UN Scientists Tie AI Water Use to the Power Grid

UN University scientists put the 2030 water footprint of data-centre electricity at 9.3 trillion litres, equal to the basic yearly needs of 1.3 billion people.

The figure covers water used to cool servers and to generate the power they draw, a bill that moves with the grid mix as much as with the gear on site.

The Water Bill Attached to Every Kilowatt-Hour

The United Nations University Institute for Water, Environment and Health released the study on June 3, 2026, from Richmond Hill, Ontario. The report, Environmental Cost of AI’s Energy Use: Carbon, Water and Land Footprints, was led by Professor Kaveh Madani, the institute’s director, with Dr. Miriam Aczel as lead author.

Global data centres used an estimated 448 TWh of electricity in 2025. Treated as a country, that load would have ranked 11th, behind France and ahead of Saudi Arabia. The International Energy Agency’s Base Case, which the UN team takes as its 2030 path, has that demand double to around 945 TWh by 2030, just under 3 percent of world electricity and nearly triple the combined yearly use of Pakistan, Bangladesh and Nigeria, countries with more than 650 million people.

The IEA puts 2024 use at about 415 TWh, or 1.5 percent of global electricity, after five years of growth at 12 percent a year. From 2024 to 2030 it sees demand rising about 15 percent a year, more than four times the pace of every other sector, even though data centres still account for less than 10 percent of the extra global electricity needed in that window. The United States and China supply nearly 80 percent of the growth: about 240 TWh more in the United States (up 130 percent from 2024) and about 175 TWh more in China (up 170 percent). Europe adds more than 45 TWh, up 70 percent.

US use per person was about 540 kWh in 2024 and is on course to pass 1,200 kWh by 2030, roughly a tenth of a typical American household’s yearly electricity. Africa stays under 1 kWh per person now and under 2 kWh by the end of the decade.

THE 2030 FOOTPRINTS ON THAT ELECTRICITY

The 1.3 billion comparison is a minimum, not a household average. It lines up with WHO guidance that 20 litres per person per day covers essential drinking, cooking and hygiene. Typical city taps run far higher. The rhetorical unit is Sub-Saharan Africa; the water is not being taken from there.

UN scientists say every kilowatt-hour used to train or run an AI system carries a water cost from cooling and from power generation, plus a land cost from energy kit and supply chains. Those three costs do not rise and fall together.

Low-Carbon Power Can Still Run High on Water

Switching from coal to bioenergy can cut the carbon of a kilowatt-hour by 70 percent on average, while raising its water use more than thirty-fold and its land use a hundred-fold, the report finds. A carbon-only scorecard can therefore move the water bill onto regions already short of it.

What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land. If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean but that is solving one problem while creating other problems, often in places that didn’t ask for it.

Dr. Miriam Aczel, lead author, UNU-INWEH

Inside the halls, cooling is only part of the draw. The IEA puts cooling at about 7 percent of electricity in efficient hyperscale sites and over 30 percent in older enterprise rooms, with servers taking about 60 percent. Most of the kilowatt-hours, and therefore most of the UN water number, sit in the IT load and in the plants that feed it.

Viral explainers still treat the 9.3 trillion litres as water poured on chips to make pictures. That is the wrong building. Direct cooling is the fight neighbors can see. The UN total is dominated by the grid behind the fence.

Ireland, Querétaro and a Dry Summer in Uruguay

Madani said the work is a call to handle side effects while the buildout is still young, and to include the places that host mines, halls and e-waste among those who gain. The report’s site notes show how that split already looks on the ground.

WHERE THE DRAW IS ALREADY LOCAL

• Ireland: Data centres took 21 percent of metered electricity in 2023, more than all urban households, and the grid operator has paused new approvals around Dublin until 2028.
• Querétaro, Mexico: Expanding compute is drawing on water supplies during prolonged droughts.
• Uruguay: Plans for a water-heavy data centre coincided with a 2023 drought that depleted Montevideo’s freshwater reserves and made tap water unsafe to drink.

Dr. Mir Matin, who manages UNU-INWEH’s geospatial, climate and infrastructure analytics programme and co-wrote the report, put the pattern in one map problem.

If you map where data centres are getting built against where water stress is worst, you tend to see the same regions in some instances. And the communities living near these sites are not necessarily the ones using the AI being run there. That asymmetry is the issue. Without fixing it, we’ll just be repeating older patterns, where some places carry the costs and other places capture the benefits.

Dr. Mir Matin, co-author, UNU-INWEH

A popular counter is that a cotton shirt uses thousands of litres and fashion dwarfs servers. That comparison is true at planetary scale and still misses the basin. Textile water is spread across farms and dye houses on several continents. A hall in a drought year competes with a city’s taps in one place, in one season, which is why Montevideo’s unsafe tap water is a different kind of fight from a landfill of jeans.

What Closed-Loop Cooling Does Not Change

Operators have cut the water they use on the plot. Microsoft cloud operations executives Judy Priest and Steve Solomon wrote in June 2026 that average water-use effectiveness across owned halls fell from 2.3 litres per kilowatt-hour in the early 2000s to 0.27 litres per kilowatt-hour in 2025, a drop of nearly 90 percent. The company has a separate 40 percent intensity cut by 2030 against a 2022 baseline and said it was 25 percent of the way there by 2025.

Direct air cooling with evaporative assist, in place since 2008, uses water only when outside air is above 85°F (29.4°C). In parts of Northern Europe no water is needed all year. Dublin and Amsterdam use it less than 5 percent of the time, Virginia about 10 percent of the year, and Phoenix as much as 40 percent of the year. About 90 percent of the 2025 owned fleet already runs on low- to zero-water cooling. In 2024 Microsoft introduced a closed-loop, direct-to-chip design that evaporates no water in operation. Phoenix halls improved WUE 23 percent in FY25. Quincy, Washington, uses 74 percent recycled or non-potable water, Singapore 99 percent, and San Antonio 79 percent. New halls in Quebec are expected to collect up to 1.5 million litres of rain a year. Microsoft also said that in FY25 it replenished more water than it withdrew across global operations.

Those are real local gains. They barely touch the UN total, because that total is the water intensity of the electricity, not only the tower next to the rack. A dry-cooled hall on a thirsty thermal grid can still carry a large off-site water bill. A wet tower on a wet hydro system can look worse at the fence and better on the UN ledger. Siting and procurement are the levers the report asks operators and utilities to treat as footprint decisions, with cumulative impact checks rather than a single carbon badge.

Specialized AI Compute, 90 Percent in Two Countries

Only 32 countries host AI-specialised data centres, and more than 90 percent of that capacity sits in the United States and China. More than 150 countries have little or no sovereign AI compute. Professor Tshilidzi Marwala, rector of the United Nations University and a UN under-secretary-general, called that split a governance problem, not a wiring problem, because the minerals and the scrap still land in poorer jurisdictions.

AI-related electronic waste could reach 2.5 million tonnes a year by 2030, equal to discarding nearly 250 Eiffel Towers each year, much of it processed where safeguards are thin. The same two-country concentration shows up in the IEA growth path: almost four-fifths of the extra terawatt-hours to 2030 are American and Chinese. The water and land attached to those kilowatt-hours will follow that map, not the map of ChatGPT users.

Madani, named the 2026 Stockholm Water Prize laureate, told colleagues the window to set rules is narrow. The report asks governments to fold AI halls into energy, water and land permits and to require one standard for carbon, water and land together. It asks operators to treat model defaults and routing as footprint choices, and communities to be in the siting room with grievance rights that can actually halt a draw.

Daily Prompts Now Dwarf the Training Bill

Public argument still fixates on training runs. Training GPT-3 took an estimated 1.3 GWh. Estimates for GPT-4 sit between 50 and 70 GWh. Once a model is live, inference, the answers to ordinary prompts, takes 80 to 90 percent of AI energy, the UN team finds.

ChatGPT alone is estimated at 2.5 billion prompts a day, or about 383 GWh a year for one product. That yearly running cost is several times a single GPT-4 training estimate. The water attached to that one product equals the minimum yearly domestic needs of about 500,000 people in Sub-Saharan Africa.

Task type moves the needle more than brand slogans. A typical chat query uses around 200 times the energy of basic text classification. One AI image uses around 1,450 times that baseline. A short AI video can match 200,000 spam classifications. Model choice, prompt length, output format and resolution all change the bill, and most of those settings are product defaults a user never sees.

Madani warned that cheaper, more efficient models get used more, so per-query savings vanish in volume unless someone caps tokens, resolution or default length. Efficiency without a limit is how the 945 TWh path stays intact even as each answer looks greener on a slide.

Nuclear Plants Still Beat Server Halls in Local Polls

Neighbors are not scoring carbon versus water on a spreadsheet. They are scoring the campus next door. A Gallup survey taken March 2 to 18, 2026, found that 71 percent oppose a local AI data centre, including 48 percent who are strongly opposed. Only 7 percent are strongly in favor, and about a quarter favor the projects at all. In the same survey, 53 percent opposed a local nuclear plant. Since Gallup first asked the nuclear question in 2001, opposition has never topped 63 percent. Jeffrey M. Jones wrote up the results for the firm.

THE MARCH 2026 LOCAL BUILD QUESTION

• AI data centres: 71 percent oppose a local build, 48 percent strongly.
• Strongly in favor: 7 percent, outnumbered on the strong-oppose side by about seven to one.
• Nuclear plants: 53 percent oppose a local plant, below the 63 percent peak since 2001.
• Why it bites: Gallup tied the new hostility to land, power, water for cooling, and local electric bills.

That ranking collides with the UN trade-off. A nuclear-backed hall can cut carbon and still raise the water attached to each kilowatt-hour, depending on how the plant is cooled. A gas-backed hall can look worse on carbon and lighter on water. Voters are rejecting the visible campus while the 9.3 trillion litres hide in the generation stack they just ranked as more acceptable.

Ireland’s freeze on new Dublin-area approvals runs through 2028. Until carbon, water and land are disclosed on the same sheet, and until siting tests the basin as well as the plug, the 1.3 billion-person comparison will keep being hung on the wrong fence.