A 100-word email written with ChatGPT may use about 17.6 fluid ounces of water, roughly the size of a standard bottle, once researchers count both server cooling and the water used to generate electricity. That estimate is not a meter reading attached to every prompt, but it gives everyday users a way to picture a hidden cost that usually stays out of sight.
The bigger concern is not one email. It is billions of AI requests layered across search tools, office software, customer service bots, image generators, and apps we now use without thinking twice. Researchers Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren estimate that global AI demand could require 1.1 trillion to 1.7 trillion gallons of water withdrawal per year by 2027, close to half the United Kingdom’s annual water withdrawal. Small bottle. Big pipe.
Why AI gets thirsty
Data centers generate heat, and AI servers can run especially hot. To keep them working, many facilities rely on cooling systems that use water because water can move heat efficiently, much like sweat cools the body on a sticky summer day.
That does not mean every AI answer has the same footprint. The actual number depends on the data center, the cooling system, local weather, the electric grid, and the length or complexity of the response. In practical terms, asking for a quick answer is different from requesting a long report, just as running one lightbulb is different from cooling a whole office building.
The numbers are getting harder to ignore
Google’s 2025 environmental report says Alphabet consumed about 8.1 billion gallons of water in 2024 across data centers and offices. Data centers accounted for about 7.8 billion gallons of that total, while offices and other facilities used about 348 million gallons.
The company also said its overall water consumption rose 28% from 2023 to 2024, even as it pointed to local cooling choices and water stewardship projects as part of its response. Google reported replenishing about 4.5 billion gallons in 2024, equal to roughly 64% of its freshwater consumption.
Microsoft’s latest data sheet shows a similar tension. In FY24, the company reported 5,807 megaliters of water consumption, equal to about 1.5 billion gallons, and said 42% of that consumption came from areas with high or extremely high water stress under the World Resources Institute’s Aqueduct tool.
Electricity hides more water
Cooling towers are only part of the picture. A 2024 Lawrence Berkeley National Laboratory report estimated that U.S. data centers directly consumed 66 billion liters of water in 2023, or about 17.4 billion gallons. Hyperscale and colocation sites accounted for most of that total.
Then comes the water hidden inside electricity generation. The same report estimated that U.S. data centers used about 176 terawatt-hours of electricity in 2023 and carried an indirect water footprint of nearly 800 billion liters, or roughly 211 billion gallons, because power plants also consume water. That’s why the electric bill and the water bill are more connected than they look.
Drought turns data centers into local fights
Water is not like carbon dioxide, which spreads through the atmosphere as a global problem. Water stress is local. A gallon pulled from a rainy region is not the same as a gallon pulled from an aquifer already under pressure.
That is why projects are facing closer scrutiny. In Chile, Google revisited plans for a $200 million Santiago-area data center after an environmental court raised concerns about impacts on the city’s aquifer during a long drought. Reuters reported that Google planned to start over with a new proposal using air-cooled technology instead of the original water-cooled system.
Google has also pointed to air cooling in places it identifies as higher water risk, including Mesa, Arizona, and Canelones, Uruguay. That matters because in dry regions, a data center can become more than a tech project. It can become a kitchen-table issue for families wondering about rationing, utility costs, and the future of local growth.
What needs to change now
The first fix is transparency. Companies should report water withdrawal, water consumption, peak water demand, direct cooling water, and indirect electricity-related water at the facility level. A global total looks tidy, but it does not tell residents whether their local aquifer is being stretched.
The second fix is smarter siting and cooling. Air cooling, closed-loop systems, recycled water, non-potable water, and workload shifting can all help, but each comes with tradeoffs. Air cooling may save water while raising electricity use. Evaporative cooling may save energy while consuming water.
At the end of the day, AI may still help society manage water better through drought forecasting, leak detection, irrigation planning, and climate modeling. But that promise only holds up if the industry measures its own footprint honestly.
The study was published on Communications of the ACM.
