AI Data Center Water Usage: How Much Water Does AI Really Use?

There is no single number, because water use depends on climate, cooling design and how busy the servers are. The most useful reference points:

For scale, one large evaporatively cooled campus can use as much water as a town of 10,000 to 50,000 people. Company disclosures show the trend: Google reported about 8.1 billion gallons of data center water consumption for 2024, and later reporting points higher again for 2025.

Every watt a GPU uses turns into heat, and an AI rack can draw 100 kW or more. The cheapest way to remove that heat at scale is evaporation: warm water is sprayed through a cooling tower, part of it evaporates, and that evaporation carries the heat away. The evaporated water is gone, which is why operators talk about water consumption rather than just withdrawal.

• Direct water is what the site's cooling system uses. Evaporative systems consume most of what they withdraw; the rest leaves as blowdown.
• Indirect water is used by the power plants that make the electricity. Thermal plants need cooling water too, so a data center on a coal or gas grid carries a hidden water footprint.
• Water Usage Effectiveness (WUE) is the industry metric: litres of water per kilowatt-hour of IT energy. Lower is better, and liquid-cooled halls can approach zero on-site.

Does a single AI chat use much water? OpenAI's CEO Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.000085 gallons, roughly one-fifteenth of a teaspoon. Small per query, enormous across billions of queries and the training runs behind them.

Cooling design is the biggest lever on water. The four architectures in our AI Build Configurator show the trade-off between energy, density and water:

The catch is that dry coolers, which reject heat without evaporating water, need more fan energy on hot days. In a desert you often trade water for electricity. In a cool climate such as British Columbia or Quebec, free cooling makes both low, one reason Canada is attracting AI data center investment.

Chipmakers are pushing the same way: NVIDIA says its next platforms can run on warm-water liquid cooling that needs little or no evaporative water, a vendor claim worth checking against real deployments.

For most operators the water bill itself is small next to electricity. The real cost of water is risk: drought restrictions, permits, municipal capacity and public opposition. Projects from Arizona to New Jersey have faced pushback over water, including residents near the DataOne Vineland AI data center.

That makes cooling a planning decision, not just an engineering one. Liquid cooling raises capital cost by roughly C$3.7 to C$4.5 million per MW in our model, but it cuts water, lowers PUE and allows denser racks, which can shrink the building. Our data center cooling cost calculator and the full configurator demo show how that trade plays out in each state and province.