How Much Water Does a Chatbot Really Drink?

A striking claim has followed AI chatbots around for a couple of years now: that every question you type costs a bottle of fresh water. It turns up in headlines and in the guilty asides of people who use the tools daily. The real figure is worth pinning down, because the honest answer is both smaller and more complicated than the slogan suggests.

Start with the numbers the companies themselves publish. Google released a technical breakdown putting the median text prompt to its Gemini app at 0.24 watt-hours of electricity and 0.26 millilitres of water, roughly five drops. OpenAI's Sam Altman has offered a similar order of magnitude, putting an average ChatGPT query at about 0.32 millilitres, a fifteenth of a teaspoon. These are self-reported figures, and Google notes its data has not been checked by an outside auditor, so treat them as the low end rather than gospel.

So where does the bottle of water come from? Largely from a different and broader way of counting. Researchers at the University of California, Riverside, led by Shaolei Ren, include not just the water that cools a data centre on site but the water consumed off site to generate the electricity the servers draw. Under that fuller accounting, a medium-length exchange can run to tens of millilitres, and the viral bottle figure tends to describe a whole session of ten to fifty questions rather than a single prompt. Neither camp is lying. They are measuring different things.

Two caveats keep the picture from being reassuring. First, heavier models cost more. A long reasoning task that churns for thirty seconds draws far more power, and water, than a one-line reply, so the median hides a wide spread. Second, scale is the real story. A few drops per prompt sounds trivial until it is multiplied by billions of prompts a day, which is why the strain shows up at the level of regional grids and reservoirs rather than your own conscience. The surge in data-centre construction, and the scramble to power it, is a genuine environmental question.

The useful correction is simple. A single chatbot reply is closer to a sip than a swig, and the honest worry is aggregate demand, not the cup of water you have been told sits behind each message.