What Your Chatbot Is Actually Drinking

Generating a 100-word email with OpenAI's GPT-4 consumes roughly 519 millilitres of water, just over one bottle, according to research from the University of California, Riverside. Send one such email a week for a year and the total reaches about 27 litres, a small individual footprint that becomes substantial once multiplied across hundreds of millions of users sending billions of queries.

That individual footprint is being multiplied at extraordinary speed. Claude's mobile app overtook ChatGPT's in US App Store downloads in early 2026, reaching roughly 149,000 daily downloads against ChatGPT's 124,000, while Claude's own daily active users climbed past 11 million, up more than 180% since the start of the year, according to Similarweb. ChatGPT remains far larger in absolute terms, with around 250 million daily active users worldwide, 900 million weekly active users, and 2.5 billion prompts sent each day, per OpenAI's own figures. What began, for most people, as occasional curiosity has become routine, repeated behaviour at a scale of billions of queries daily, each one drawing a small, invisible amount of water and electricity that no individual user has much reason to notice.

The water is not used by the model itself. It cools the data centres where queries are processed, facilities that generate significant heat running thousands of processors simultaneously. Global data centre electricity demand is on track to more than double by 2026, pushing operators toward whichever cooling method is cheapest, with water-intensive evaporative systems still common across the industry.

The strain shows up locally before it shows up in any company's annual report. In Santa Clara, California, data centres now consume 60% of the city's electricity, raising the risk of blackouts and pushing infrastructure costs onto residential utility bills. More than 90% of AI-specialised computing capacity is concentrated in just two countries: the United States and China.

Where the resulting data centres get built compounds the strain rather than spreading it evenly. In Uruguay, a data centre project went ahead in the same period a 2023 drought had already depleted Montevideo's freshwater reserves to the point that tap water briefly became unsafe to drink. A 2026 United Nations assessment of AI's environmental footprint concluded that the world has moved beyond a water crisis and into a state of global water bankruptcy, warning that AI's benefits tend to flow across borders and sectors while the burden, water withdrawals, land use, and electricity demand, concentrates in the specific communities where the infrastructure is sited. Some municipalities in water-stressed regions have begun refusing planning permission for new liquid-cooled facilities, a sign the tension between rising data demand and local water security is no longer theoretical.

Large technology companies are not ignoring the criticism. Microsoft has pledged to be water positive by 2030 and has rolled out immersion cooling systems that cut data centre water use by 31% to 52%. Whether voluntary efficiency gains can outpace underlying demand growth remains the open question, since even as individual queries become more efficient, the volume of queries keeps climbing.

What makes the water story harder to register than the energy one is its invisibility. A chatbot response arrives instantly, with no signal that water was drawn from a regional supply or electricity pulled from a strained grid to produce it. The bill, when it eventually arrives, tends to land on whoever lives nearest the data centre rather than whoever sent the prompt.