How Much Water Does a ChatGPT Prompt Really Use?

Google estimates that a typical question submitted to its Gemini chatbot consumes about 0.26ml of water, or roughly five drops. However, if researchers factor in how much water is spent to generate the electricity that powers data centres, that figure rises to about 0.725ml, or 179% more. Earlier research suggests the footprint can be considerably larger when put into context at scale. A 2023 study coauthored by Shaolei Ren, a professor of electrical and computer engineering at the University of California, Riverside who studies AI’s environmental footprint, estimated that an older ChatGPT model could expend half a litre of water for every 10 to 50 user interactions with the chatbot.

AI’s water footprint is difficult to put into perspective. Producing a single pair of jeans can require 3,800 litres of water, according to a Levi’s estimate cited by the United Nations Environment Programme (UNEP) in 2026. A cotton T-shirt has an estimated 2,720-litre water footprint, while a year’s worth of coffee and tea consumption accounts for roughly 140 trillion litres globally.

While those figures tower over the estimates for a single AI chatbot prompt, they also include factors like rainwater, irrigation and water tied to agricultural and manufacturing pollution. The larger concern is how quickly AI’s cumulative demand is growing, and where that water is being consumed.

Research coauthored by Ren in 2025 projected that global AI operations could withdraw between 4.2 trillion and 6.6 trillion litres annually by 2027, which is the equivalent to four to six times Denmark’s annual water withdrawals. That scale of demand raises questions about how much water AI actually uses, especially as data centres expand into regions already facing water shortages.

But if AI’s water demands are growing so rapidly, why is it still so difficult to determine how much water a single prompt uses?

The estimates depend largely on what researchers measure. In a September 2026 study, researchers from Northeastern University and Boston University developed a framework called the Water Cost of Intelligence (WCI) that accounts for water consumed directly at data centres, water used to generate electricity and regional water scarcity. That helps explain why Google’s estimate is lower: it only accounts for water consumed at its data centres, not the electricity needed to power them.

But another problem with assigning a fixed number to a prompt is that not every request requires the same amount of work.

“We can estimate it for a clearly defined task under specified operating conditions, but ‘one prompt’ is no longer a consistent unit of work,” Ren says. “It can produce a short answer or initiate extensive reasoning and multiple model calls to complete a complex task.”

His earlier estimates were based on relatively straightforward GPT-3 responses. He explains that while newer models process individual tokens more efficiently, a single request today can trigger substantially more computation. Even a short answer doesn’t necessarily require less work to produce.

Ren says it’s important to distinguish between water consumed at data centres and water used to generate electricity. “A data centre may consume little water on-site while relying on more electricity generation that consumes additional water elsewhere,” he says. “The trade-off between on-site water use and additional power usage is particularly significant when it’s hot in the summer.”

Arif Gasilov, a partner at Gasilov Group, who specialises in natural resources, energy and water, questions how much a per-prompt estimate tells us in the first place.

Servers can handle thousands of requests simultaneously, but they also draw electricity simply by remaining operational. The water attributed to an individual prompt is typically calculated using averages rather than measured directly.

“A lot of the usage comes from it being powered on and active rather than that little bit that gets added on per prompt,” Gasilov says.

Water consumption isn’t the only concern. A data centre drawing from abundant water supplies may place less pressure on its surroundings than one relying on a limited municipal system or a handful of aquifers in a desert community.

“That can have a much, much larger effect on the local community and the local area because it accounts for a lot more of the percentage of the local usage,” he says.

Efforts to reduce water use can also create trade-offs. Gasilov points to closed-loop cooling systems, which reuse water rather than constantly drawing in new supplies. While they can reduce water consumption at data centres, some require more electricity, potentially increasing water use at the power plants supplying them.

He also questions whether improvements in efficiency can keep pace with the rapid growth in AI use.

“I doubt that the efficiency increases will be enough to outweigh more people being on the platform,” he says.