Does One Chatbot Query Really Drink a Bottle of Water?

You have probably seen the claim, phrased as a gotcha. Every time you ask a chatbot a question, it drinks a bottle of water. Sometimes it is a glass, sometimes half a litre, sometimes the figure comes with a tidy illustration of a plastic bottle tipping into a server rack. The numbers are presented as settled fact. They are not, and the gap between the viral version and the measured one is wide enough to be worth walking through.

The honest answer starts with a different question: what does a single query actually cost in energy, since water use follows from that. In June 2025 OpenAI's Sam Altman published the company's first official figure, putting an average ChatGPT query at about 0.34 watt-hours of electricity and roughly one-fifteenth of a teaspoon of water. Independent work landed in the same neighbourhood. Researchers at Epoch AI estimated an average query at around 0.3 watt-hours. One-fifteenth of a teaspoon is not a bottle. It is a rounding error against a cup of coffee.

Where the scary numbers come from

So why do people quote figures hundreds of times larger? Three things are usually being blurred together.

The first is training versus use. Building a frontier model takes an enormous, one-off burst of energy and water spread across months of computation. Answering your question afterwards does not. The viral claims often take the total footprint of training and quietly attach it to a single prompt, which is a bit like charging one bus passenger for the cost of building the bus.

The second is the type of request. A short text reply is cheap. Generating a high-resolution image or a video clip can cost far more, and a long reasoning task that runs for minutes costs more again. Averaging these together, then presenting the result as what happens when you ask a simple question, inflates the number for the request most people actually make.

The third is cooling assumptions. Water use is not measured at the chip. It is calculated from energy use multiplied by how a given data centre is cooled, and that varies enormously. A facility using evaporative cooling in a hot, dry climate looks very different from one on a cold coast using outside air. Pick the thirstiest design and you can make any query look extravagant.

What is actually true

None of this means AI's resource use is a non-issue. It plainly is not. The aggregate demand is real and growing fast, which is why communities are fighting over data-centre water in places like Visakhapatnam in India, and why the strain these facilities put on local grids has become a genuine political question rather than a technical footnote. A single data centre can consume as much power and water as a small town, and there are a great many of them being built.

The point is that the right unit of concern is the industry, not your individual prompt. Treating one chatbot question as an act of environmental vandalism gets the scale wrong and, worse, aims the worry in a useless direction. The decisions that matter are about where these facilities are built, how they are cooled, and what powers them. Those are choices made by companies and regulators, not by whether you ask one more question today.

Altman's figures, it should be said, have not been peer reviewed, and the company has an interest in the number looking small. But the independent estimates sit close to it, and the burden of proof runs both ways. If a claim says your query drinks a bottle of water, ask where the figure came from. The answer is almost always a training cost, a worst-case cooling assumption, or a number with no source at all.