Data centers use water to cool the servers and keep them from overheating.
Why
Servers produce heat. Cooling systems use water to carry heat away, through cooling towers that evaporate water.
How much
A large data center can use 3 to 5 million gallons of water a day. In 2014, US data centers used about 626 billion liters (Lawrence Berkeley National Laboratory estimate).
AI
AI training and inference raise water use.
Alternatives
Air cooling, liquid cooling and using recycled water.
Impact
Local water stress is a concern.
Cooling demand in numbers
Servers turn nearly all electricity into heat, so data centers use cooling towers that evaporate water to remove it. A 2021 study from Lawrence Berkeley National Laboratory estimated US data centers directly consumed about 66 billion liters of water in 2014 and projected higher growth. A large facility can use hundreds of thousands to several million gallons per day.
- Water use effectiveness (WUE) is liters per kWh of IT energy, and an average around 1.8 L/kWh was reported by the Uptime Institute.
- Researchers at UC Riverside estimated training GPT-3 in Microsoft’s data centers could evaporate about 700,000 liters of freshwater.
- Indirect water use at power plants generating the electricity is often larger than on-site use.
Alternatives
Air cooling uses little water but needs more electricity. Liquid immersion and direct-to-chip cooling cut both. Facilities can use recycled or non-potable water, seawater or cold climates for free cooling. Some operators have pledged to be water positive by 2030, such as Microsoft, Google and Amazon, though accounting methods vary.
Water use
Data centers use water in cooling towers to remove heat from servers. A 2023 study by researchers at UC Riverside estimated that training some large AI models can consume hundreds of thousands of liters of water. A large data center can use millions of gallons a day, with Google reporting about 5.2 billion gallons of water use across its data centers in 2022.
Alternatives
Liquid cooling, air cooling in cool climates, reuse of reclaimed water and siting near renewable sources reduce water use. Companies measure water usage effectiveness (WUE) in liters per kWh.
Scale of the problem
A 2023 study by the University of California, Riverside estimated that global AI demand could account for 4.2 to 6.6 billion cubic meters of water withdrawal in 2027. US data centers consumed about 66 billion liters of water directly in 2023, according to a Lawrence Berkeley National Laboratory report, with projections of 2 to 4 times by 2028. Microsoft reported 6.4 million cubic meters of water consumed in fiscal 2022.
Reducing the water footprint
Options include closed loop liquid cooling, using recycled or reclaimed water, siting facilities in cool climates, immersion cooling and running warmer. Water usage effectiveness (WUE) averages about 1.8 liters per kWh in the industry. Locations in drought areas draw community concern.
Based on data from Lawrence Berkeley National Laboratory and the University of California, Riverside
