AI did not just demand more electricity — it demanded water. By 2025, U.S. data centres, increasingly expanded for AI workloads, were consuming nearly one trillion litres of water a year, much of it used in cooling systems that shed heat by evaporating water into the air.

The public argument about artificial intelligence has mostly been an argument about electricity. Bigger models, denser chips and new data centres have put fresh pressure on power grids that were not built with this pace of computing demand in mind.

But the same buildings also demand something quieter, more local and often more politically sensitive: water.

By 2025, estimates cited in investor and media reporting put North American data-centre water use at around one trillion litres a year, with the United States accounting for the centre of that build-out. The figure is not a single official meter reading. It sits inside a wider problem: data-centre water use is still measured unevenly, disclosed inconsistently and often hidden behind company-level reporting that tells the public little about individual sites.

Even so, the direction is clear. A 2021 paper in npj Clean Water by David Mytton estimated that U.S. data centres were already consuming about 1.7 billion litres of water a day, while noting that less than a third of operators were then measuring water consumption. That older estimate alone works out to more than 600 billion litres a year.

The AI boom did not create the data-centre water issue from nothing. It intensified an existing one.

Why computers need water

A data centre is, physically, a heat problem. Servers convert electricity into computation, and much of that electricity eventually becomes heat. The more densely servers are packed, and the harder they are driven, the more heat must be removed to keep equipment within operating limits.

Many facilities do this with systems that use evaporation. Water absorbs heat, then some of it is evaporated into the air through cooling towers or related systems. The method can be efficient from an electricity point of view, especially compared with some purely mechanical cooling systems. But the trade-off is that water leaves the immediate system. It does not simply circulate forever.

That is why data-centre water accounting can become slippery. A facility may withdraw water from a municipal system, a river, an aquifer or a reclaimed-water supply. Some of that water may be discharged later. Some may be consumed through evaporation. In addition, electricity generation can have its own water footprint, especially where power comes from thermal plants that use water for steam cycles and cooling.

So the water demand of AI infrastructure is not only the water visible at the data-centre fence line. It can also include the water used to generate the electricity that keeps the servers running.

AI changed the scale of the load

The International Energy Agency’s 2025 Energy and AI report said data centres accounted for about 1.5 percent of global electricity consumption in 2024, or 415 terawatt-hours. The United States represented the largest share of that demand, and the IEA projected that global data-centre electricity consumption would more than double by 2030, with AI as the most important driver of growth.

Electricity demand is not the same as water demand, but the two are linked. More computation means more heat. More heat means more cooling. More cooling, depending on the system and local climate, can mean more water.

A 2026 preprint by Gianluca Guidi and colleagues, focused on U.S. hyperscale data centres, estimated that 403 such facilities operating between May 2024 and April 2025 consumed roughly 68 to 99 terawatt-hours of electricity under different load scenarios. The authors also found that more than half of attributed generation came from fossil-fuel sources under their central scenario.

That matters for water because a data centre’s footprint is shaped not only by the cooling system inside the building, but also by the grid behind it. A facility cooled with relatively little on-site water can still be tied to off-site water use if its electricity comes from water-intensive generation.

The transparency problem

One reason the water debate is difficult is that the numbers are not always public in useful form.

Large technology companies publish sustainability reports, but those reports often give global or company-wide totals rather than facility-level water figures. They may separate direct water use from broader environmental claims, or they may report efficiency metrics without showing total use at a particular site. Third-party data centres, leased capacity and electricity-related water consumption can further obscure the picture.

That is why investor pressure has begun to focus on site-level disclosure. Reporting in 2026 described investors pressing Amazon, Microsoft, Google and others for more detailed water and energy data, arguing that broad sustainability claims do not tell local communities how much water a nearby facility will require.

The question is not only whether a company is using less water per unit of computation. Efficiency can improve while total water use rises if the number and size of facilities grows quickly enough. This is the rebound problem in infrastructure form: each unit becomes more efficient, but the total system expands faster.

Why location matters

A million litres of water does not mean the same thing everywhere.

In a wet region with strong wastewater recycling and spare municipal capacity, a data centre may place a manageable burden on local systems. In a dry region, or a town already facing strained aquifers and ageing pipes, the same demand can become a public dispute.

This is where the national numbers can mislead. U.S. data-centre water consumption may be small beside agriculture or thermoelectric power in aggregate, but data centres are not spread evenly across the country. They cluster around fibre routes, cheap land, tax incentives, grid connections and existing technology corridors. Their impacts are therefore local before they are national.

A mid-sized facility can be a large new customer for a local water utility. During heatwaves, when cooling demand rises and households also use more water and electricity, that pressure can become more visible. Communities do not experience “the cloud” as an abstraction when a project arrives with requests for power, pipes, backup generators and secrecy around the end user.

Not all cooling is the same

There are ways to reduce data-centre water demand. Some operators use reclaimed wastewater instead of drinking water. Some facilities rely more heavily on outside air in cooler climates. Others are moving toward closed-loop liquid cooling systems that recirculate water rather than continuously evaporating it.

Those designs matter, especially for new AI campuses where power density is high. But they do not erase the need for careful accounting. A closed-loop system may reduce direct water consumption at the site, while still requiring electricity whose generation has water and carbon consequences. A facility using reclaimed water may be preferable to one using treated drinking water, but that reclaimed supply may also have other possible local uses.

The real question is therefore not whether AI data centres can ever be made less water-intensive. They can. The question is whether transparency, siting and infrastructure planning will keep pace with construction.

The hidden physicality of AI

AI is often sold as weightless intelligence: software running somewhere beyond the user’s attention. The water issue makes that image harder to sustain.

Every generated answer, image, code suggestion or recommendation is part of a physical system. It depends on chips, buildings, substations, transmission lines, cooling equipment and water infrastructure. The point is not to pretend that AI is uniquely wasteful among modern industries. It is to recognise that the technology is being built at a scale where ordinary local resources become part of the global computation stack.

This is also why the water debate should be handled carefully. It is easy to turn litres into shock figures without asking what is being counted, where the water comes from, how much is consumed rather than withdrawn, and what alternatives exist. But it is just as easy for companies to hide behind efficiency ratios that do not answer the community’s simpler question: how much water will this site use, and from whose supply?

By 2025, AI had made that question harder to avoid. The new data centres did not only ask grids for more electricity. In many places, they also asked watersheds, utilities and local residents for a share of the cooling system.