A glass of water evaporates somewhere when you ask a chatbot “What’s the weather today.” That claim has circulated online for two years , and is misleading, though not entirely baseless. This report looks at AI’s water and electricity footprint using measured figures and names the source for each. Much of the confusion comes from comparing numbers that count different things.
Google said in an August 2025 technical report that a median Gemini text prompt uses 0.24 watt-hours of electricity and 0.26 millilitres of water , about five drops. The company’s estimate counts only data-centre cooling water. A 2023 University of California, Riverside study estimated 10–50 millilitres for a medium-length answer because it also counted water used at power plants generating the electricity. On that broader estimate, a half-litre bottle represents 10–50 answers. The “one bottle per prompt” claim applies the high end of that range to one answer.
Five drops are small; billions of five-drop servings are not. Companies do not publish daily prompt totals, so this is an illustration: one billion prompts a day at Google’s rate would mean about 95 million litres of on-site cooling water a year , around 38 Olympic-size pools. Adding power-plant water makes the footprint much larger. It is also concentrated in a small number of places, often in areas already facing water stress.
The International Energy Agency’s April 2026 report estimates global data centres used 415 terawatt-hours in 2024 and 485 TWh in 2025, a 17% rise. AI-focused centres’ use grew 50% in 2025 to 155 TWh. In the IEA’s central scenario, total use reaches 950 TWh by 2030 , about 3% of global electricity , with AI-focused use at 465 TWh. Two-thirds of 2025 data-centre electricity still went to non-AI services such as video, cloud storage, banking and email: AI was driving growth, but was not yet the majority of the load.
The IEA’s 2023 estimate divides 560 billion litres of data-centre water use into 373 billion litres at power plants, 140 billion for on-site cooling and 47 billion in chip and server manufacturing. Two-thirds is used at power plants, which may be far from data centres. Corporate claims about lower on-site cooling water can therefore be accurate while leaving indirect use out. The IEA projects total use could reach 1,200 billion litres by 2030.
Google said it used 10.9 billion gallons (41 billion litres) in 2025, a 34% annual rise and more than double its 2021 level. It says watershed projects replenish part of its consumption; its 2024 report put that share at 64%. Amazon reported 2.5 billion gallons for 2025 and Microsoft about 2.7 billion for 2024, but their reporting methods differ, so the figures are not directly comparable.
Location can matter as much as volume. A Houston Advanced Research Center and University of Houston study estimates Texas data centres will use 49 billion gallons in 2025 and 399 billion in 2030 , an eightfold increase. In Loudoun County, Virginia, one of the world’s largest data-centre hubs, facilities use 8% of municipal water; one projection puts the share at 29%. A sector report says more than 40% of planned AI data centres are in areas with high or extremely high water stress.
There is a countervailing trend: the IEA says energy per AI task has fallen by at least a factor of ten a year in recent years, a pace it calls unprecedented in energy history. It estimates that converting every search-engine query into an AI query would add less than 4 TWh annually, under 1% of current use. Text chat is relatively cheap. Video generation, long reasoning and agent tasks can use far more energy, and companies do not disclose the mix. The five largest technology companies spent more than $400 billion on data-centre investment in 2025, with a 75% increase expected in 2026.
Closed-loop cooling can reduce direct water use by 70–90%, at the cost of higher installation expenses and somewhat more electricity. Microsoft says it has used zero-water cooling designs in new data centres since 2024. Other options are siting facilities in cooler climates and low-stress watersheds, and requiring clearer water-use reporting. At least five U.S. states were considering rules in 2026; an Idaho proposal would require closed-loop systems for centres built after July 2026.
AI itself does not drink water; the electricity and cooling systems that run it use water, concentrated in particular places. One prompt may amount to five drops; one data centre can use as much as a town. The key question is not only “how much?” but “where , and whose water?”
Verified: IEA (April 2026) electricity figures, growth rates and efficiency estimates; IEA 2025 water split; Google technical report (0.24 Wh / 0.26 mL); UC Riverside study (10–50 mL); Google environmental reports (10.9 billion gallons, 34% rise); HARC / University of Houston estimate via Lincoln Institute. Uncertainty: Texas figures came from a draft not checked against a final report. Loudoun figures and the “40%” estimate rely on one source, Soeteck. Google’s 2021 figure is inferred from “more than doubled.” Amazon and Microsoft figures cover different years and methods. The pool calculation is our arithmetic. Not available: Most companies, including OpenAI and Anthropic, do not publish water-use data; public data on data-centre water use in Turkey was not found.
Leave a comment