Does AI use too much energy and water?

Does AI use too much energy and water? The debate goes round in circles because the two camps are not talking about the same thing. One side looks at the query: a question put to a chatbot uses a tiny amount of energy. The other looks at the aggregate: billions of queries, the training of models and the building of new data centres weigh heavily on power grids and, locally, on water.

Both findings are true. The useful question for a business is not “should we stop using AI?” but “are we using the right size of model for each task?”, because that is where both the footprint and the bill are decided.

Position 1: data centres are making consumption explode

The strongest source on the aggregate is the special report Energy and AI from the International Energy Agency (IEA), published in April 2025. It estimates data centre consumption at the time of the report, then in 2030 in its base case. In round figures:

On electricity, the report adds that:

• data centre consumption has grown by about 12% a year since 2017, more than four times faster than total electricity consumption;
• the 2030 total would be slightly more than the entire electricity consumption of Japan at the time of the report, and AI is the main driver of this growth, alongside other digital services;
• the United States accounted for 45% of data centre consumption in 2024, China 25% and Europe 15%.

On water, in 2023 about two thirds of consumption was linked to energy and electricity generation, a quarter to direct cooling, and the rest to chip manufacturing.

Finally, the report stresses concentration: an AI-dedicated data centre can use as much electricity as a very energy-hungry plant such as an aluminium smelter, but these centres are far more geographically concentrated, and nearly half of US capacity sits in five regional clusters. The sector already weighs heavily in the electricity consumption of some local markets: that is where the strain is felt, on the grid as on water, depending on climate and cooling method.

Position 2: a query uses very little, and efficiency is improving fast

The other camp puts the individual query first. In August 2025, Google published a measurement of the footprint of its Gemini assistant. For the median text query, the company reports:

It is Google measuring its own services: this is not an independent assessment.

For its part, Stanford University’s 2025 AI Index report observes that the inference cost of a system as capable as GPT-3.5 fell by more than 280 times between November 2022 and October 2024, thanks to increasingly capable small models. A cost is not an energy consumption, but the same report notes that the energy efficiency of hardware is improving by 40% a year: doing the same thing takes fewer and fewer resources.

This camp adds that the IEA itself puts things in perspective: data centres account for only about a tenth of the growth in global electricity demand to 2030, less than industrial motors, air conditioning of homes and offices, or electric vehicles.

Why both are right

The gains per query are real, but they say nothing about the number of queries. When a resource gets cheaper, people use more of it: every efficiency gain opens up new uses, more agents, more models that “think” at length before answering. The two trends coexist: each query becomes more frugal, and yet the IEA projects a doubling of total consumption.

Two honest limits to keep in mind. The per-query figures for commercial services come from their vendors, who choose their method: the IEA notes that the lack of data on the consumption of commercial models prevents assessing it, and its own measurements cover open models, under test conditions. And its projections are scenarios, with uncertainties that the report itself says grow after 2030.

What we think

Making users feel guilty for every question put to a chatbot seems to us to miss the point: at the scale of one person, the effect is small, and the IEA shows that other uses weigh more heavily in demand growth. On the other hand, at the scale of a business that automates thousands of operations a day, the choice of models becomes a real lever, and it is the business that holds it.

The IEA tests give the order of magnitude:

The larger of the two uses more than 40 times as much electricity. A model that “reasons”, used on a simple task, uses twice as much as a model of comparable size. The trap is therefore to send everything by default to the most powerful model, including tasks a small model would do just as well, such as sorting an email or extracting an amount from an invoice. The biggest model is also the most expensive: frugality and savings point the same way here, which is rare enough to be worth underlining.

What would change our mind: independent measurements, not published by the vendors, showing that the per-query consumption of mainstream models is no longer falling, or that it goes back up with models that reason at length.

What this changes in practice for a business

1. Route tasks to the right model. Each automated task should go through the smallest model that gets it right: a table matches each type of task to a model, tested on real examples. Our guide to the cost of AI in business, from tokens to routing details the method.

2. Avoid unnecessary calls. Some processing does not need AI at all: a simple rule or a database query does the job without a model. And a result that does not change is cached rather than recomputed.

3. Look at where the model runs. The footprint depends on the electricity of the place where the computation happens. Hosting some processing in Europe, or an open model on your own server, is also a question of sovereignty, which we cover in our analysis of Europe’s dependence on the American cloud giants.

4. Measure before communicating. A business that wants to talk about the footprint of its uses must count its calls and the model used for each. Without that record, any figure displayed would be a second-hand estimate.

Where to start

The footprint of AI is decided less by a question put to a chatbot than by the processing that repeats thousands of times. Start with the list of your automated uses, with the model each one calls and their volume: those that run most on a large model are the first to test on a smaller one, and the same sorting lightens the bill as well as the footprint. Our guide to integrating AI in business takes these choices one by one, from the model to the cost. If you want to know what each use costs before deciding, our offer to pay for AI what it is worth, use by use starts with an audit of your bill and a benchmark on your own content.

Frequently asked questions

Sources

1. Energy and AI (new tab)iea.blob.core.windows.net
2. A measurement of the footprint of its Gemini assistant (new tab)arxiv.org
3. Stanford University’s 2025 AI Index report (new tab)hai.stanford.edu