Is AI Really Draining the World's Water?

Few statistics about AI have travelled as far as the one about water. You have probably seen a version of it: that a short exchange with a chatbot drinks a whole bottle of water, or that generating a single image swallows ten gallons. The numbers are alarming, they are easy to share, and on the scale of an individual query they are almost certainly wrong.

Start with the figure the companies themselves put forward. OpenAI's Sam Altman has said an average ChatGPT query uses about 0.34 watt-hours of electricity and roughly 0.000085 gallons of water, as Data Center Dynamics reported. That is a fraction of a millilitre, closer to a drop than a bottle. Independent analysis by Epoch AI arrived at a similar order of magnitude for the energy, roughly the amount an oven draws in a second or so.

Where the scary numbers come from

So how does a drop become a bottle? Two things inflate the viral figures. The first is a confusion between water withdrawn and water consumed. A data centre may cycle large volumes of water through its cooling systems and return most of it, but reports often quote the withdrawal figure as though every drop evaporated. The second is aggregation dressed up as a single query. Take a plant's total annual water use, divide by an estimate of queries, fold in assumptions about the electricity that powers it, and you can land almost anywhere. Change the assumptions and the number swings by a factor of hundreds.

It helps to put the quantities beside familiar ones. One analysis noted that even a generous estimate of daily water use for AI specifically sits far below what American golf courses and residential lawns consume every day. That does not make AI's footprint zero. It does mean the individual-query panic is misplaced.

The part that is actually worth watching

Here is where honesty cuts both ways. The per-query numbers are tiny, but the aggregate is not, and it is growing. Altman's figures have not been peer reviewed, the definition of an average query is fuzzy, and heavier requests can use ten to twenty times the energy of a simple one. None of those figures include the enormous cost of training a model in the first place. Multiply even a small per-query number by billions of queries and add a construction boom in data centres, and you have a real strain on local grids and water systems in the specific places these facilities are built.

Most researchers who study this argue that the electricity source matters more than the water. A data centre running on a coal-heavy grid does far more environmental damage than one on hydro or wind, whatever its cooling looks like, a point IEEE Spectrum has made in its own accounting.

So the myth and the concern can both be true. Your single chatbot question did not empty a water bottle. The industry's total appetite for power and cooling is still a genuine issue, and it is why companies are already racing to make data centres easier on the grid. The right response is neither to feel guilty about asking a question nor to wave the problem away, but to keep the argument attached to numbers that hold up.