Research·note·Apr 2025·3 min read
AI’s Thirst: The Hidden Water Cost of the Digital Age
Applied AISustainability
Artificial intelligence is rewriting software, media, and work. It is also quietly thirsty. Training a model the size of GPT-3 can evaporate as much as 700,000 liters of drinking water. By 2027, global AI demand could withdraw 6.6 billion m³ of water — more than the annual use of four to six countries the size of Denmark — and evaporate 0.6 billion m³ that does not come back. Where does that water go, and what, if anything, changes the bill?
The water bill behind every prompt
A University of California, Riverside study estimates that training GPT-3 in Microsoft data centers used between 700,000 and 5.4 million liters of water, depending on cooling efficiency and the energy source. Each interaction with the model — one ordinary reply — uses between 10 and 50 ml of potable water, varying by region and time of day. An 800-word answer is about a tenth of a 500 ml bottle.
Text is the small end of the ledger. Generating an image in a Ghibli-like style can take up to 34.77 liters: 0.57 liters to cool servers and 34.2 liters to produce the electricity (using a U.S. average of 34.7 L/kWh). Multiply that by billions of images made for amusement, and the total stops looking like a rounding error.
Why AI drinks so much
Water shows up in three places:
- Server cooling: Data centers that host large models dump enormous heat. Cooling towers evaporate water to carry that heat away — as much as 2.1 million liters a day at sites such as Google’s.
- Electricity generation: Thermoelectric plants that still feed many data centers withdraw and evaporate water at scale. In the United States, producing 1 kWh is estimated to use 3.14 L of water.
- Hardware manufacturing: Chips and servers need ultrapure water. The indirect load adds millions of liters a year.
The “thirst” is not the same everywhere. In Washington state, a single AI reply uses about 47 ml, a mix of warm climate and evaporative cooling. In Mexico it is about 24 ml. In Ireland, with cool weather and a greener grid, it drops to 7 ml.
A thirsty decade ahead
An arXiv paper projects that by 2027, global AI demand will withdraw between 4.2 and 6.6 billion m³ of water, about half of the United Kingdom’s annual use. Of that, 0.38–0.60 billion m³ will evaporate, in a world where WWF expects 66% of people to live under water stress in 2025. Microsoft’s water use rose 34% in 2022; Google’s rose 22%. Both have pledged to replenish more water than they use by 2030. Outside reviewers are not convinced the accounting holds.
What actually helps
- Transparency: Companies already publish carbon figures. Water used to train and serve a named model should sit next to them.
- Siting: Cold climates and grids heavy on hydro or wind cut the water footprint.
- Cooling design: Air cooling and recycled water beat once-through evaporative towers.
- Use: Developers and users can stop treating generation as free. Idle prompts still evaporate something.
Larger models such as GPT-4o only raise the load. AI can still be a tool against climate risk. Unchecked, it becomes another claimant on rivers that are already spoken for.