Two numbers that seem to argue
In 2023 a group of researchers at the University of California, Riverside estimated that a model like GPT-3 "drinks" about half a litre of water for every 10 to 50 medium-length answers.1 In August 2025 Google published its own data: the median text prompt to the Gemini apps uses 0.24 watt-hours, emits 0.03 grams of CO₂ equivalent and uses 0.26 millilitres of water, about five drops.2 A few months earlier Sam Altman had written that an average ChatGPT query uses about 0.34 watt-hours.3
Five drops or half a bottle? These are serious numbers, published by serious sources. The point is that they do not measure the same thing.
What is counted, and what is not
When two estimates differ by an order of magnitude, it is worth asking what was left inside the boundary and what was left out. Here there are at least four differences.
- Which water. The Riverside study adds the water evaporated to cool the servers to the water used by the power stations that generate the electricity. Google's figure covers the water used in the data centres.
- Which year. Between 2023 and 2025 hardware and software changed a great deal. Google says that in twelve months the energy of the median prompt fell 33-fold and its emissions 44-fold.2
- Which request. A short text question is not a long conversation, nor an image, nor a model that "reasons" for tens of seconds. Median and average also tell different stories: a few very heavy uses move the average, not the median.
- Who measures. Company figures are the most detailed we have, but they are self-reported: nobody outside those companies can check them.
None of these choices is wrong. Each answers a different question. The problem starts when a number is quoted without its boundary and becomes a slogan.
A price that does not appear on the user's bill, but is paid by the atmosphere, by rivers, by soils, by the most exposed communities.
From "AI: L'Impronta Fantasma" (translated)
The bill that never reaches the user
Even granting that a single request weighs very little, the real bill is the total. According to the International Energy Agency, data centres used about 415 terawatt-hours in 2024, 1.5% of the world's electricity, and by 2030 demand could exceed 945.4 Artificial intelligence is the main driver of that growth.
This is the paradox I have been following for years: every request becomes more efficient, and precisely for that reason we make more of them. It is a dynamic as old as the steam engine, and a lower number in a press release does not solve it.
What to ask, as users and as companies
- Public, comparable methods. Every estimate should state its boundary: which energy, which water, which period, which kind of request.
- Independent checks. Today the best data come from the companies that build the models. Third-party measurements are needed, as already happens in other industries.
- Precise questions to suppliers. Anyone bringing AI into a company can ask for the energy per request, the energy mix of the data centres and the water use. These are legitimate questions, and those who ask first get better answers.
- The right tool for the task. A small model is enough for many everyday jobs. Using the most powerful one for everything is like taking a lorry to buy bread.
Between five drops and half a bottle there is no truth and no lie. There is a question we have not yet learned to ask well: how much does it cost, all in? That is why I built a calculator anyone can use, and why I keep following the invisible weight of artificial intelligence.