Your steak weighs two thousand hours of AI

You have been told that generative AI consumes water, electricity and carbon. That is true. You have been told, in the same breath, to watch your queries. That is where the reasoning breaks.

I have been training professionals in generative AI for three years. The question of environmental impact comes up in every session, usually towards the end of the first half-day, often with a note of guilt in the voice. The question is legitimate. What is far less legitimate is the state of the figures handed to us to answer it: one number per article, never the same one, never the same scope, and nobody to say which number measures what.

This piece sets out the orders of magnitude. Not to absolve AI. To make the question usable.

Two correct figures, a factor of one hundred

Before arguing about a figure, insist on knowing what it counts.

Google reports 0.26 mL of water per Gemini query. Mistral reports 45 mL. A factor of one hundred and seventy. Neither is lying. The first counts the water used to cool the data centre. The second adds the water used to generate the electricity and to manufacture the chips.

Almost every contradiction between articles is not about the data. It is about the scope. On water, three different quantities travel under the same word: what is pumped, what evaporates and never returns to the local basin, and what the electricity cost upstream. In the United States, that last one weighs roughly twelve times the cooling water. Two thirds of the water behind digital technology is not in the data centre.

On electricity, it is a stock against a flow. Data centres mean 415 TWh in 2024, about 1.5% of world electricity: the figure used by those who minimise. They also account for close to 10% of the growth in world demand to 2030: the figure used by those who alarm. Both sentences are true, and they are not talking about the same thing.

On carbon, the electricity mix decides. The same server emits roughly ten times more in Virginia than in France. The gap between training GPT-3, 552 t of CO₂e, and training BLOOM, 25 t at a comparable size, comes from that first, not from any feat of engineering.

A figure without a scope is not data. It is an argument in disguise.

What one hour costs

Come down to the unit. One query, then one hour.

Take one hour of active use, twenty queries, an explicit and debatable assumption. On the data centre side, that is roughly 6 Wh, 25 mL of water and 2.7 g of CO₂e on the world electricity mix. Six watt-hours is an LED bulb left on for forty-five minutes. Twenty-five millilitres is a mouthful.

These figures have fallen fast, and that fall needs to be read correctly.

Google reports dividing the energy per query by thirty-three in twelve months. Over the same period, the electricity consumed by AI-oriented data centres jumped 50% in the single year 2025. Unit efficiency collapses, volume explodes, and the total still climbs. That is the rebound effect, and that is the real subject.

One piece of honesty few articles take on. No provider publishes its total query volume. Without that denominator, nobody, including the authors of the most serious studies, can compute the real aggregate impact of generative AI. Every global estimate in circulation rests on assumptions about volume, never on declared data.

Count the screen

The data centre is not alone in the room. There is the screen in front of you.

A comparison is only honest if it counts both ends of the chain, the server room and the device. And it is very often the device that dominates. ADEME and Arcep measure this at French national scale: devices account for between 65% and 80% of the digital footprint, and three quarters of their impact is already settled at manufacture.

One hour of gaming on a gaming PC uses roughly seventy times more electricity than one hour of generative AI on the data centre side. One hour of video streaming on a television, thirteen times more. In both cases, the weight sits in the living room, not in the server room.

The item everyone points at is rarely the one that weighs.

What data centres actually run

AI has not emptied data centres of everything else.

Generative AI accounts for roughly 15% of data centre electricity in 2024 according to the IEA, and should reach 35 to 50% by 2030. That is considerable, and it is still a minority. Enterprise cloud and streaming together carry more than half the load.

One actor is missing from this picture, and the absence is instructive. Bitcoin uses roughly 120 TWh of electricity a year, 2,237 billion litres of water and 40 Mt of CO₂e, for a service that produces no useful computation for anyone. On water, it weighs four times every data centre on the planet combined. It almost never appears in articles about the footprint of AI.

Zooming out

Step outside digital. Look at the world balance sheet.

On all three resources, digital technology as a whole sits in the same order of magnitude as commercial aviation. It is a sector that counts. It is not the sector that decides.

Globally small, locally decisive.

That is the key to the whole file, and it is what the world balance sheet will never tell you. Data centres account for 0.015% of world freshwater, yet 38% of the American fleet sits in areas of high water stress, as do two thirds of the sites launched since 2022. At The Dalles, Oregon, a single site accounts for close to 40% of the town's water. The world percentage says nothing about what is happening in one valley.

The scale that puts it back in place

Put AI next to your ordinary gestures. Look.

Saving a hundred queries amounts to not driving seven metres. The gesture is symbolic. It is not material.

Where it actually plays out

Putting the volumes in proportion puts nothing else in proportion.

It would be dishonest to conclude that all is well. The best documented problems are simply elsewhere than where the debate looks for them. Local competition for water and for land. The chemical pollution of cooling blowdown, loaded with salts and biocides. Upstream manufacturing, with the ultrapure water of the foundries (TSMC used 101 million cubic metres in 2023), PFAS, solvents, mining. The mismatch in timetables between a data centre that takes two to three years to build and a grid that takes five to ten to reinforce. In the United States, fossil plant closures deferred to absorb the demand.

Your real levers come down to three gestures. Keep your devices longer: most of a device's impact is already settled at manufacture, and going from two years of use to four halves it without changing a single habit. Look at meat and at new textiles. Look at your travel, where ten kilometres in a petrol car are worth eight hundred hours of generative AI.

The false levers, meanwhile, occupy the ground. Cleaning out your inbox has almost no climate effect, and this is documented: storage accounts for roughly 0.5% of an email's life cycle, which led ADEME to reorient its own recommendations. Giving up a few queries is not an environmental policy, it is a ritual.

Lowering the resolution of a stream on a large screen, or switching off a gaming PC, does have a real effect, precisely because the device dominates.

Jean-Jérôme DANTON