Every time you type a question into a chatbot, the reply feels weightless. But the machines that produce it run hot, and cooling them takes water.

One 2025 estimate does the maths: multiply a single chatbot reply by roughly 700 million a day, across a year, and the water lost to cooling would match a year’s drinking supply for about 1.2 million people. The figure comes from a study titled “How Hungry is AI?”, posted online in May 2025 by Nidhal Jegham and colleagues.

It is a calculated estimate, not a meter reading, so hold it at that distance. Even so, the scale is hard to shrug off.

What 1.2 million people’s water actually means

The number starts from one small building block. The authors estimate that one short GPT-4o query uses about 0.42 watt-hours of electricity, give or take a little. On its own, that is trivial, less than a low-energy bulb burning for a minute or two.

The weight comes from repetition. Scaled to about 700 million queries a day across a year, the study estimates the total reaches electricity comparable to 35,000 U.S. homes, water equal to the drinking needs of 1.2 million people, and enough carbon that offsetting it would take a forest the size of Chicago. The water figure is the one users never see.

Where the water goes

Data centres run hot. The servers behind each reply throw off heat, and that heat has to go somewhere. Many buildings use evaporative cooling: water is drawn across warm air, some of it turns to vapour, and the heat leaves with it. That vapour does not return. It is counted as water consumed, not borrowed.

There is a trade-off built in. According to the Environmental Law Institute, cooling with air uses little water but burns more energy, while evaporative cooling saves energy but loses significant water to the atmosphere. Tuning a data centre to be power-efficient can quietly make it more thirsty. The same review notes U.S. data centres used an estimated 21.2 billion litres of water in 2014, rising to 66 billion litres in 2023.

Water is one leg of a three-legged problem

Looking at water alone understates the picture. The electricity figure, 35,000 homes for a year, and the carbon figure, a forest the size of Chicago, come from the same scaling. The study also found a wide spread across the 30 models it tested: the most demanding used more than 29 watt-hours for a long prompt, over 65 times the most efficient. Which model you use, and how long your prompt runs, can shift the footprint by more than tenfold.

The authors put the underlying tension plainly. The study argues that “as AI becomes cheaper and faster, global adoption drives disproportionate resource consumption”. Cheaper per query does not mean cheaper in total when the number of queries keeps climbing.

Why the weightless image sticks

None of this shows up at the keyboard. The reply appears, the cursor blinks, and nothing hints at a cooling tower venting vapour. The water side of AI stayed largely hidden until a 2023 study from Shaolei Ren’s group at UC Riverside estimated that training GPT-3 in Microsoft’s U.S. data centres could directly evaporate around 700,000 litres of freshwater.

Ren, an associate professor of computer science at UC Riverside, has explained why the water cost went unnoticed for so long. He told The Markup in 2023 that “Water footprint has been staying under the radar for various reasons, including the big misperception that freshwater is an ‘infinite’ resource and the relatively lower price tag of water.” In the same interview he said that “Many AI model developers are not even aware of their water footprint.” That was his read on the field two years before the Jegham benchmark, and it goes some way to explaining why the 1.2 million figure still lands as a surprise.

What the numbers don’t settle

These are estimates, and they carry the usual caveats. The Jegham figures rest on an assumed 700 million queries a day, not on data from the companies themselves. The per-query energy sits within about 19% of the 0.34 watt-hours OpenAI’s Sam Altman has publicly cited.

Water use also depends on where and when a model runs: a data centre in a cool, water-rich region behaves nothing like one in a hot, dry one. 

The open question is a race between two trends. Hardware keeps getting more efficient per query, which pulls the footprint down. Query volume keeps climbing as more people fold these tools into daily work, which pushes it back up. Whether efficiency gains outpace adoption, or adoption swamps the savings, will decide whether the 1.2 million figure looks high in a few years, or merely a baseline.