An AI service can publish 2 different electricity figures for the same workload without either being a mistake. In August 2025, Google put the median Gemini Apps text prompt at 0.10 watt-hours when counting active AI chips alone. Add the host computers, idle capacity and data-center overhead, and its estimate became 0.24 watt-hours. The difference helps explain why AI’s environmental cost remains so hard to judge. We need to know what the number includes before deciding what it means.
Google deserves credit for showing its work. That disclosure also makes the case for a common set of figures that every major provider should have to supply.
What the published numbers include
A spare machine waiting for a traffic spike is doing something useful. It keeps a service ready when demand jumps or another machine fails. It also uses power while waiting. If the count starts only when a chip begins to produce an answer, the cost of being ready is left out. It remains on the electricity bill.
Accountants deal with this kind of problem all the time. But a buyer can be misled when one company reports the narrower figure and another reports the broader one, then both ask buyers to judge their use of resources.
Jeff Dean and Amin Vahdat, who presented Google’s figures, have a strong answer to anyone who says the industry knows nothing about its footprint. They describe how to measure a service used at scale. Their estimate includes machines and overhead that a chip-only count misses. It is also a dated median for a particular service, with limits the company acknowledges. Google says no independent third party had checked the claims.
The public ought to reward that care. Yet even a report with clear limits can be used to compare things that do not match.
Mistral’s July 2025 disclosure, for example, used a 400-token response and a broader lifecycle approach that included upstream impacts. It left out the user’s device and noted gaps in what is known about the footprint of making graphics processors. The result cannot fairly be ranked against a Gemini prompt measured on different terms. A ratio between those water figures tells the buyer little when the companies have counted different things.
A business choosing an AI supplier should be able to ask:
- What job was measured?
- Over what period?
- Which equipment was counted?
Those questions should have clear answers. A buyer should not have to find out that a reassuring average came from short text responses when the planned use calls for long documents and repeated calls to a model.
The model itself can stay the same while the system around it changes. Grouping requests, keeping spare machines ready and meeting a strict time limit for responses are all part of the service a customer buys. A test can measure one part of that system for a good scientific reason. Trouble starts when that part is presented as the whole environmental cost.
Water and carbon need their own accounts
Power use tells us something worth knowing, but it cannot settle the water or carbon question by itself.
Google’s technical paper uses market-based electricity emissions, which account for the environmental attributes of the electricity a company buys. It also assigns part of the emissions from making its AI accelerators to their use. The carbon figure therefore includes some manufacturing emissions, though it falls short of a complete lifecycle account.
A location-based carbon account asks about the electricity generated in the relevant area. A market-based account reflects contractual purchases. Both can serve a purpose, but a buyer should be able to tell which one is used in the claim. Buying renewable electricity attributes does not make a machine stop drawing power. Reports should leave no room for that mistake.
Water brings another set of boundaries. A measure of water consumed for direct cooling leaves out water used elsewhere to produce electricity. Water withdrawn from a source is also different from water consumed. Consumption measures the portion not returned to that source within the accounting period. Comparing either figure with water use in a home does little to explain what was counted.
For someone trying to assess a local effect, the place and season matter. A total for the whole company does not tell a town how much water it must supply. Even a precise average may say little about the risk of shortages.
Arcep, which oversees France’s communications sector, argues that these reports should cover each stage of the lifecycle. They should also cover the various parts of the system and the harm each can cause. That is a sensible starting point. It leaves room to ask separate questions rather than forcing power, water and materials into a single number that looks reassuringly tidy.
I would accept a less tidy report if the figures could be compared. A disclosure could show the electricity consumed, identify the carbon method, and distinguish direct water use from wider water impacts. It could say what is still unknown. Where a component cannot yet be measured well, providers could give an estimate and explain its limits. Leaving it out makes the total smaller without helping the reader judge it.
Disclosure has to work both ways
There is a cost to going first. A provider that publishes a broad account can face criticism that a rival that publishes nothing avoids. It may also give rivals useful facts about how its business works. Those are real concerns when firms compete for the same buyers but face different reporting rules.
Mistral raised that first-mover problem in evidence considered by the French Senate. The Senate’s report, filed September 16, proposed common disclosure while protecting trade secrets. Those proposals are not European law. They offer more hope than waiting for each company to decide which figures it would like customers to see.
The rules should apply to all providers in the same market. Each should report against common definitions, and independent reviewers should have enough access to check the claims. A regulator or qualified auditor can see more detail in confidence than the public report would show. Keeping a design secret does not mean every fact about resource use must also be secret.
Those rules would also protect companies doing serious work to measure their costs. A broad disclosure should not become a competitive penalty because another supplier is allowed to omit the inconvenient parts. Common rules give a buyer a chance to spot real gains and give an engineer a reason to make them known.
We should expect revisions. Models change, workloads change, and estimates improve. A report should state its period and version so that a claim from last year cannot become a promise about the service being sold today. If the method changes, the provider should explain whether the gains came from changes to the service or a different way of counting.
Buyers need enough information to compare suppliers and check their claims, even if no one can account for every keystroke. The buyer opening 2 competing proposals should be able to put their resource accounts beside each other and understand the difference. The model’s inner workings can remain proprietary. What the buyer is being asked to believe should be open to inspection.
