Researchers have uncovered the staggering environmental toll of artificial intelligence, revealing that generating a single 100-word email consumes approximately 519 millilitres of water.
A 2025 peer-reviewed paper titled “Making AI Less Thirsty,” published in Communications of the ACM, outlines these findings, reports Space Daily.
The research, conducted by Pengfei Li, Shaolei Ren, and their colleagues at the University of California, Riverside, establishes a new methodology for estimating the “per-query water footprint” of large language models.
The study clarifies that the 519-millilitre figure roughly equivalent to a standard bottle of water includes more than just the direct liquid used to cool data centre servers. It also accounts for the “indirect water” required to generate the massive amounts of electricity these servers demand.
The environmental impact intensifies as user engagement grows. While a single response triggers the half-litre cost, the research group estimates that a sustained conversation consisting of ten to fifty exchanges falls within the same 500-millilitre order of magnitude.
However, the water consumption scales by a factor of one every time a user extends the conversation further.
As AI integration becomes ubiquitous in daily digital tasks, these findings highlight a growing ecological challenge for the technology sector.
Why AI requires water
The most widely used method for cooling data centres is evaporative cooling. In this process, water circulates through pipes running close to heat-generating equipment, absorbs thermal energy, and is then released into the air where part of it evaporates, carrying heat away as vapour.
Around 80 per cent of the water used in such systems is lost through evaporation, while the remainder returns to local water networks, sometimes at elevated temperatures and with chemical traces from cooling treatments.
New-generation AI-focused data centres are larger, denser, and significantly more thermally demanding than traditional cloud facilities built in the 2010s. A single hyperscale AI campus can consume more water in a day than a town of 10,000 people uses collectively for drinking, cooking, sanitation, and agriculture.
Water use rising sharply across tech giants
According to the latest Environmental Report from Google covering the 2024 financial year, total water consumption reached about 8.1 billion gallons, with roughly 95 per cent used in data centres.
This marks an 8 per cent rise from 2023, which itself was up 17 per cent from 2022, and 2022 rose 20 per cent from 2021.
Overall, Google’s water use nearly doubled between 2021 and 2024, with the company attributing the surge mainly to AI workload expansion.
Microsoft reported similar trends, though on a smaller scale, with about 1.7 billion gallons of water consumed in 2022—up 34 per cent year-on-year.
Investigations into its West Des Moines, Iowa data centre cluster, where GPT-4 training occurred in 2022, found that a single training run used 11.5 million gallons in July and 13.4 million gallons in August that year.
The same cluster has since expanded to five facilities, collectively drawing around 68.5 million gallons annually from the local municipal system, making it the region’s largest industrial water user.
Meta reported around 813 million gallons of global water use in 2023, with 95 per cent consumed in data centres.
Amazon, which operates the world’s largest cloud infrastructure network, does not publish total water consumption figures.
A 2024 report by the Lawrence Berkeley National Laboratory, prepared for the US Department of Energy under the Energy Act 2020, estimated that US data centres directly consumed around 17.4 billion gallons of water for cooling in 2023.
It also found that indirect water use linked to electricity generation reached about 211 billion gallons—around twelve times higher than direct consumption. The report warned that direct use could double or quadruple by 2028, with indirect demand rising in parallel.
Global pressure on freshwater systems
The study projects that global AI demand could require between 4.2 and 6.6 billion cubic metres of water withdrawals annually by 2027.
That lower estimate equals roughly four times Denmark’s annual water withdrawals, while the upper estimate approaches half of the United Kingdom’s total annual use. The projections assume current growth trends and do not account for faster-than-expected AI expansion.
The study also highlights that water must be sourced locally, meaning regional pressure varies significantly. World Resources Institute classification systems show that in 2023, about 42 per cent of Microsoft’s water use came from water-stressed regions, while Google reported around 15 per cent of freshwater withdrawals from areas of high scarcity. Both figures are expected to rise if current trends continue.
Local impacts in drought-hit regions
Several regions already face visible strain from expanding data centre infrastructure.
In Chile, Google paused its planned $200 million data centre project in Cerrillos near Santiago in 2024 after a court ruled that environmental impacts on the Central Santiago Aquifer had not been properly assessed. The country has endured a drought lasting over 15 years, with water rationing introduced in 2022.
In Querétaro, Mexico—where 32 new data centres are planned—severe drought in 2024 affected 17 of 18 municipalities. Microsoft has secured rights to about 25 million litres of water annually from a local aquifer already facing a 60-million-litre deficit.
Uruguay, also experiencing its worst drought in 70 years, is expected to host a Google data centre in Canelones that could consume around 7.6 million litres of water daily in its first phase—equivalent to the daily needs of about 55,000 people.
In the United States, a $14 billion data centre project in Arizona was withdrawn in 2024 after local opposition led to rezoning denial amid water concerns. In Spain’s Aragón region, multiple projects are advancing despite ongoing disputes over agricultural water rights.
Transparency gaps in reporting
Industry water disclosures remain inconsistent and incomplete. The main gaps include differences between withdrawal and consumption, limited reporting of indirect water use linked to electricity generation, and a lack of facility-level breakdowns.
Withdrawal refers to total water drawn from sources, while consumption reflects water lost permanently through evaporation. Depending on reporting methods, this alone can change estimates by a factor of three.
Indirect water use, largely excluded from corporate disclosures, can be up to twelve times higher than direct cooling use, according to the Lawrence Berkeley analysis. Facility-level data is also rarely disclosed, making it difficult to assess local impacts in water-stressed areas.
Companies cite methodological complexity, competitive concerns, and reputational risk as reasons for limited transparency.
Growing infrastructure and uncertain balance
The AI sector is expanding at unprecedented speed, with McKinsey estimating global investment in AI infrastructure could reach around $5.2 trillion by 2030. These facilities operate, in effect, as large-scale industrial cooling systems housing vast computing power.
Each individual AI query uses minimal water, but cumulative demand is accelerating rapidly.
With global freshwater scarcity worsening and around one-quarter of the world’s population expected to face severe water stress by 2030, according to United Nations projections, competition between digital infrastructure and human water needs is becoming increasingly acute.
AI technologies may also help address water challenges through improved climate modelling, irrigation efficiency, weather forecasting, and drought management. However, whether such benefits scale fast enough to offset consumption remains uncertain.
On current trends, the outcome remains unresolved. What happens by 2027 will depend on decisions already being made in corporate boardrooms, government policy forums, and local planning authorities.
