Each time somebody makes use of ChatGPT to write down a 100-word electronic mail, roughly 519 millilitres of water is consumed, nearly equal to an ordinary water bottle.
That estimate comes from a peer-reviewed 2025 paper printed in Communications of the ACM by Pengfei Li, Shaolei Ren and their colleagues on the College of California, Riverside.
The research accounts for each the direct water used to chill information centre servers and the oblique water required to generate the electrical energy that powers them.
When that is scaled to hundreds of thousands of customers making a number of queries on daily basis, the numbers change into staggering.
By 2027, international AI infrastructure is projected to eat between 4.2 billion and 6.6 billion cubic metres of water yearly. That’s practically equal to half of the UK’s whole annual water withdrawal.
Extra worryingly, a lot of this water is being drawn from areas which are already dealing with water stress.
Why AI Knowledge Centres Eat So A lot Water
Knowledge centres generate monumental warmth. The highly effective chips that run trendy AI methods, particularly high-end graphics processing items, can every eat and launch between 300 and 700 watts of warmth whereas working underneath heavy load.
To manage this warmth, many information centres depend on evaporative cooling. On this system, water is pumped by means of the ability to soak up warmth from servers. A portion of that water then evaporates into the environment.
Round 80 p.c of the water drawn into an evaporative cooling system is completely misplaced by means of evaporation.
The remaining water might return to the system, however usually at increased temperatures and with chemical residues.
The brand new technology of AI-focused hyperscale information centres is bigger, denser and way more heat-intensive than the overall cloud infrastructure constructed within the 2010s.
A single massive information centre campus can now eat extra water in a day than a city of 10,000 folks makes use of for ingesting, sanitation, cooking and agriculture mixed.
In keeping with the 2024 US Knowledge Centre Vitality Utilization Report by Lawrence Berkeley Nationwide Laboratory, ready for the US Division of Vitality, information centres consumed round 17.4 billion gallons of water instantly for cooling in 2023.
One other 211 billion gallons had been consumed not directly by means of electrical energy technology.
The report additionally famous that information centre load development has tripled over the previous decade and is projected to double or triple once more by 2028.
Google, Microsoft And Meta’s Water Use
Main expertise corporations have began disclosing their water consumption in annual sustainability experiences. The development is obvious: water use is rising.
Google’s 2024 Environmental Report stated the corporate consumed round 8.1 billion gallons of water in the course of the yr, with about 95 p.c used at information centres. This marked an 8 p.c enhance from 2023.
The earlier years had additionally seen sharp will increase, with Google’s water consumption practically doubling in three years.
Microsoft has reported smaller numbers, however the sample is analogous. The corporate consumed round 1.7 billion gallons of water in 2022, a 34 p.c enhance from the earlier yr.
Unbiased reporting on Microsoft’s information centre cluster in West Des Moines, Iowa, the place GPT-4 coaching runs had been performed in 2022, discovered {that a} single coaching run consumed 11.5 million gallons of water in July 2022 and 13.4 million gallons in August.
The identical cluster has since expanded to 5 amenities, drawing 68.5 million gallons yearly from the native municipal water system.
Meta consumed round 813 million gallons of water globally in 2023. Amazon, which operates the world’s largest cloud infrastructure, doesn’t publish mixture water consumption figures.
AI Is Increasing In Water-Burdened Areas
The Li and Ren paper initiatives that by 2027, international AI demand might result in water withdrawal equal to greater than 4 occasions Denmark’s annual utilization, or practically half of the UK’s whole yearly withdrawal.
The issue just isn’t solely the quantity of water used, but in addition the place that water is coming from.
Microsoft acknowledged in its 2023 sustainability report that round 42 p.c of its water consumption got here from areas categorised as water-stressed underneath the World Sources Institute’s ranking system.
Google reported that 15 p.c of its freshwater withdrawals in 2023 got here from areas dealing with excessive water shortage.
The results are already seen. In Chile, Google paused a deliberate $200 million information centre close to Santiago after an environmental courtroom dominated that the corporate had not adequately assessed the affect on the Central Santiago Aquifer.
The nation had been dealing with drought for 15 years and had begun rationing residential water in 2022.
In Querétaro, Mexico, the place 32 new information centres are deliberate, the state skilled its worst drought in a century in 2024.
Microsoft has secured rights to round 25 million litres of water yearly from a neighborhood aquifer that’s already operating a 60-million-litre annual deficit.
In Arizona, a $14 billion information centre undertaking was withdrawn in 2024 after native residents opposed the rezoning.
What Corporations Are Not Disclosing
The out there figures are primarily based solely on what corporations have chosen to reveal. The precise water footprint of the AI business is probably going a lot bigger.
There are three main disclosure gaps.
The primary is the distinction between water withdrawal and water consumption. Withdrawal refers back to the whole quantity of water taken from a supply, whereas consumption refers back to the water completely misplaced, principally by means of evaporation.
Corporations usually report solely certainly one of these figures, which might make their footprint seem a lot smaller.
The second hole is the distinction between direct cooling water and oblique water used for electrical energy technology.
The Li and Ren analysis estimates that oblique water use might be round 12 occasions increased than direct cooling water use. Only a few company experiences embody this determine.
The third hole is the absence of facility-level information. An organization-wide annual whole doesn’t inform native communities whether or not their very own aquifer or water system is underneath strain.
The UC Riverside paper is essential as a result of it makes use of publicly out there data to estimate these hidden gaps.
As unbiased researchers produce credible estimates, expertise corporations might face rising strain to reveal extra detailed water information.
The Huge Query For AI
The worldwide infrastructure for AI is being constructed at a rare tempo. Behind the thrill over synthetic intelligence are large industrial amenities that rely closely on cooling methods and electrical energy.
Every particular person AI question could seem small. However when multiplied by billions of interactions, the environmental value turns into vital.
Supporters argue that AI might assist clear up local weather and water issues by means of higher local weather modelling, improved irrigation methods and extra correct drought prediction.
However the actual query is whether or not AI can ship these advantages sooner than its personal water consumption grows.
For now, that query stays unanswered.
