The Water Behind Artificial Intelligence

The Water Question Sounds Simple

How much water does an AI data centre use?

It is the question many citizens ask first because water is not an abstract resource. It comes from a municipal system, groundwater, a river, a reservoir or another source shared with homes, farms, businesses and ecosystems.

Water was also one of the strongest concerns Erin Brockovich raised during her appearance with Theo Von. The discussion helped bring a technical infrastructure issue into public view. It also demonstrated why citizens want direct answers before large facilities receive long-term access to local resources.

Watch or listen to Erin Brockovich with Theo Von

The problem is that the direct answer is not one universal number.

A figure may describe one proposed facility’s maximum permitted withdrawal. Another may represent one company’s annual consumption. A third may combine onsite cooling water with water used by distant power plants. A fourth may estimate the water associated with a single text prompt under one model, facility and measurement method.

The question is not whether a widely shared number once came from a credible source. The question is whether that number describes the facility, model, climate, electricity supply and time period being discussed now.

Why Data Centres Need Cooling

Computing equipment turns electricity into useful processing work, but nearly all of that electricity eventually becomes heat. The more densely a facility packs advanced processors, the more heat it must remove.

Cooling is not optional. Excess heat can reduce performance, shorten equipment life, damage hardware and interrupt services.

A traditional system may use fans to move heat from server racks into chilled-water equipment. The heat then moves through another loop to a cooling tower, where evaporation releases it into the atmosphere. The United States Department of Energy identifies evaporation and cooling-tower blowdown as major sources of water demand in these systems.

Other facilities use outside air, air-cooled chillers, direct-to-chip liquid cooling, immersion systems, closed loops or hybrid designs that change according to outside temperature.

Review the Department of Energy’s data-centre cooling guidance.

The Most Important Water Distinction

A facility can withdraw a large volume while returning much of it. Another facility can withdraw less but evaporate a greater share. Neither figure by itself explains the effect on the watershed.

The period also matters. A yearly average can hide a high summer peak. A maximum permitted amount can be much higher than normal use. A campus figure can combine several buildings. A company-wide total can combine facilities operating in different climates.

The European Union now requires covered operators to report total water input and potable-water input using defined measurement methods. That is progress because it creates common labels. Public reporting, however, can still be aggregated in ways that make a specific facility difficult for local citizens to evaluate.

Some of the Water May Be Used Far From the Data Centre

Direct water use occurs inside the facility boundary. It can include cooling, humidity control, landscaping, sanitation and other site operations.

Indirect water use can occur through electricity generation. Thermal power plants may withdraw or consume water for cooling. Hydroelectric reservoirs can lose water through evaporation. Fuel production can also carry a water footprint.

That means a facility using little water onsite may still depend on a water-intensive electrical system. A water-cooled facility connected to a less water-intensive grid can show a different balance.

A 2025 Lawrence Berkeley National Laboratory study found that the water consumed per computing workload can vary by more than 10,000 times. The leading factors included server efficiency, the water intensity of the electrical grid, server utilization, cooling system, facility efficiency, climate, inactive equipment and hardware replacement cycles.

The research did not find one universal cooling or siting formula that minimizes water use everywhere.

Read the Lawrence Berkeley National Laboratory study.

The Problem With “One Prompt Uses This Much Water”

Per-prompt estimates are attractive because they turn a large infrastructure issue into something one person can imagine. The problem is that the result changes according to the model, task, hardware, cooling design, location, weather, server utilization and electricity source.

A short text response is not the same workload as generating video, producing a long reasoning trace or operating an agent that completes several steps. A heavily utilized modern server is not the same as older equipment sitting partly idle.

In a 2025 preprint, Google researchers reported that the median text prompt in the measured Gemini Apps production system consumed approximately 0.26 millilitres of water under their methodology—the equivalent of several drops.

That finding applies to the Google system, workload and measurement period studied. It is not a universal number for every AI model, every company, every type of prompt or every data-centre location.

It also does not prove that aggregate AI water use is insignificant. Per-task efficiency can improve while total water demand rises because more people use the services, workloads become more intensive and more facilities are built.

Using Less Water Can Require More Electricity

Evaporative cooling can remove heat efficiently and may use less electricity than some dry-cooling systems. Its main drawback is the water lost through evaporation, particularly during hot weather.

Air cooling and air-cooled chillers can reduce direct water demand. In some climates, they may require more electricity to move the same amount of heat. If the electrical grid itself is water-intensive or carbon-intensive, reducing onsite water may increase another environmental cost.

Direct-to-chip liquid cooling can move heat away from dense AI processors more efficiently than fans alone. In a closed loop, the coolant can circulate continuously rather than being evaporated.

Microsoft says its newer AI-optimized data-centre design uses a closed loop and consumes zero water for cooling during operation. The company estimates that the design can avoid more than 125 million litres of water per facility each year compared with its earlier global average. Microsoft also acknowledges a nominal increase in energy use compared with evaporative designs.

That is a company estimate tied to a specific design. It should not be generalized to every data centre. It does demonstrate that cooling choices can substantially change the trade-off.

Review Microsoft’s zero-water cooling design.

Chile: A Data Centre Forced Back to the Water Question

Chile offers one of the clearest examples of cooling design becoming a public and legal issue.

A proposed Google data centre in Cerrillos, near Santiago, was challenged over the effect of groundwater use on the Central Santiago Aquifer and whether the environmental review had adequately considered climate change.

In February 2024, Chile’s Second Environmental Court partially annulled the project’s approval and ordered the environmental review process to return to an earlier stage so that climate-change effects on the water component could be considered.

The court record also notes that the project had proposed replacing groundwater-dependent cooling towers with air-condensing cooling equipment, eliminating groundwater consumption from three wells for which it held water rights.

The case does not prove that every data centre in Chile threatens an aquifer. It proves that cooling design, water rights, climate conditions and public review can change the shape of a project.

Chile then moved beyond one dispute. In 2026, its Environmental Assessment Service issued national criteria describing the principal components, equipment and environmental impacts that data-centre proposals must address when entering the environmental assessment system.

Read the Chilean Environmental Court summary and Chile’s 2026 data-centre assessment criteria.

A Redesigned Future Does Not Erase Current Use

New cooling systems may reduce future demand, but existing facilities still matter.

Google reported that its Quilicura data centre in Chile consumed 121.8 million gallons of water in 2024. The company says the water came from the local utility, Aguas Andinas, and compared the annual amount with the consumption of a golf course.

The figure is useful because it is tied to one facility and one year. The golf-course comparison is less useful without additional context.

Which golf course? In what climate? Does the number describe withdrawal or consumption? Was demand concentrated during the hottest months? How stressed was the watershed? Which other users depend on the same system?

Corporate comparisons can be numerically accurate while still being weak measures of local water impact.

Review Google’s Quilicura water disclosure.

Can Data Centres Give Water Back?

Data-centre operators increasingly describe reclaimed water and water replenishment as part of their environmental strategy.

Reclaimed water

A facility can use treated wastewater instead of potable water. That can reduce demand on drinking-water supplies and provide a customer for municipal reuse systems.

Reclaimed water still requires treatment and pipes. It may still be consumed through evaporation. It can also compete with agricultural, industrial or environmental reuse options.

Water replenishment

Companies may fund wetland restoration, irrigation efficiency, groundwater recharge, leak reduction or drinking-water projects. These investments can produce real benefits.

Replenishment is not automatically the same as reducing direct local consumption. A project may operate in another watershed, produce benefits during a different season or rely on modelled savings rather than measured water returned to the original source.

When Water Access Becomes Corporate Power

Water is already priced, licensed, allocated and contracted in different ways around the world. Data-centre demand adds another large industrial user to systems that may already serve households, farms, manufacturing and ecosystems.

Citizens should ask whether a company holds long-term extraction rights, whether rates are confidential, whether the allocation survives drought restrictions and whether the rights transfer if the facility or property is sold.

There is not enough evidence to say that global data-centre expansion represents one coordinated plan to control the world’s water.

Citizens do not need proof of a global conspiracy to examine whether one long-term private agreement reduces public control over one local resource.

The most important question is practical: when scarcity arrives, who has the legal right to continue drawing water, who must reduce consumption and who made that decision?

Data Centres Can Be Designed for Water Responsibility

The strongest case made by developers and engineers is that newer facilities can deliver much more computing work with less water than earlier generations.

Responsible approaches include choosing water-secure locations, avoiding potable water, using reclaimed water, installing closed-loop or direct-to-chip cooling, publishing facility-level measurements, accepting binding drought reductions and restoring the same watershed.

Efficiency gains should not be dismissed. They also should not be treated as proof that total demand will fall.

A more efficient facility can still increase local consumption if the campus becomes larger, adds more processors or operates more intensive workloads. Several efficient facilities can still create a cumulative burden on the same water system.

The question is not whether efficient data centres exist. It is whether efficiency, location and public protection are required before approval—or merely promised after opposition begins.

What We Know—and What Must Be Proven Facility by Facility

Creators Should Understand the Resources Behind Their Tools

AI music depends on remote infrastructure. Creators cannot calculate a song’s exact water footprint from the information most companies currently provide.

Guilt-based claims built around one prompt or one song are not useful. They turn a policy and infrastructure question into a personal accusation without giving the individual enough information to make a meaningful decision.

Creators can still support facility-level reporting, responsible siting, better cooling systems and enforceable public protections.

Your position is not proved by repeating the most alarming number. It is proved by being willing to replace that number when better evidence becomes available.

Questions to Ask Before a Data Centre Receives Water Approval

1. Which water source will the facility use?
2. Is the source potable, reclaimed, groundwater or surface water?
3. What is the permitted maximum withdrawal?
4. What is the expected normal withdrawal?
5. What is the expected actual consumption?
6. What are the monthly and seasonal peaks?
7. How much water is returned, and where?
8. Which cooling technology will be used?
9. Why was that system chosen for this climate?
10. How does demand change during extreme heat?
11. What restrictions apply during drought?
12. Does the facility receive priority over homes or agriculture?
13. What rate will the operator pay?
14. Is the rate confidential or discounted?
15. Can the water rights be transferred?
16. Is future expansion included in the agreement?
17. Is indirect electricity-related water use estimated?
18. Will monthly facility-level data be public?
19. Who independently verifies the figures?
20. Is replenishment in the same watershed?
21. What penalties apply if commitments are missed?
22. What is the cumulative demand from all nearby data centres?

Primary Sources Used

• Lawrence Berkeley National Laboratory: The Water Use of Data Center Workloads
• United States Department of Energy: Cooling Water Efficiency for Data Centers
• European Union: Data-Centre Sustainability Reporting Regulation
• Google Researchers: Measuring the Environmental Impact of Delivering AI at Google Scale
• Microsoft: Next-Generation Zero-Water Cooling Design
• Chile Second Environmental Court: Cerrillos Data Center Decision
• Chile Environmental Assessment Service: Data-Centre Assessment Criteria
• Google: Quilicura Data Centre Water Disclosure

The Water Question Ends at the Data Centre—but Begins in the Watershed

Artificial intelligence does not drink water. Data centres, power systems and cooling equipment do.

That distinction matters because it tells citizens where to look.

Not at a universal claim about one prompt. Not at a global company total detached from local conditions. Not at a maximum permit presented as normal use.

The evidence is found in the cooling design, water source, monthly measurements, drought agreement and watershed expected to support the facility for decades.

The intelligence may be delivered around the world. The water is taken from somewhere specific.

That place is the community surrounding the data centre.