Each time the AI uses ChatGPT to write a 100-word email, it consumes approximately 519 milliliters of water, which is almost the volume of a standard water bottle. This number comes from Peer-reviewed 2025 papers Pengfei Li, Shaolei Ren of the University of California, Riverside, and colleagues publish in ACM Communications. It includes both the direct water used to cool data center servers and the indirect water needed to generate electricity for those servers. If you scale this up to millions of users making dozens of transactions every day, the numbers become staggering. The global infrastructure handling AI queries is expected to consume 4.2 to 6.6 billion cubic meters of water per year by 2027, equivalent to half of the UK’s annual water withdrawals. Most of the water is pumped from areas that have dried out.
Why AI data centers consume so much water and where it goes
Data centers generate a lot of heat. As the chip at the core of modern AI high-end graphics processing units, each chip can consume between 300 and 700 watts of power and operate under high load as long as queries continue to come in. The most common method of managing heat is evaporative cooling: water is pumped into the facility, absorbs heat from the servers, and then releases some of that heat to the atmosphere in the form of water vapor. Approximately 80% of the water drawn into an evaporative cooling system is lost permanently through evaporation. The rest is recycled back, sometimes at higher temperatures and with chemical residue.The new generation of AI-specific hyperscale data centers are larger, denser, and more thermally intensive than the general-purpose cloud infrastructure built in the 2010s. A large campus now consumes more water in a day than a town of 10,000 that uses water for drinking, sanitation, cooking and agriculture combined. this 2024 U.S. Data Center Energy Use Report A report produced by Lawrence Berkeley National Laboratory for the U.S. Department of Energy estimates that by 2023, data centers will consume approximately 17.4 billion gallons of water directly through cooling and an additional 211 billion gallons indirectly through generating electricity to power the same facilities. Data center loads have tripled over the past decade and are expected to double or triple again by 2028.
Google , Microsoft and Yuan : What the numbers actually show
The largest tech companies have begun disclosing water consumption data in annual sustainability reports, and all are on a consistent trajectory.Google’s 2024 environmental report states that total annual water consumption is approximately 8.1 billion gallons, approximately 95% of which is used in data centers. This number is an 8% increase from 2023, a 17% increase from 2022, and a 20% increase from 2021. Google’s water consumption has nearly doubled in three years, and the company has cited AI workload growth as a key driver in successive reports.Microsoft’s consumption data is smaller, but has a similar shape. The company reported production of approximately 1.7 billion gallons in 2022, a year-over-year increase of 34%. Independent reports on GPT-4 training at a Microsoft data center cluster in West Des Moines, Iowa, in 2022 showed that a single training session consumed 11.5 million gallons of water in July 2022 alone, and 13.4 million gallons of water in August. The cluster has since expanded to five facilities, pumping 68.5 million gallons of water annually from local municipal water systems. In 2023, Meta consumed approximately 813 million gallons globally. AmazonThe company that operates the world’s largest cloud infrastructure does not release total water consumption figures.
Artificial intelligence is being built in the world’s most water-scarce regions
Li and Ren’s paper predicts that by 2027, global AI demand could require water withdrawals equivalent to more than four Denmarks, or close to half of the UK’s total annual water withdrawals. The problem isn’t just the volume, it’s where the volume comes from.Microsoft acknowledged in its 2023 sustainability report that about 42% of its water use that year came from areas classified as “water-scarce” in the World Resources Institute’s rating system. Google estimates that by 2023, it will be equivalent to 15% of the fresh water extracted from areas with severe water shortages.The practical consequences are already apparent. In Chile, Google suspended plans for a $200 million data center near Santiago after an environmental court ruled that the company had not adequately considered the impact on Santiago’s central aquifer, which has been experiencing a 15-year drought and is rationing residential water starting in 2022. Thirty-two new data centers are planned in the Mexican state of Queretaro, which in 2024 suffered its worst drought in a century. Microsoft has secured the rights to obtain approximately 25 million liters of water per year from the local aquifer, which currently has a deficit of 60 million liters per year. In Arizona, a $14 billion data center project was withdrawn in 2024 after local residents successfully opposed rezoning.
What is not being disclosed and why it matters
The numbers above are the ones the company chooses to make public. According to all available assessments, the true water footprint of the AI industry is much larger.Three disclosure gaps persist. The first is the difference between water withdrawal and consumption, that is, the amount of water permanently lost through evaporation versus the amount of water returned to the local system. Most reports only mention one number, and choosing between them can change the apparent footprint by three times or more. The second is the gap between direct cooling water and indirect power generation water. Li and Ren research estimates that this number is about twelve times the direct number, and few company reports include this number. The third is a lack of facility-level data: Company-wide annual totals cannot tell local communities whether their particular aquifers are under stress.The core contribution of the UC Riverside paper is that it provides reliable estimates of these gaps using publicly available proxies. Data that the AI industry refuses to release are increasingly being estimated within reasonable limits by independent academic researchers, making voluntary disclosures harder to avoid over time.
The industry has not yet given an answer
The global AI infrastructure is being built faster than any similar technology in modern history. The physical buildings built out of trillions of dollars of investment are, at their most basic level, large industrial-scale evaporative cooling systems with computing equipment inside.Each query is small. The total amount is not. It is an open question whether the technologies being developed within these facilities, such as better climate models, more efficient irrigation, and more accurate drought predictions, can begin to contribute to large-scale solutions sooner than the acceleration of water consumption that will define the actual environmental legacy of the present. Judging from the current trajectory, this question remains unanswered.
