A medium-length AI chat exchange consumes roughly 10 to 50 millilitres of water. That figure comes from a peer-reviewed paper whose authors were careful to say it holds only “depending on when and where” the model runs. Strip out the condition and you get the headline everyone repeats. Keep it and you get the actual subject, which is that AI’s water footprint is a local and seasonal quantity being argued about with national annual averages.
Two confident camps have formed. One says every prompt drinks a bottle of water. The other says the concern is manufactured, because data centres are a rounding error next to agriculture and golf courses. I spent several weeks with the underlying research instead of the coverage, including two papers published in June 2026 that change the picture, and I came out holding a less comfortable position than either camp offers.
This piece covers what the evidence supports, where the honest uncertainty sits, and what a UK organisation adopting AI should take from it. Energy and carbon are handled separately in the per-vendor audits of Anthropic, OpenAI and xAI.
Key takeaways
- Around three-quarters of AI’s water is consumed at power stations, not inside the data centre. Water and energy are the same question.
- National averages cannot answer this. Fixed national factors misestimate water stress by up to 75% low and 999% high depending on region.
- The sceptics are broadly right about US national totals and broadly wrong that this settles the question.
- The UK is a different problem. 84% of proposed data centres sit in water-stressed areas, and water companies’ 2050 plans do not count them at all.
- Only two-fifths of operators measure their water use, so the alarm and the dismissal rest on data that mostly does not exist.
Where the water actually goes
Most of the confusion comes from four words being used interchangeably when they mean different things. Getting them straight resolves more than any single statistic.
| Term | What it means | Why it matters |
|---|---|---|
| Withdrawal | Water taken from a source | Large numbers, mostly returned. Most alarming headlines use this. |
| Consumption | Water evaporated and not returned | The number that affects local availability. |
| Onsite (Scope 1) | Cooling inside the building | Roughly a quarter of the total. Often potable. |
| Offsite (Scope 2) | Evaporated generating the electricity | Roughly three-quarters of the total. Usually invisible in reporting. |
That last split is the one people miss. Mapping 472 US hyperscale facilities against grid regions and river basins, Guidi and Dominici at Harvard put total operational consumption at roughly 300 gigalitres a year, with about three-quarters attributable to electricity generation, not cooling.
The practical consequence is that a data centre’s water footprint is decided substantially by the grid it plugs into. A facility on fossil-heavy power has a different water profile from an identical building on wind, whatever its cooling design. Water efficiency and energy efficiency are the same project, which is why the workflow-level compute reductions I have measured in my own practice show up on both ledgers at once.
The bottle of water, and what its source actually says
The half-litre figure comes from “Making AI Less Thirsty” by Li, Yang, Islam and Ren, published in Communications of the ACM. It estimates that training GPT-3 consumed around 5.4 million litres, and that inference costs roughly 500ml per 10 to 50 medium-length responses.
The range is a factor of five, and the authors attribute it to when and where the model runs. Same model, same prompt, five times the water depending on grid mix, cooling design and season. The number was never a constant, and quoting it as one is how a reasonable estimate became a misleading meme.
The case that this is overblown
The most rigorous sceptical account is Andy Masley’s “The AI water issue is fake”, which deserves engaging with properly. His argument runs roughly as follows.
Composition first. Around 90% of AI’s water is withdrawn non-consumptively offsite and returned, about 7% is consumed offsite, and only about 3% is potable water consumed onsite. Then scale. All US data centres consumed 200 to 250 million gallons a day in 2023, around 0.2% of national freshwater consumption, with AI’s share at roughly 0.008%. Even a tenfold increase leaves it near 0.08% by 2030.
Then the local test. In Maricopa County, Arizona, the most water-stressed county hosting significant data centre capacity, he calculates data centres at 0.12% of county water against golf courses at 3.8%. He documents cases where data centre fees funded municipal water upgrades, in The Dalles, Council Bluffs, Quincy and Goodyear. And he catalogues recurring distortions in media coverage, including a widely shared story about wells drying near a Meta facility where the cause was construction sediment rather than operational water use.
Much of this holds up. The media criticism is largely fair, the composition analysis matches the academic literature, and the US national arithmetic is not seriously disputed. He also states his limits plainly, conceding that construction can cause local groundwater problems, that siting in water-scarce regions raises real policy questions, and that he does not address non-US contexts in depth.
Why national averages cannot settle a local question
The weakness in the sceptical case is not its arithmetic. A national annual average is the wrong instrument for measuring something local and seasonal, and the research published since that article quantifies how wrong it gets.
In June 2026, Talukder, Ren, Islam and colleagues analysed 4,147 US data centre locations using county-level monthly water-stress weightings instead of national factors, tracing both onsite cooling and the electricity supply behind each site. Three findings matter.
- Fixed national factors misestimate stress-adjusted water by up to 75% low in some regions and 999% high in others. That is not a margin of error. It means aggregate reasoning produces the wrong ranking of which facilities cause harm.
- Roughly 75% of US data centres experience at least one month of elevated water stress, and about 30% hit extreme stress in at least one month. An annual average smooths away precisely the periods when availability is binding.
- Minimising volumetric water can increase actual harm, by shifting load toward a more water-stressed electricity supply. Optimising the headline number and optimising the outcome are different operations.
This is consistent with earlier peer-reviewed work. Siddik, Shehabi and Marston found in Environmental Research Letters that the US data centre water scarcity footprint is more than double its volumetric footprint, meaning the sector draws disproportionately from stressed watersheds. Around a fifth of direct water came from moderately to highly stressed watersheds, and nearly half of servers were powered by plants sitting in water-stressed regions.
Guidi and Dominici add the geographic concentration. Just 3 of 24 balancing authorities account for 59% of all electricity-related water use. Diffuse at national level and concentrated where it lands are both true statements about the same system.
Claim by claim
| Sceptical claim | What the evidence supports | What it misses |
|---|---|---|
| AI is ~0.008% of US freshwater | Broadly correct as a national figure | Water scarcity is a basin-level and monthly property |
| Most water is withdrawn, not consumed | Correct, and important | Consumption still concentrates in specific stressed basins and grid regions |
| Data centres use less than golf courses | True in Maricopa County | Golf being worse does not make a second demand costless in a constrained basin |
| No documented harm to local supply | Fair for operational US cases so far | Absence of documented harm is weak evidence when most operators measure nothing |
| Media coverage is misleading | Largely correct | Bad journalism about a subject does not resolve the subject |
| Treating more water is cheap, so demand helps | Treatment economics are real | Treatment cost is not the binding constraint where the raw resource is short |
AI Water Usage in the UK
The sceptical case is explicitly about the United States, and the UK inverts most of its premises. A report published on gov.uk, lead-authored by Rich Kenny of the GDSA Planetary Impact Working Group, sets out the position.
| United States | United Kingdom | |
|---|---|---|
| Headroom | Large national freshwater base | Projected shortfall of nearly 5 billion litres a day by 2050, over a third of current public supply |
| Siting | Mixed, some in stressed basins | 84% of proposed water-intensive data centres in areas stressed now or by 2040 |
| Planning | Contested locally | Water companies’ 2050 deficit plans do not account for data centres at all |
| Measurement | Partial, Google most transparent | No reliable national dataset. Two-fifths of operators track water use |
The concrete version is Culham in Oxfordshire, the first government-designated AI Growth Zone, sited about seven miles from a planned reservoir at Abingdon intended to supply the Thames Valley, London and Hampshire. Thames Water, the utility for that area, is classified as seriously water stressed and estimates that data centres could account for around 30% of all new water demand in its region, roughly 270 million litres a day.
Scale that to a single building. A 100MW hyperscale facility consumes around 2.5 billion litres a year, equivalent to the needs of about 80,000 people. That is a meaningful number in a country where water scarcity is already estimated to be holding up housing development at a cost of around £25 billion over five years, and where drought was declared across half of England in 2026.
The Environment Agency has described future demand from AI and data centres as “highly uncertain” and has asked operators to forecast their own consumption. That is a regulator saying, in the politest available language, that it does not have the data it needs to plan.
What nobody can measure
Underneath both camps sits the same hole. Research by McCauley and Scanlan at the University of Wisconsin-Milwaukee, forthcoming in the Rutgers Computer and Technology Law Journal and summarised in The Conversation, compared what the major operators actually publish.
| Operator | What is disclosed |
|---|---|
| Amazon | No water usage disclosed |
| Microsoft | Company-wide operations, no data centre breakdown |
| Meta | Company-wide aggregate, with data centres separated |
| Individual figures per data centre, the most transparent of the four |
Even the best of these omits what is needed to interpret the number: facility size, cooling technology, power source, seasonal variation, and indirect water from electricity generation, which the same researchers estimate at around twelve times direct use. David Mytton reported in npj Clean Water in 2021 that fewer than a third of operators measured water consumption at all. Five years later the UK figure is two-fifths. That is the pace of improvement.
Which is why I find confident dismissal as unpersuasive as confident alarm. Both extrapolate from a dataset that the regulators, the academics and the government’s own report agree does not exist.
Where I land
I build AI products and run my consulting practice on these tools, so I have an obvious interest in the comfortable answer. I have tried to argue against it instead.
The transparency gap is the story. Not the bottle of water, not the golf course comparison. When two-fifths of operators measure their own consumption and one of the largest publishes nothing, the honest statement is that nobody can currently tell you whether a given data centre is a problem. Mandatory location-specific reporting, of the kind the EU rating scheme already requires and the gov.uk report recommends for UK sites above 1MW, is the intervention that makes every other argument possible.
Efficiency comes before supply, including renewable supply. The standard industry answer is that renewables solve the water problem, since wind and solar consume almost none in operation. That is true, and I still do not fully buy it as a strategy. Building more supply to meet unconstrained demand pushes the impact upstream into mining and materials. Reducing demand first is the option that avoids relocating the damage, and it is available now, at the workflow level, to anyone using these tools.
Local beats national, and the UK case is not the US case. A technology can be immaterial nationally and genuinely damaging in the basin where it lands during the month that matters. The 2026 research puts numbers on how badly aggregate reasoning fails, and the UK’s combination of water stress, siting patterns and planning blind spots makes that failure mode more likely.
I remain conflicted about working in this space, and the concentration of power in a handful of labs worries me at least as much as the resource question. I have landed on the view that I can do more by staying in it with my eyes open than by looking away. That is a position I hold provisionally, and would revise if the evidence moved.
What to do with this
If you are adopting AI in an organisation that takes its environmental commitments seriously, the water question resolves into four practical items.
- Ask vendors where the workload runs, and on what grid. Location and power source drive most of the footprint. A vendor who cannot answer is telling you something.
- Ask for location-specific water reporting, not company-wide totals. Aggregates are unfalsifiable by design.
- Reduce compute before offsetting anything. Caching, smaller models for routine tasks and reusable workflow components cut the bill on every ledger at once.
- Treat local context as the deciding factor. The same deployment carries different consequences in Slough and in Stockholm.
None of that requires resolving the argument. It requires refusing to accept either camp’s confidence as a substitute for data we do not yet have.
Where this belongs organisationally is inside the governance perimeter, alongside the confidentiality and accuracy rules, which is how I structure it in AI training for leadership teams. If you want help putting that perimeter in place, our services page explains how we engage.
Frequently asked questions
How much water does one AI prompt use?
Roughly 500ml per 10 to 50 medium-length responses according to the peer-reviewed estimate in Communications of the ACM, so somewhere between 10ml and 50ml per exchange. The fivefold range depends on where and when the model runs. Most of that water is evaporated at the power station rather than in the data centre.
Is AI’s water use actually significant?
Nationally in the United States, no. AI accounts for well under 0.1% of freshwater consumption. Locally and seasonally it can be significant, because around 75% of US data centres face at least one month of elevated water stress and about 30% face extreme stress in some month. Both statements are true, which is why the debate persists.
Do data centres use drinking water?
Frequently yes, because cooling systems need high-purity water to avoid corrosion and bacterial growth, and mains water is the cheapest way to get it. Roughly 57% of direct data centre water was potable in Mytton’s 2021 analysis. Reclaimed and non-potable sources are growing, with Google using reclaimed water at around a quarter of its campuses, though adoption remains partial.
Does using less water always help?
No. Air and dry cooling reduce onsite water while increasing electricity demand, which increases offsite water and carbon. The 2026 research shows that minimising volumetric water can increase stress-adjusted harm if it shifts load onto a more water-stressed grid. The right choice depends on the local grid and the local basin.
What would actually fix this?
Mandatory location-specific reporting is the precondition for everything else. After that, siting decisions that treat water stress as a constraint rather than an afterthought, water-efficient cooling required in stressed regions, and demand reduction pursued before supply expansion.