What Your Chatbot Is Actually Drinking

Ask a chatbot to draft a short email and it will answer in a second, with no sense that anything was spent to produce it. Yet researchers at the University of California, Riverside, have put a number to that exchange: roughly a bottle of water for every hundred-word reply, drawn not by the model but by the systems that keep it cool. On its own it is a footprint barely worth noting. It becomes interesting only when you consider how many of these exchanges now happen, and how quickly that figure has grown.

The pace is the striking part. Over the past year, AI assistants have moved from novelty to daily habit, with Claude's US app downloads overtaking ChatGPT's in early 2026 and its active audience expanding several times over, even as ChatGPT remains by far the larger platform, handling billions of prompts a day. What was once occasional curiosity has become routine, a rhythm of questions asked and answered at a scale most users never pause to picture.

The water itself does the unglamorous work of cooling. Data centres run vast banks of processors that generate real heat, and evaporative systems remain a common and effective way to manage it. As demand for computing climbs, so does the demand on the power and water that support it, which is why the industry's environmental conversation has shifted from energy alone to the fuller picture.

That picture is most visible locally. In parts of California, data centres now draw a substantial share of a city's electricity, enough to shape how the grid is planned. The world's AI computing capacity is concentrated in a small number of countries, which means the question of where facilities are built carries as much weight as how many. In Uruguay, a project advanced during a period when drought had already strained Montevideo's freshwater supply, a reminder that the same facility lands differently depending on where it sits. A recent United Nations assessment framed the tension plainly: AI's benefits travel widely, while its resource demands settle in the places that host the infrastructure. Some water-stressed regions have begun weighing new permissions more carefully, a sign the conversation is now practical rather than hypothetical.

The industry is engaging with it directly. Microsoft has committed to becoming water positive by the end of the decade and has introduced cooling methods that meaningfully reduce water use. The open question is whether efficiency can keep pace with growth, since each query is becoming leaner even as the total keeps rising. It is a solvable problem, and one the sector is increasingly measured against.

What makes the water story harder to hold in mind than the energy one is its quietness. The reply appears; the cost does not. Bringing that cost into view, so that the people who send the prompts and the communities that host the machines belong to the same picture, is the work now underway, and it is work the industry has already started.