The enormous water consumption of AI data centers is a recurring argument, especially among critics, against the unbridled expansion of digital infrastructures. But how high the actual demand is only becomes apparent upon closer inspection: climate, cooling technology and electricity mix play a key role in determining the ecological balance.
Water is used in data centers primarily to cool servers, which generate large amounts of waste heat when computing. Depending on the technology, it is used directly in evaporative cooling systems or indirectly through the water consumption of the power plants that supply the required electricity.
However, there are currently numerous calculations circulating around the water consumption of AI data centers, some of which clearly contradict each other. Estimates range from comparatively moderate amounts to alarming scenarios.
Just recently, journalist Karen Hao had to backtrack on a calculation from her bestseller “Empire of AI”. She published in her book that a planned Google data center near Santiago de Chile could require “more than a thousand times the water consumption of the entire population.” However, there was a calculation error – the number was much too high.
Because of such incidents, experts warn against absolutizing individual values. Location factors, the technical equipment of the data centers and the electricity mix used can have a major influence on how water-intensive AI actually is.
KI: What factors influence water consumption?
The water consumption of AI data centers depends heavily on numerous factors. For example, the use of more water can mean that there is no need to operate electrical cooling systems.
If more electricity is used to cool the data centers, water consumption decreases. However, at the same time – depending on the electricity mix – greenhouse gas emissions can also increase again.
“Every site is different,” explains Fengqi You, a professor of energy systems engineering at Cornell University Wired. “How much water you need for the same amount of AI depends on the climate, the technology used and the [Energie-]Mix it up.”
Problematic water consumption depends on the location
At the same time, what makes things even more difficult is that some calculations also include indirect water consumption. However, the entire water footprint is only estimated, which also includes figures for electricity generation. As a result, it could be that the assumed figures from estimates are much larger than the actual water consumption on site.
Nevertheless, the water consumption of AI data centers should not be underestimated. “In the short term, this is not a problem or a national crisis,” says Cornell professor You. “But it depends on the location. In places where there is already a water shortage, building these AI data centers will be a big problem.”
However, calculations are very complex due to the numerous factors. Estimates that, for example, writing an email with ChatGPT uses an entire bottle of water are therefore hardly applicable for an “average” query.
Also interesting: