Compute has a price and no unit
It is sold by the hour of a named chip. That tells the buyer how long the machine ran, not what it did.
Electricity is sold by the megawatt-hour. Water is sold by the cubic metre. Both are measures of the thing delivered. Compute is sold by the chip-hour, which is a measure of how long a machine was rented. Two buyers who each pay for an hour of the same chip can receive very different amounts of work, depending on the memory, the network, the cooling and the software around it.
The price is real
There are daily price indices for chip-hours and forward curves built on them. An exchange has announced futures in which each contract “will represent a month’s worth of rent” for one chip, and says the contracts “will now turn compute into a standardized, tradable commodity”. The launch date was given as “pending regulatory review”.[1] We have not confirmed that they trade.
The unit is not
A consultancy that expects this market to grow also writes that “the same GPU on the same day can fetch significantly different prices across indices”.[2] That is what a market looks like before its unit is settled. Each index has to decide what counts as the same thing, and each decides differently.
Other markets have met this and solved it. Crude oil is not one substance. It trades because there are agreed grades and agreed tests. Electricity differs by place and hour. It trades because there are agreed nodes and agreed meters. Compute has benchmarks, but no agreed grade that a contract can point to.
Why this is an environmental question
Compute is the newest large load on the other three systems. Data centres used about 415 terawatt-hours in 2024, around 1.5% of the world’s electricity, and are expected to more than double by 2030.[3] The power, the cooling water and the emissions are all bought and reported per facility.
What nobody can yet state on common terms is the resource cost per unit of work. One developer has published its own figure: a median text prompt uses 0.24 watt-hours and 0.26 millilitres of water.[4] Others have published theirs, with different boundaries. Without a shared unit of work, those numbers cannot be compared, and a claim of efficient compute cannot be checked.
What a receipt would need
- Claim. This system delivered this much work.
- Unit. Work on a defined task, not hours on a named chip.
- Issuer. Someone other than the seller, using a test anyone can rerun.
- Time. When it was measured, since hardware and software both change the answer.
Then the small print: where it ran, on what power and with what water.
Until that exists, every price for compute contains a guess about performance, and every lender, buyer and regulator is making that guess separately.
Sources
- CME Group (2026). CME Group and Silicon Data to Launch Compute Futures on October 5.
- Boston Consulting Group (2026). The New Economics of AI Compute Markets.
- International Energy Agency (2025). Energy and AI: executive summary.
- Google Cloud (2025). Measuring the environmental impact of AI inference.