There is no single answer, because published figures count different things. One measurement by Google puts the water consumed on site for a typical text request to its Gemini assistant at about five drops. Other estimates cited in the same paper, of 45 mL and 50 mL, are more than a hundred times higher, because they use different methods and boundaries. To read any water figure, ask three questions: is it water taken in or water used up, is it water at the data centre or also water used to make its electricity, and what work does it cover?
Taken in, or used up?
Water experts separate two ideas. The US Geological Survey defines a withdrawal as water removed from the ground or diverted from a surface-water source for use, and consumptive use as the part of water withdrawn that is evaporated, transpired or incorporated into products or crops. Water that is withdrawn but returned, for example as treated discharge, can be used again downstream. Water that evaporates cannot, at least not locally.
At the site, or at the power station too?
A data centre uses water in two places. The first is on site, mostly in cooling. The second is off site, in the power stations that generate its electricity. The US Department of Energy's best practice guide, written by NREL, describes a site-based measure, the annual site water usage divided by the energy used by the IT equipment, and a source-based measure that also counts water used off-site in the production of energy used on-site. Lawrence Berkeley National Laboratory's 2024 report on US data centres uses the same split, and gives an average US on-site figure that stays just over 0.36 L/kWh through 2023 in its estimates.
The measure for on-site water is Water Usage Effectiveness, or WUE, defined in an international standard, ISO/IEC 30134-9, in litres per kilowatt hour of IT energy.
One published measurement of a single AI request
In August 2025, Google researchers published a measurement of its own systems. For the median Gemini Apps text prompt they report energy of 0.24 Wh and water consumption of 0.26 mL, which they describe as about five drops. Their method counts water consumed, meaning withdrawn minus returned, using Google's fleet WUE of 1.15 L/kWh, and notes that on average Google consumes 80% of the water it withdraws. The measurement covers serving the request in Google's data centres. It does not cover training the model, external networks or the user's own device.
| Question to ask | Why it changes the number |
|---|---|
| Withdrawal or consumption? | Withdrawal includes water later returned, so it is larger |
| Site only, or electricity too? | Water used in power generation can be larger than on-site water, depending on the power mix |
| Which task? | A short text answer, an image, a video or model training use very different amounts |
| Whose data centre and where? | Cooling design and climate change on-site water use per unit of energy |
What is not certain
- The Google figure is one company's measurement of its own service, in a paper written by its own researchers and not yet peer reviewed. It is a useful reference point, not an industry average.
- Water used in electricity generation depends on the local power mix and is not included in the Google figure.
- Estimates of total AI water use rely on assumptions about how much computing is done and where; they change as AI use grows.
Sources and method
- Elsworth and others, Google, Measuring the environmental impact of delivering AI at Google Scale, arXiv:2508.15734, 21 August 2025, operator-authored preprint. https://arxiv.org/abs/2508.15734
- US Geological Survey, Circular 1268, Estimated Use of Water in the United States in 2000, glossary, 2004. https://pubs.usgs.gov/circ/2004/circ1268/htdocs/text-glossary.html
- Van Geet and Sickinger, NREL, for the US Department of Energy Federal Energy Management Program. Best Practices Guide for Energy-Efficient Data Center Design, revised July 2024, section 8.3. https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design_0.pdf
- Shehabi and others, Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, 19 December 2024. Landing page https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report ; full text https://climate.mit.edu/sites/default/files/2025-02/lbnl-2024-united-states-data-center-energy-usage-report.pdf
- ISO/IEC 30134-9:2022, Data centres key performance indicators, Part 9: Water usage effectiveness, March 2022. https://www.iso.org/standard/77692.html
Method: each link was opened on 29 October 2026 and every figure was taken from the page cited beside it. No figure was estimated.
Read next
This article provides general information, not investment advice or an offer of securities.