Google estimates that a typical query sent to its chatbot Gemini consumes approximately 0.26 millilitres of water, equating to around five drops. However, when researchers factor in the water used to generate the electricity that powers data centres, this figure rises to approximately 0.725 millilitres, representing a significant increase of 179%. Previous studies suggest that this water footprint could be considerably larger when evaluated on a broader scale. A 2023 study co-authored by Shaolei Ren, a professor of electrical engineering and computer science at the University of California, Riverside, who investigates the environmental impact of AI, estimated that an earlier model of ChatGPT might consume half a litre of water for every 10 to 50 user interactions with the chatbot.
The Challenge of Understanding AI’s Water Footprint
The water footprint associated with AI is challenging to contextualise. For instance, producing a single pair of denim jeans may necessitate as much as 3,800 litres of water, according to an estimate from Levi’s cited by the United Nations Environment Programme (UNEP) in 2026. A cotton t-shirt is estimated to have a water footprint of 2,720 litres, while the annual consumption of coffee and tea accounts for around 140 trillion litres globally. Although these figures vastly exceed the estimates for a single AI chatbot query, they also encompass factors such as rainwater, irrigation, and water linked to agricultural and industrial pollution. The primary concern lies in the rapid growth of accumulated AI demand and the locations where this water is consumed.
Research co-authored by Ren in 2025 projected that global AI operations could extract between 4.2 trillion and 6.6 trillion litres of water annually by 2027, representing four to six times the annual water extractions of Denmark. Such a substantial demand raises questions about the actual water consumption of AI, particularly as data centres expand into regions already experiencing water scarcity.
The Complexity of Measuring Individual Queries
Determining the water consumption linked to a single prompt remains a complex task due to the variability in measurement criteria employed by researchers. In a study conducted in September 2026, researchers from Northeastern University and Boston University developed a framework named the “Water Cost of Intelligence” (WCI), which accounts for water consumed directly in data centres, the water used for electricity generation, and regional water scarcity. This framework elucidates why Google’s estimate is lower, as it only considers the water consumed at its data centres without factoring in the electricity required to power them.
Another challenge in assigning a precise figure to a query is that not all requests necessitate the same amount of computational effort. “We can estimate it for a clearly defined task under specific operating conditions, but ‘a request’ is no longer a consistent unit of work. It may generate a brief response or lead to extensive reasoning and multiple calls to the model to complete a complex task,” explains Ren. His earlier estimates were based on relatively straightforward responses from GPT-3. He elaborates that, although newer AI models process individual tokens more efficiently, a single request today may demand a substantially greater computational effort, as even a brief response does not necessarily require less work to generate.
Understanding Water Consumption in Data Centres
A data centre might consume little water on-site but could rely on extensive electricity generation, which, in turn, consumes additional water elsewhere. The dilemma between on-site water consumption and the supplementary energy consumption is particularly significant during hot summer months, argues Ren. Arif Gasilov, a partner at Gasilov Group and an expert in natural resources, energy, and water, questions how much insight an estimate per query truly provides. Servers can handle thousands of requests simultaneously, yet they also consume electricity simply by remaining operational. The water consumption attributed to a specific request is often calculated using averages rather than measured directly.
“Much of the consumption arises from being powered on and active, rather than from that small amount added per request,” Gasilov notes. Water consumption is not the sole concern; a data centre that draws from abundant water resources may exert less pressure on its environment than one reliant on a limited municipal system or a handful of aquifers in a desert community. “That can have a far greater impact on the local community and area, as it represents a much larger percentage of local consumption,” he explains.
The Trade-offs of Water Conservation Efforts
Gasilov highlights the role of closed-loop cooling systems, which recycle water rather than continuously extracting new supplies. While these systems can reduce water consumption in data centres, some require more electricity, potentially increasing water consumption at the power plants that supply them. He also questions whether efficiency improvements will keep pace with the rapid growth of AI usage: “I doubt that the increase in efficiency will be sufficient to offset the fact that there are more users on the platform.”