The figure for a single email comes from a 2025 peer-reviewed paper in Communications of the ACM by Pengfei Li, Shaolei Ren, and colleagues at the University of California, Riverside. The paper, titled “Making AI Less Thirsty,” sets out the methodology by which the per-query water footprint of large language models can be estimated. The figure for the 100-word email is approximately 519 millilitres, which is close enough to the volume of a standard bottle of water for the bottle to be the practical comparison. The number includes both the direct water used to cool the data centre’s servers and the indirect water used to generate the electricity those servers consume.
The 519 millilitre figure assumes a single response. Most users do not send a single response. They have conversations.
The same research group estimates that a single sustained conversation with a chatbot, defined as somewhere between ten and fifty exchanges, consumes approximately the same 500-millilitre order of magnitude. The figure scales by a factor of one each time the conversation extends.
Why AI needs water at all
Data centres generate heat. The servers processing AI queries are essentially small radiators running at high intensity for as long as the workload continues. The chips at the heart of contemporary AI training and inference, the high-end graphics processing units manufactured primarily by Nvidia, dissipate between 300 and 700 watts each, depending on the model. A single training run for a large language model uses tens of thousands of these chips simultaneously, for weeks or months at a time. The heat has to go somewhere.
The most common method for moving that heat out of a data centre is evaporative cooling. Water is pumped through pipes that run alongside or directly across the heat-producing equipment, absorbs the heat, and is then exposed to the air. A portion of the water evaporates, carrying the heat into the atmosphere as water vapour. The remaining water cycles back through the system. Approximately 80 per cent of the water drawn into an evaporative cooling system is lost to evaporation. The rest returns to local water systems, sometimes at higher temperatures and with chemical residues from the cooling process.
The newer generation of data centres built specifically for AI workloads are larger, more dense, and more thermally intense than the data centres built for general cloud computing in the 2010s. A single large hyperscale AI campus can now consume more water in a day than a town of ten thousand people uses for everything: drinking, washing, cooking, sanitation, agriculture, and irrigation combined.
This video explains exactly what big tech promised and how AI is doing the opposite of that.
How much, in actual numbers
Google’s most recent Environmental Report, covering the 2024 financial year, sets out the water consumption of the company’s global operations in detail. The combined figure for 2024 was approximately 8.1 billion gallons, of which approximately 95 per cent was used at data centres. The 2024 figure was an 8 per cent increase on 2023. The 2023 figure had been a 17 per cent increase on 2022. The 2022 figure had been a 20 per cent increase on 2021. The cumulative result is that Google’s water consumption nearly doubled between 2021 and 2024, with the company itself naming AI workload growth as the primary driver in successive environmental reports.
Microsoft’s figures are similar in shape, smaller in absolute scale. The company reported water consumption of approximately 1.7 billion gallons in 2022, a 34 per cent year-on-year increase. The growth has continued. The independent investigative reporting on Microsoft’s data centre cluster in West Des Moines, Iowa, where the GPT-4 training runs were conducted in 2022, has documented that a single training run consumed 11.5 million gallons of water in July 2022 and another 13.4 million gallons in August. The same cluster has, in subsequent years, expanded to five separate facilities collectively drawing 68.5 million gallons annually from the West Des Moines municipal water system, more than any other industrial user in the metropolitan area.
Meta consumed approximately 813 million gallons globally in 2023, with 95 per cent of that volume used at data centres. Amazon, which operates the largest cloud infrastructure in the world, does not publish aggregate water consumption figures.
The Lawrence Berkeley National Laboratory’s 2024 Data Center Energy Usage Report, prepared for the United States Department of Energy under the Energy Act of 2020, estimated that data centres in the United States consumed approximately 17.4 billion gallons of water directly through cooling in 2023. The same report estimated that an additional 211 billion gallons of water were consumed indirectly through the electricity required to power the same data centres. The indirect figure is approximately twelve times larger than the direct figure. The report projects that the direct figure could double or quadruple by 2028. The indirect figure scales in the same proportion.
Where the water comes from
The Li and Ren paper projects that global AI demand will require somewhere between 4.2 and 6.6 billion cubic metres of water withdrawal annually by 2027. The lower estimate is approximately the total annual water withdrawal of four Denmarks. The higher estimate approaches half the total annual water withdrawal of the entire United Kingdom. Both estimates assume current trajectories of AI workload growth and current water-efficiency practices. Neither estimate accounts for the possibility that AI demand continues to grow faster than the modelled trajectory.
The water has to come from somewhere. In Microsoft’s 2023 sustainability report, the company acknowledged that approximately 42 per cent of its water consumption that year came from regions classified as “water-stressed” under the World Resources Institute’s standard rating system. Google’s equivalent figure for 2023 was 15 per cent of freshwater withdrawals from regions of “high water scarcity.” Both figures, on the trajectory of the past three years, are likely to increase rather than decrease.
The concrete consequences of those abstract percentages are now visible in specific locations. In September 2024, Google announced it was pausing its planned 200-million-dollar data centre in Cerrillos, near Santiago, Chile, after a Chilean environmental court partially reversed the project’s original 2020 permit. The court ruled that the company had not adequately accounted for the impact on the Central Santiago Aquifer in a country that had been in a continuous drought for fifteen years and had begun rationing residential water in 2022. The project is now under revision.
In Querétaro, Mexico, where 32 new data centres are currently planned, the state suffered its worst drought in a century in 2024, with seventeen of eighteen municipalities affected and the drinking water supply for thousands of families at risk. Microsoft has secured rights to approximately 25 million litres of water annually from a local aquifer that is currently running a 60-million-litre annual deficit. In Uruguay, currently experiencing its worst drought in 70 years, Google’s proposed data centre in Canelones would, in its first operational phase, consume approximately 7.6 million litres of water per day, equivalent to the daily residential water needs of 55,000 people. In Goodyear and Buckeye, Arizona, a 14-billion-dollar data centre project was withdrawn in 2024 after local resident organisations successfully pressed elected officials to deny the necessary rezoning. In Aragón, Spain, multiple data centre projects are advancing in regions where agricultural water rights are already contested.
The pattern, on the available evidence, is that the cooling infrastructure for global AI is being built preferentially in regions where freshwater is cheap, regulatory oversight is loose, and the local population is least positioned to negotiate.
What companies don’t disclose
The figures cited above are the figures the companies have made public. The full water footprint of the AI industry is, by every available assessment, larger than the figures voluntarily disclosed in sustainability reports.
Three specific gaps recur across the disclosure landscape. The first is the gap between water withdrawal, which is the volume drawn from local sources, and water consumption, which is the volume permanently lost to evaporation. Most corporate reports name only one of these figures, and the choice between them can shift the apparent footprint by a factor of three or more depending on which is reported. The second is the gap between direct cooling water and indirect electricity-generation water. Almost no corporate report includes the indirect figure, despite the Lawrence Berkeley estimate that the indirect figure is approximately twelve times the direct one. The third is the gap between aggregate global figures and facility-level figures. A company-wide annual total tells a stakeholder nothing about whether the company’s data centre in a drought-stressed Arizona town is straining the local aquifer.
The reasons for the disclosure gaps are several. Some are methodological: the per-facility water footprint of a data centre depends on cooling technology, local climate, electricity-grid mix, and seasonal demand variation, none of which the company necessarily measures with precision. Some are competitive: detailed facility-level water disclosure could give competitors useful intelligence about a company’s infrastructure plans. Some are reputational: a company that discloses its full water footprint and is then criticised for the size of it is exposed to public-relations risk in a way that a company reporting only aggregate figures is not.
The Li and Ren paper’s contribution to the literature is, in significant part, that it produces credible estimates of the gaps. The figures that the AI industry has not been willing to publish are figures that academic researchers, using publicly available proxies for cooling efficiency and electricity-grid water intensity, are now able to estimate within reasonable bounds.
What is at stake
The global infrastructure for processing AI queries is being built faster than any new technology infrastructure in modern history, on a financing trajectory that McKinsey has projected at approximately 5.2 trillion US dollars by 2030. The physical buildings the trillion-dollar investment is producing are, in their fundamental operational requirements, large industrial-scale evaporative cooling systems with computing equipment inside them.
Each query is small. The aggregate is not.
Half the United Kingdom’s annual water withdrawal, evaporating into the atmosphere from cooling towers across the world’s data centres by 2027, is not a marginal correction to a global water balance that is otherwise stable. Global freshwater scarcity is increasing on every measured trajectory. Approximately one-quarter of the world’s population, by United Nations projections, will face severe water stress by 2030. The water the AI industry is now drawing from aquifers, rivers, and reservoirs, increasingly in the regions least able to spare it, is competing directly with that population.
The technologies the AI industry is developing have, by any reasonable analysis, the potential to contribute to solving some of the same water-management problems they are now exacerbating, through better climate modelling, more efficient irrigation, more accurate weather prediction, and more sophisticated drought response. Whether the contribution arrives at scale faster than the consumption does is the open question that determines whether the trade-off, on the long view, is worth it.
On the present trajectory, the answer is unclear.
What the trajectory will look like by 2027 depends on decisions being made, in board rooms and government offices and local zoning meetings, now.
