AI’s Silent Water Bill
Written by: Daniela Cabello
In 2019, as engineers trained one of the most advanced AI models in history, utilities in Iowa noticed something unusual. Water withdrawals spiked far beyond normal levels, not because of agriculture or industry, but because the supercomputers powering artificial intelligence needed cooling. That local story reflects a global reality. AI may exist in the cloud, yet its expansion relies on very real natural resources on the ground. Every query to a generative model sets off a physical chain reaction. Servers heat up, cooling systems switch on, and water evaporates to carry the heat away. Researchers estimate that a short conversation with an AI system corresponds to about half a liter of water consumed indirectly through cooling. Individually, this seems negligible, but multiplied across millions of daily users, it becomes a reminder that digital services are not weightless. They carry a resource cost as concrete as electricity or steel.
The most interesting detail is that this cost is not uniform. The geography of a data center can determine how thirsty AI becomes. In Finland or Sweden, operators can rely on long stretches of cold air to keep servers at safe temperatures with almost no freshwater. In Arizona, the same workload needs evaporative towers that consume large volumes. Two identical models trained in different places therefore leave very different water footprints. For business leaders, this is not a technical footnote. It is a strategic factor. Location decisions now influence not only performance and cost but also the sustainability and resilience of operations.
Scale magnifies the challenge. Microsoft disclosed that its global water use rose 34% in a single year, while Google reported a 20% jump over the same period. In 2022 alone, Google withdrew about 25 billion liters of water and consumed close to 20 billion, an amount comparable to the annual supply of a mid-sized city. Projections suggest that by 2027, the global water footprint of AI could reach up to 6.6 billion cubic meters each year, more than the yearly consumption of entire countries such as Denmark. These numbers do not signal catastrophe, but they do highlight how rapidly digital growth can turn into a question of resource strategy.
What makes this discussion compelling is that solutions are emerging in parallel. Google has pioneered the use of treated wastewater, with one of its U.S. facilities reaching 97% reliance on recycled water for cooling. Microsoft has pledged to become water positive by 2030 and is investing in air-to-water condensers as well as the use of reclaimed water in places like Texas and California. These moves are not just corporate gestures. They are hedges against operational risk in regions where community acceptance and regulation will increasingly shape the future of infrastructure.
Even simple management choices can make a difference. Researchers have suggested that AI training, when possible, should be scheduled during cooler hours or seasons. Just as watering a lawn at night reduces evaporation, running workloads when outside temperatures are lower allows data centers to rely more on air cooling and less on water-intensive systems. This shows that sustainability is not only a question of technology but also of smarter timing and operational intelligence.
For professionals and entrepreneurs, the connection between AI and water is less about alarm and more about perspective. Every digital interaction draws on physical infrastructure, and that infrastructure has resource needs that vary by scale and location. As AI becomes embedded in business models and daily operations, understanding its water footprint will be as relevant as monitoring its energy demand or financial cost. What matters is not simply whether AI consumes water, but how companies manage that consumption, disclose it, and innovate to reduce it.
References
Associated Press. (2023). As AI booms, so does data centers’ thirst for water. AP News. https://apnews.com/article/artificial-intelligence-water-use
Ren, S., et al. (2023). Making AI less thirsty: Uncovering and addressing the secret water footprint of AI models. University of California, Riverside.
Vincent, J. (2023, August 30). Training large AI models uses a shocking amount of water. The Verge. https://www.theverge.com
OECD. (2023). AI, data centres and water sustainability. OECD Policy Paper.
MIT Technology Review. (2023). The hidden water footprint of AI. MIT Tech Review. https://www.technologyreview.com
