When people ask how much water AI “drinks,” the honest answer is: more than most of us can see. The water is used not by the chatbot itself, but by the data centers and power systems behind it. A 2025 study in Nature Sustainability estimated that AI-server deployment in the United States could create an annual water footprint of 731 million to 1.125 billion cubic meters across 2024-2030 scenarios, and about 71% of that total would come indirectly from electricity generation rather than from water poured directly into cooling systems. (nature.com)
Why is water involved at all? Because powerful chips get hot. The International Energy Agency says data centers used about 415 terawatt-hours of electricity in 2024 and could reach around 945 TWh by 2030; cooling alone can account for roughly 7% of power use in efficient hyperscale facilities and more than 30% in less efficient ones. In some places, water cooling is chosen because it can cut energy use: Google says water-cooled data centers use about 10% less energy and emit roughly 10% less carbon than many air-cooled sites. (iea.org)
At the level of one AI task, however, the picture is surprisingly uneven. A 2023 paper estimated that training GPT-3 in Microsoft’s U.S. data centers could directly evaporate about 700,000 liters of freshwater. Yet Google’s 2025 production study found that the median Gemini Apps text prompt used about 0.26 milliliters of water and 0.24 watt-hours of energy—roughly five drops of water—and that Google had cut energy use per median prompt by 33 times over one year. Lawrence Berkeley National Laboratory also reported in 2025 that workload-level data-center water use can vary by more than 10,000-fold depending on server efficiency, cooling design, location, and the local electric grid. (arxiv.org)
That is why the sustainability debate is shifting from simple shock numbers to smarter design. Microsoft says that datacenter designs introduced from August 2024 use zero water for cooling and can avoid more than 125 million liters of water per year per facility; pilot sites in Phoenix, Arizona, and Mt. Pleasant, Wisconsin, are scheduled for 2026, with the new sites expected to come online in late 2027. Google, meanwhile, says it aims to replenish more freshwater than it consumes by 2030 and reported replenishing more than 7 billion gallons through 165 projects in 2025. So the real question is no longer whether AI uses water. It does. The harder question is whether the next wave of AI can grow with cleaner power, better cooling, and honest local reporting. (microsoft.com)
