the first requirement of a deliverable commodity contract, and compute fails it comprehensively.
So the contracts do not deliver anything. They settle in cash against a published index. You cannot standardise the physical thing — but you can standardise a price for it, and settle the difference. The hardware stays exactly where it is.
This is not a minor implementation detail. It is the entire reason the market can exist now when earlier attempts failed. The instructive precedent is bandwidth: around the turn of the millennium there was a serious effort to trade telecom capacity as a commodity, and it collapsed. The causes were non-fungibility, no credible reference price, no workable delivery mechanism, and thin liquidity around counterparties that did not survive. Index settlement is a direct answer to the first three.
Whoever publishes the price owns the market.
A cash-settled contract is only as good as the number it settles against. This is why both exchange announcements led with an index partner rather than a product spec — and why the index providers, not the exchanges, are the interesting businesses here.
The two approaches differ instructively. Silicon Data publishes daily rental price indices spanning H100, H200, A100, B200 and MI300X, gathered across clouds, brokers and manufacturers, and tracking hyperscaler and neocloud tiers separately. Ornn builds its Compute Price Index from printed transactions rather than surveys or quotes — a stricter standard, and a harder one to achieve at coverage.
Silicon Data states the credibility conditions directly, and they are worth holding onto, because they are the tests any future entrant will be measured against:
- →Methodological transparency — anyone settling against it can see how it is built.
- →Broad coverage across providers and hardware generations, not a single venue's book.
- →Daily cadence, so the curve has something current to reference.
- →Independence from the parties trading against it — the publisher must not be a counterparty.
That last condition is the one with teeth. A benchmark cannot be published by someone with a position in what it prices. It is the lesson every rate and commodity benchmark scandal of the past two decades taught the hard way, and it is encouraging that the compute market is starting from that position rather than arriving at it after an incident.
One contract, two opposite anxieties.
A futures market needs both a natural long and a natural short, and compute has an unusually clean pair of them.
The buyer is an AI company whose largest variable cost is GPU rental. It fears the price rising. Buying futures fixes its effective cost: if the index climbs, gains on the contract offset the higher bill. Given that one-year H100 rates moved roughly $1.70 to $2.65 per hour in five months, that is not a theoretical anxiety.
The seller is whoever owns the capacity. It fears the price falling, and it fears empty hours. As we set out in the unit economics of a rack, utilization is the one term an operator barely controls and an unsold hour is gone for good. Selling forward converts part of that exposure into a fixed number.
And here is why this matters more to an owner than to a trader. A hedged revenue stream is a financeable revenue stream. Silicon Data makes the point explicitly — providers selling against inventory convert an eroding rental stream into a fixed one, stabilising cash flow for debt underwriting. It is the same reason a power project hedges its output before it reaches a credit committee. The hedge is not speculation; it is the thing that makes the asset bankable.
It clears where power clears.
The CME contracts will be listed and subject to NYMEX rules — the New York Mercantile Exchange, home of crude, natural gas and power. That placement is not administrative housekeeping. It is a statement about what kind of commodity compute is.
Consider the properties. Compute cannot be stored — an hour that goes unsold is not inventory, it is gone. Its value is tied to a location, because power price, latency and grid position all differ by site. It is produced and consumed simultaneously.
Those are not the properties of oil, which can be stored, shipped and arbitraged across time. They are the properties of electricity. And electricity markets already solved this class of problem decades ago, with forward strips, index settlement and explicit locational basis. Compute is not inventing a market structure — it is inheriting the one built for power, which is why it can arrive this fully formed this quickly.
The index is not your price.
Cash settlement solves fungibility by refusing to engage with it — and that refusal leaves a residue. The contract settles on a standardised national number. You realise a price specific to your site, your hardware, your tier and your contract book. The gap between the two is basis, and it does not disappear because you hedged.
Silicon Data says so plainly: an application running custom configurations in a specific datacenter does not hedge cleanly with a standardised H100 contract. That is a notably candid thing for an index provider to publish, and it should be read as a specification rather than a warning. The instrument reduces your exposure to the market price. It does not reduce your exposure to your own circumstances.
- →Location — power price, land cost and grid position vary enormously between sites.
- →Service tier — on-demand, reserved, spot and interruptible capacity do not clear at one price.
- →Hardware mix — a standard H100 index says nothing about your B200s, your MI300Xs or your interconnect.
- →Contract structure — a portfolio already sold on multi-year take-or-pay has little spot exposure left to hedge.
None of that makes the hedge useless. Every commodity hedger in the world lives with basis; power traders have priced locational basis for decades and built an entire product category around it. It does mean the first question before hedging is not “what is the price?” but “how closely does this index actually track what I earn?”
The financial market never delivers anything.
A commodity market has four layers, and it is worth being clear-eyed about which are now spoken for. The venue and index layers are being taken by participants with clearing infrastructure, regulatory standing and distribution that will not be dislodged by a newcomer.
Claimed. Credibility compounds — liquidity concentrates on whichever reference the contracts already settle against.
Claimed. Clearing, margin, distribution and regulatory standing are not replicable quickly.
Open. The financial market never delivers anything. Somebody still has to supply real capacity, and that somebody owns megawatts.
Open. Bridging an index hedge to a specific site, and lending against hedged compute revenue.
The two open layers are the interesting ones, and they share a characteristic: they require owning something physical. A cash-settled contract never causes a single GPU-hour to be produced. Somebody still has to build the facility, secure the megawatts and run the machines — and that somebody is now able to sell their output forward into a market with a visible price, which is a materially better position than it was a year ago.
The same is true of basis. The listed contract deliberately ignores location and tier, which means every hedger with a specific site has a residual exposure nobody is currently packaging. In power markets that gap became an entire product category. There is no obvious reason compute will be different.
Meters, finances, trades.
Read the last few years as a sequence and it is remarkably orderly. Compute became something that meters — a measured output with a published rate. Then it became something that finances, as capital organised around the contracts attached to it. On October 5, subject to regulators, it becomes something that trades.
That is the standard route by which an industrial input turns into an asset class, and it usually takes decades. Oil took most of a century to get from wellhead to futures pit. Compute has done it in about three years, which should probably make everyone slightly more careful rather than slightly more excited.
For anyone who owns capacity, though, the implication is concrete and rather good. A visible forward price means your output can be valued by someone who has never visited the site. A hedge means the revenue can be fixed. And fixed revenue, as every power developer already knows, is the difference between an asset you can finance and one you merely own. The megawatts were always the scarce thing. Now there is a public price for what they produce.
Which puts the emphasis back where our thesis has always had it. A tradable price for compute makes the enclosure more valuable, not less — because the contract settles on a number, and only a building with power, cooling and a grid position can actually produce the hours behind it.
Compute futures — questions
- What is a compute future?
- A standardised contract to settle the difference between an agreed price for GPU rental capacity and where a published rental-price index actually prints on a future date. No GPUs change hands. CME Group and Silicon Data intend to list Silicon Data H100 Rental Index Futures and B200 Rental Index Futures on NYMEX from October 5, 2026, pending regulatory review, with each contract representing one month of GPU rental cost.
- Do compute futures involve delivering actual GPUs?
- No, and that is the point. The contracts are cash-settled against a reference index rather than physically delivered. Physical GPU-hours are not fungible — hardware generation, location, interconnect and service quality all differ — so they cannot be standardised into a deliverable contract. A published price for them can be standardised, which is what makes the market possible.
- Who publishes the index the contracts settle against?
- Two index providers have exchange partnerships. Silicon Data publishes daily GPU rental price indices covering H100, H200, A100, B200 and MI300X, drawn across clouds, brokers and manufacturers, and partners with CME. Ornn publishes the Ornn Compute Price Index, built from printed transactions rather than surveys or quotes, and partners with ICE. Both settle in cash; the ICE contracts are described as Asian-style, settling on the arithmetic average of daily index values.
- Why would a datacenter operator sell compute futures?
- To convert an uncertain revenue stream into a fixed one. Utilization is the variable an operator least controls, and unsold hours cannot be recovered. Selling forward against capacity fixes the price on some of that output, which stabilises cash flow — and stable, contracted cash flow is what lenders underwrite. It is the same logic as a generator hedging power output.
- Why would an AI company buy compute futures?
- For budget certainty on its largest variable input. A company expecting to consume a large volume of GPU-hours over the next year can fix its effective cost; if rental rates rise, gains on the futures offset the higher bill. The volatility justifies it — one-year H100 contract rates moved from roughly $1.70 to $2.65 per hour between October 2025 and March 2026, about 56% in five months.
- What is basis risk in compute hedging?
- The gap between the index a contract settles against and the price a particular operator or buyer actually realises. A standardised H100 index cannot reflect one site's location, power cost, interconnect, service tier or hardware mix. Silicon Data states the point directly: an application running custom configurations in a specific datacenter does not hedge cleanly with a standardised H100 contract. The hedge reduces exposure to the market price, not to your own circumstances.
- Why are compute futures listing on an energy exchange?
- Because compute behaves like electricity rather than oil. It cannot be stored, an unsold hour is gone rather than carried as inventory, and its value is tied to a location and a moment. Those are power-market characteristics, and power markets already have the instruments — forward strips, index settlement, locational basis. Listing under NYMEX rules puts compute alongside the commodities whose market design actually fits it.
Related Whyte Consolidated research on the datacenter buildout and the capital forming around it:
- Whyte Consolidated — The GPU is the asset
- Whyte Consolidated — Cooling is part of the computer now
- Whyte Consolidated — NVIDIA's $500 billion bet: turning compute into an asset class
- Whyte Consolidated — Own or rent: the split strategy inside the AI datacenter buildout
Contract details are taken from CME Group and Silicon Data announcements of May and August 2026, from the ICE and Ornn announcement of May 19, 2026, and from Silicon Data's published description of its indices. All of these products are stated to be subject to regulatory review or approval, and specifications may change before or after listing; nothing here should be relied on as a contract specification. Rental-rate figures are market observations as published, not quotes available to any particular buyer. Payoff and basis diagrams are illustrative schematics, not data. This article describes market structure for general information only. It is not investment, legal, tax or accounting advice, not a recommendation or solicitation to buy or sell any futures contract or other instrument, and not an offer of any product. Derivatives trading involves risk of loss, and hedging programmes have regulatory, accounting and margin implications — take professional advice before acting.