The Hidden Water Footprint of the AI Revolution

Blurs.co.uk
10 Min Read
Photograph: Rsparks3 (2025), Wikimedia Commons, CC0 1.0 public-domain dedication.
Summarize with AI

The cloud has no visible pipes, reservoirs or cooling towers. Yet every search, video stream, online transaction and AI request depends on physical buildings packed with computers. Those machines generate heat, and removing that heat can require a surprising amount of water.

As data centres expand to support artificial intelligence, the water question is becoming harder to ignore. The impact is not the same everywhere: a facility in a cool, water-rich region may have a very different footprint from one operating in a hot, dry area. The type of cooling system, the electricity supply and the efficiency of the computing workload all matter.

Why servers need so much cooling

Almost all the electricity consumed by a server eventually becomes heat. A large facility may contain thousands of servers operating continuously, alongside networking equipment, storage systems and power electronics. If this heat is not removed, equipment can overheat, become unreliable or shut down.

Traditional data centres often use air-conditioning and chilled-water systems. Some facilities use cooling towers, where water evaporates to carry heat away. Evaporation can be energy-efficient, but the water that evaporates is consumed locally and must be replenished. Other systems use outside air when weather conditions allow, while newer high-density computing facilities increasingly use liquid cooling close to the chips.

Liquid cooling does not automatically mean water is consumed. Some systems circulate coolant in sealed loops, transferring heat to another system that may or may not use evaporative cooling. The details of the entire cooling chain determine the water impact.

Two different water footprints

Data-centre water use is often divided into direct and indirect consumption.

Direct water use is the water consumed at the facility, principally through cooling. It may also include water used for humidification and other building services. Water withdrawal and water consumption are not identical: withdrawn water may be returned to its source, while consumed water is not immediately available for reuse in the same local catchment.

Indirect water use occurs in the production of electricity. Power stations can require water for cooling, and the amount depends on the generation technology and local conditions. A data centre using little water on site may still be associated with substantial water consumption through the electricity it buys.

This distinction is important when comparing facilities. A headline figure for water used inside the building is not a complete account of the water footprint behind its computing.

Why location changes the equation

Cooling requirements are influenced by temperature, humidity and seasonal conditions. A facility in a cool climate may be able to use outside air or other forms of free cooling for much of the year. In hotter climates, mechanical cooling may be required more frequently.

But a cooler climate is not a universal environmental solution. A site may still be located in a water-stressed region, depend on electricity generated using water-intensive technologies, or require new power and transmission infrastructure. The local water basin matters more than a simple national average.

A 2024 study of Sweden’s data-centre service sector assessed energy use, water consumption and greenhouse-gas emissions, and developed a tool to help evaluate where facilities should be located. Its approach underlines that siting decisions need to consider several environmental factors together rather than treating cooling efficiency as the only measure.

The workload matters too

It is tempting to ask how much water one AI question uses. There is no reliable universal answer. A request may be processed on different hardware, in different facilities, using different cooling systems and electricity mixes. The model, length of the task, server utilisation and the way the operator allocates shared infrastructure all affect any estimate.

A 2025 review in Resources, Conservation and Recycling found that water use per data-centre workload could vary by more than 10,000-fold across the conditions examined. The authors identified server efficiency, electricity-grid water factors, server utilisation, cooling type, climate and infrastructure efficiency among the important determinants.

That enormous range is a warning against turning one estimate into a universal claim about every prompt or every data centre. It also shows why transparent reporting needs to explain the assumptions behind its figures.

The scale of the expansion

A study published in the Journal of Cleaner Production in November 2026 estimated that China’s data-centre water consumption rose from 229 million cubic metres in 2016 to 478 million cubic metres in 2024. The study included both direct cooling water and water associated with electricity generation, and used a spatial inventory of 823 facilities.

The researchers found that expansion in computing capacity had historically been the main driver of rising water use, while improvements in efficiency partly offset the increase. Their scenarios suggest that efficiency upgrades can reduce future demand, but efficiency alone may not fully counterbalance unconstrained growth.

These figures relate to China and to the study’s defined operational water footprint. They should not be treated as a direct estimate for the global industry.

Can the industry use less?

There are several ways to reduce water pressure. Operators can use air-side cooling when weather permits, improve the efficiency of chillers and cooling towers, use closed-loop liquid cooling, and select non-potable or recycled water where appropriate and safe. Better server utilisation can also reduce the amount of infrastructure needed for a given computing task.

However, choices involve trade-offs. A system that saves water may use more electricity; a system that saves electricity may consume more water through evaporation. The right answer depends on local water scarcity, the carbon intensity of electricity, climate and the facility’s design.

Reporting should therefore include both direct and indirect water use, explain the boundaries of the calculation, and disclose the location and water source. A low water-use figure is less meaningful if it hides water consumption in electricity generation or ignores local scarcity.

The cloud has a local footprint

Digital services are often marketed as immaterial. In reality, their infrastructure competes for land, electricity, water and cooling capacity. The environmental consequences are experienced in particular places, even when the people using the services live thousands of miles away.

The central question is not whether data centres should use water at all. Cooling is a physical requirement, and different systems have different impacts. The question is whether new capacity is being built where resources can support it, whether operators are measuring the full footprint, and whether the benefits of digital growth justify the demands placed on local communities.

The cloud may be everywhere. Its water footprint is somewhere specific.

A global warning about water-stressed locations

A global study published in Environmental Science & Technology in September 2026 adds an important geographical dimension. The researchers assessed 9,347 identified data centres worldwide and estimated that around one quarter would have experienced at least one month of water-gap conditions each year under the study’s 2001–2010 hydrological baseline. A “water gap” means local water demand exceeds available supply; it is not a measurement of how much water each data centre actually consumed. The result indicates exposure to water stress, rather than proving that every facility in the group is currently causing a shortage.

The distinction matters because water risk is intensely local. A facility’s impact depends on its cooling system, water source, electricity supply and the condition of the catchment around it. The study strengthens the case for publishing site-level water data and assessing water availability before approving new capacity. It does not provide a universal figure for the water used by an individual AI query.

Sources and further reading

Share This Article
Leave a Comment

Leave a Reply