Your query has a cost. Aside from the myriad ethical and social dilemmas that have sprung up since generative artificial intelligence hit the mainstream almost three years ago, there’s a hidden cost to asking ChatGPT to write you an email: the water we drink. It’s no small amount either; the average 100-megawatt data centre, where AI workloads happen, slurps up about two million litres of water each day.

“When we run AI servers, they burn a lot of energy, and this energy turns into heat. So, to get rid of the heat, we have to use a massive cooling system,” says Shaolei Ren, an associate professor at the University of California, Riverside, who researches AI. The heat needs to be moved from the servers to the building and then outside. “Many data centres are using open-loop cooling towers that are just evaporating water 24/7,” says Ren. Water is also used indirectly through electricity generation to power the servers.

Ren and his team calculated that for 10 to 50 medium-length responses, GPT-3 needs to “drink” 500 millilitres of water. This depends on the data centre — there’s a trade-off between water and electricity efficiency. Geography also matters. Hotter climates require more direct water consumption, making Canada an attractive place to put data centres. Electricity is also cheap here and often renewable, especially in British Columbia and Quebec.

Canada now has 239 data centres — a number set to grow, largely because of AI. Bell plans to build six AI data centres in British Columbia, Alberta is jostling to be “North America’s destination of choice for AI-enabled data centre investment,” and the federal government is investing $2 billion in computing infrastructure over the next five years.

And yet, a Bloomberg investigation in May 2025 found most data centres are in areas that are already water stressed. Between 80 and 90 per cent of water consumed by data centres is potable. “They don’t have to use drinking water, but usually most of the cooling systems require clean water,” says Ren.

With climate change already altering the global water cycle, Ren is hoping for more transparency and accountability from the AI industry on its environmental toll. “But at the same time, it’s not something we should panic about,” he says. “We should evaluate the benefit versus the cost. And environmental cost is just part of the cost.”