Four billion years in motion: the one substance every living thing depends on. Borrowed, returned, and never truly used up. Until now, the asking was ours alone.
Before there was anything to be thirsty, there was water. It moves through stone and skin without asking. Everything alive is mostly the water it has borrowed, the same few drops circling the earth for an age. We were only ever the thirst it learned to carry.
The turn
Every drop now has a rival.
The surface begins to warm. What was only ever ours to ask for now answers to something else.
The AI boom
The boom is real. Its bill arrives, quietly, as water drawn down and power burned to keep the machines cool. What follows is what one industry is learning to drink, accounted the way the water is: by scope.
Scope 1: on-site cooling
The water the machines drink where they stand
Servers run hot. To stay alive they are cooled, and much of that cooling is evaporative: fresh water turned to vapour on site and not returned. It starts, contested, at the scale of a single reply.
How much water does one answer cost? The estimates disagree:
- About 500 mL per 20 to 50 ChatGPT queries (UC Riverside, Li et al., 2023)
- 519 mL per 100-word GPT-4 response (The Washington Post, 2024)
- About 0.32 mL per typical query (Sam Altman, OpenAI, 2025)
- 0.26 mL per median text query (Google, Gemini, 2025)
The gap is not an error. It is a question of what you agree to count.
A single large data centre can draw up to five million gallons of water a day, the daily use of a town of 10,000 to 50,000 people (Environmental and Energy Study Institute).
One Google-owned data centre's on-site cooling exceeded 6.07 billion gallons, about 23 billion litres, of fresh water in a single year, 2023 (ACM research, via BGR).
Google's Cerrillos facility in Santiago, Chile draws an estimated 7.6 million litres of drinking water a day, in a region already in severe drought (IE Insights).
Scope 2: electricity generation
And the water the grid drinks to keep them running
Every watt behind a query is generated somewhere, and most generation is thirsty too. The power bill is a second water bill, one grid removed, and it is climbing fast.
Data centres used about 415 TWh in 2024, roughly 1.5 per cent of global electricity. Demand grew about 17 per cent in 2025 to about 485 TWh, with AI-focused data-centre electricity up around 50 per cent (IEA, Energy and AI).
On the central projection, data-centre electricity roughly doubles to about 945 TWh by 2030, close to the entire electricity demand of Japan (IEA).
A single ChatGPT query was estimated to draw about 2.9 Wh, roughly ten times the 0.3 Wh of a conventional Google search (Brookings, 2024).
Scope 3: embodied, in chip manufacture
And the water already spent before a query is ever asked
Long before inference, the silicon itself is washed clean. Embodied water is spent once, upstream, in the fabrication plant, counted against the machine before it does a thing.
Manufacturing a single microchip can consume around 2,200 gallons of ultra-pure water (OECD.AI).
Training GPT-3 alone was estimated to evaporate roughly 700,000 litres of freshwater (UC Riverside).
Embodied cost lands in the fab, not the data centre, which is exactly why it so often falls outside the number a company chooses to report.
Every model runs on a river
The cloud is not weightless. Cooling is just water, spent: one industry's thirst, re-priced as compute.
Reflection
The question was never whether the machines would think. It is what they will drink while they learn.
Morgan Stanley projects AI-related data centres could consume more than a trillion litres a year by 2028, about an elevenfold rise from 2024 (Morgan Stanley, via IE Insights).
The World Resources Institute puts AI-driven data-centre water withdrawal at 1.1 to 1.7 trillion gallons a year by 2030 (World Resources Institute).
In Aragón, Spain, residents protested new data centres under a single line: "Your cloud is drying my river."
The arithmetic is contested; the thirst is not.