AI's water bill is real — but it's not your chatbot prompt

The meme has a point, it just lands in the wrong place. Global AI data centres consumed roughly 560 billion litres of water directly in 2025, and a UN University report projects 9.3 trillion litres a year by 2030 if AI grows to 40% of data centre electricity — the basic domestic water needs of 1.3 billion people. Google alone reported 10.9 billion gallons last year, up 34% in twelve months.

But the bigger draw isn't the cooling tower, it's the power plant. Indirect water — evaporated generating the electricity that runs the racks — dwarfs on-site use, roughly 211 billion gallons in the US in 2023 against a fraction of that for direct cooling. Any single prompt's share is measured in drops. The damage is aggregate and, critically, local: one large facility can pull up to 5 million gallons a day, frequently in counties that are already water-stressed.

A fix is partly underway. The newest AI chips require closed-loop liquid cooling, which recirculates coolant and can run at near-zero freshwater draw. What's still missing is numbers — the UN report's lead author's ask is blunt: make providers disclose.

Why it matters: the water fight over AI gets settled county by county, and right now almost nobody outside the operators knows the real figures.