AI uses so much water because the chips running it get extremely hot, and the cheapest way to pull heat out of a room full of them is to evaporate water. That evaporation happens on site, and the water does not come back.
The bigger cost is the one nobody sees. Generating the electricity those servers consume also burns through water at the power plant, and that indirect footprint dwarfs what the buildings themselves use.
Evaporative cooling is where the on-site water goes
A server converts almost all the electricity it draws into heat. Blowing that heat away with air alone gets expensive fast, so many large facilities run cooling towers instead.
Warm water flows across the tower, a fraction of it evaporates, and the phase change carries the heat off. The evaporated portion is gone from the local supply, which is why engineers call it consumption rather than withdrawal.
Lawrence Berkeley National Laboratory measured the scale of it. American data centres consumed 66 billion litres of water directly in 2023, up from 21.2 billion litres in 2014. Hyperscale and colocation sites accounted for 84 percent of that.
The power plant drinks far more than the building
Thermoelectric plants need cooling too, and hydro reservoirs lose water to surface evaporation. Every kilowatt-hour a server pulls carries that hidden cost upstream.
The same Berkeley Lab report puts the indirect water footprint of US data centres at nearly 800 billion litres, roughly twelve times the direct figure. The national average works out to 4.52 litres per kilowatt-hour, slightly above the 4.35 litre US average for electricity overall.
That ratio is why the industry chases new generation capacity rather than better plumbing. It also explains the corporate interest in fusion and other firm low-carbon power projects.
What a single prompt actually costs
Google published measured figures for its own infrastructure in 2025. The median Gemini text prompt used 0.24 watt-hours of energy, emitted 0.03 grams of CO2 equivalent, and consumed 0.26 millilitres of water, about five drops.
Those numbers are small enough to feel harmless. The problem is not any single query, it is billions of them running against hardware that keeps getting denser.
The demand curve is bending sharply upward
US data centres used 176 terawatt-hours in 2023, or 4.4 percent of national electricity. Berkeley Lab’s scenarios put 2028 somewhere between 325 and 580 TWh, which lands at 6.7 to 12 percent of forecast US consumption.
Globally the International Energy Agency counted 415 TWh in 2024, roughly 1.5 percent of world electricity, and projects around 945 TWh by 2030.
Buildout at that scale reshapes supply chains as well as watersheds. It already shows up in memory chip pricing for consumer devices, and it drives proposals for gigawatt-class compute factories.
The fixes involve real trade-offs
Closed-loop and direct-to-chip liquid cooling cut on-site evaporation sharply. Siting facilities in cold climates does the same.
But cutting water often means spending more electricity on mechanical chillers, which pushes the indirect footprint back up. There is no version of this where a data centre stops consuming both.
Does one AI prompt really use a bottle of water?
Not according to the only large-scale measured disclosure available. Google reported 0.26 millilitres for a median Gemini text prompt, far below the higher estimates that circulated before. Image and video generation are heavier workloads and were not covered by that figure.
Is the evaporated water destroyed?
No, it returns to the atmosphere. It leaves the local watershed though, which is what matters to a municipality sharing an aquifer with a hyperscale campus.
How much of the grid do data centres actually take?
They reached 4.4 percent of US electricity in 2023 and about 1.5 percent worldwide in 2024. The forecast range for the US in 2028 runs from 6.7 to 12 percent, depending on how many accelerators ship and how they are cooled.
Joyce Lewis covers consumer technology, gaming, and online privacy for 3Zebras. She focuses on the practical side of modern tech, from smartphone tips and app guides to smart device security and the questions readers actually search for.