Why AI is So Thirsty
The heart of the issue is heat. AI models, especially large language models like the ones behind ChatGPT, require immense computational power. This is provided by thousands of high-performance servers packed into buildings called data centres. These servers generate
an enormous amount of heat, and if they overheat, they fail. To prevent this, data centres need constant cooling, and one of the most energy-efficient ways to do this is with water. Many facilities use evaporative cooling systems, where water absorbs heat and evaporates, a process that can consume millions of gallons per day. A single large data centre can use as much water as a small city of 50,000 people.
The Scale of Consumption
The numbers are staggering and growing rapidly. In 2023, US data centres used an estimated 228 billion gallons of water, factoring in both direct cooling and the water used for electricity generation. Projections show that by 2028, AI data centres globally could consume over a trillion litres of water annually. This surge is driven by the AI boom, with tech giants like Google and Microsoft reporting significant increases in their water usage. For instance, Google's data centres used over 5 billion gallons in 2023, while Microsoft's water use jumped 34% in a single year as it expanded its AI capabilities. The amount of water used for a simple AI interaction is debated, with estimates ranging from a few drops to an entire water bottle for a handful of queries, depending on what is measured.
The Hidden Water Footprint
Direct cooling is only part of the story. A much larger, often overlooked, aspect of AI's water footprint is indirect consumption. Over 90% of the water attributed to AI is used not in the data centre itself, but in the power plants that generate the vast amounts of electricity needed to run it. Many power plants, especially those using fossil fuels or nuclear energy, also rely on water for cooling. Furthermore, manufacturing the specialised semiconductor chips that power AI is an incredibly water-intensive process, requiring millions of gallons of ultrapure water daily for a typical factory. When you add up on-site cooling, power generation, and manufacturing, the true water cost of an AI query becomes much higher.
Local Impacts and Global Concerns
While the benefits of AI are global, its environmental costs are intensely local. Data centres act like giant straws, pulling massive volumes of water from local municipal systems, aquifers, and rivers. This becomes a critical issue when these facilities are built in water-stressed regions. Alarmingly, about two-thirds of data centres constructed since 2022 are in areas facing water scarcity, including hot, dry climates like Arizona. This puts the tech industry in direct competition with communities and agriculture for a finite resource, especially during heatwaves or droughts when demand from all sides peaks.
The Search for Solutions
The industry is aware of the problem and is exploring solutions. Companies like Microsoft and Google have pledged to become "water positive," meaning they aim to replenish more water than they consume by 2030. Innovations in cooling technology are a key focus. Closed-loop cooling systems, which recirculate the same water instead of constantly consuming fresh supplies, can reduce water usage by 50-70%. Other emerging methods include direct-to-chip liquid cooling, which is more precise, and even air-cooled systems that use virtually no water, though they may require more electricity. Using recycled or treated wastewater instead of fresh, potable water is another promising strategy being adopted to lessen the strain on local supplies.
