AI Data Centers’ Water Use Puts Host Cities on the Hook

AI data centers’ water use is now an end-of-decade planning problem: a United Nations University report projects the global data centers powering artificial intelligence will carry a 9.3 trillion-liter water footprint by 2030, enough to match the basic annual domestic needs of 1.3 billion people in Sub-Saharan Africa. The report also puts data center electricity demand at 945 terawatt-hours (TWh) by that year, close to 3% of projected world power use.

The local bill sits in the towns and watersheds that host the machines. The authors point to Ireland, Querétaro in Mexico and Uruguay as places where grids or water supplies have already been tested by data center growth, then ask regulators to measure carbon, water and land together before permits are granted.

Host Communities Enter the AI Ledger

The United Nations University Institute for Water, Environment and Health (UNU-INWEH, the UN university’s water research institute) published the AI carbon, water and land footprints report on June 3. The publication page lists Aczel, Chamanara, Matin, Farsi, Marwala and Madani as authors.

Kaveh Madani, UNU-INWEH’s director, led the investigation team, and Miriam Aczel, a UNU-INWEH researcher, is named by the institute as the lead author. Their release says global data centers consumed an estimated 448 TWh of electricity in 2025. Treated as a country, that level would have ranked 11th in the world, behind France and ahead of Saudi Arabia.

The global system building artificial intelligence must also govern it sustainably and fairly.

Professor Tshilidzi Marwala, United Nations University rector and UN under-secretary-general, said that in the release. His comment appears in the section on compute concentration, where the report says the countries hosting AI infrastructure and the communities carrying the resource costs can be different groups.

Power Demand Sets the Water Math

The power figure has a second source. The International Energy Agency (IEA, the Paris-based intergovernmental energy body) said in its Energy and AI demand analysis that data centers used 415 TWh in 2024, about 1.5% of global electricity consumption, and are projected to reach 945 TWh by the end of the decade in its base case.

The agency says the United States accounted for 45% of global data center electricity consumption in 2024, followed by China at 25% and Europe at 15%. It also says half of U.S. data centers under development are in existing large clusters, where grid queues and local bottlenecks can decide project timing.

Oton Technology has tracked the squeeze in Europe’s AI power grid bottleneck, where project timing depends on clean megawatts reaching compute campuses.

Inference Makes the Prompt a Load Source

UNU-INWEH points to everyday use after a model ships. The report says inference, the running of deployed models in response to user prompts, accounts for 80 to 90% of total AI energy use. It estimates ChatGPT, OpenAI’s chatbot, handles about 2.5 billion prompts a day, translating to roughly 383 gigawatt-hours (GWh, one billion watt-hours) a year.

Format changes the load. The report compares common AI tasks with basic text classification, a low-energy baseline such as spam sorting:

• Text chat – about 200 times the energy of basic text classification for a typical conversational query.
• AI images – about 1,450 times the same baseline for a generated image.
• AI video – roughly the electricity of 200,000 spam classifications for a single short generated clip.

At the product layer, design reaches the water system. Longer default answers, larger images and video generation push more work into data centers; facilities then need power, cooling and backup capacity sized for the load users create at scale.

Local Opposition Has a Water Meter

Gallup put the local politics in plain numbers. In its survey of U.S. adults on AI data centers, conducted March 2 to 18 with 1,000 adults, 71% opposed building one in their area, including 48% who strongly opposed it.

The same survey found 53% opposed a local nuclear plant, a lower share than for data centers. Gallup’s open-ended April panel work found opponents most often cited environmental concerns; 18% mentioned water use, another 18% mentioned energy use and 16% mentioned pollution, including noise, air and water pollution.

Water systems give those objections a technical vocabulary. The U.S. Energy Department’s Federal Energy Management Program says water usage effectiveness for data centers is measured as annual site water use in liters divided by information technology equipment energy use in kilowatt-hours. A tower-cooled facility lowers heat by evaporating water into the atmosphere, and the department says cooling tower demand is especially high for constant loads such as data center cooling.

Heat now has its own research thread. A March arXiv preprint titled the data heat island effect study estimated land surface temperature rose by 2°C on average after an AI data center began operating, with more than 340 million people potentially exposed to that increase. The paper had been revised by April 21 and should be read as preprint evidence until peer review lands.

Compute Concentration Leaves a Long Tail

The UNU-INWEH release says only 32 countries host AI-specialized data centers, with more than 90% of that capacity in the United States and China. More than 150 countries, the report says, have little or no access to sovereign AI compute.

Those figures change the beneficiary map. A government can host mines, e-waste processing or water-intensive infrastructure without owning much of the compute capacity that creates the returns. The report estimates AI infrastructure could generate up to 2.5 million tonnes of e-waste each year by the end of the decade, much of it processed in lower-income economies with limited safeguards.

The local choices are already concrete. Oton Technology’s earlier report on the proposed AI data center on Alaska’s North Slope described a $500 million plan covering more than a square mile of tundra and targeting up to 3 gigawatts of power from stranded natural gas. The water question varies by site. The permitting question travels with every project: which public resource carries the load, and who gets the benefit?

A Permit Ledger for Water, Power and Land

The governance section of the report is practical. It asks governments to fold AI infrastructure into energy and water planning before land-use permits are granted, and to require standardized environmental reporting. It asks industry to treat model choice and default output length as part of the footprint.

The Green Grid, a data center sustainability group under the Information Technology Industry Council, moved in the same direction last year with its Water Usage Impact metric. The metric scores consumed water against local water stress, which is the part missing when a company reports a global average for water efficiency.

On the industry side, the shift is already visible in water-focused AI coalitions. Oton Technology previously covered the Water-AI Nexus Council’s attempt to curb AI’s water drain, a project that brought Autodesk into a group looking at AI’s own footprint and its use in water systems.

A permit file for a new AI campus now needs a site-level ledger covering grid source, cooling design, annual water consumption, local water stress, land footprint, backup generation and e-waste plan. The June 3 report leaves regulators with one measurable demand: publish that ledger before the first rack is installed.