How Much Water Do AI Data Centers Use? New Cooling Technology Could Reduce Their Impact

As artificial intelligence expands across the United States, communities are raising concerns about the water and energy required to operate AI data centers. Protesters in states including Texas, New Mexico and Arizona have called for stronger protections for local water supplies, with signs reading “Protect our water” and “Water for people, not AI.”

Data centers require substantial electricity to run AI systems, and many also use water to remove the heat produced by high-performance computer processors. However, the environmental impact of AI data centers is complex. Water use varies widely depending on the facility’s cooling system, location, climate, energy source and workload.

Some online claims suggest that a single chatbot query uses as much as 500 milliliters of water. Other commentators argue that AI has little or no meaningful water footprint. Neither statement tells the complete story. Estimates differ by data center, model, season and method of accounting, while technology companies do not always publish consistent or complete water-use data.

Why AI data centers need water

AI processors generate intense heat while performing calculations. Some processors can reach temperatures of approximately 176 degrees Fahrenheit, similar to the heat generated by an overloaded laptop but on a much larger scale.

Many data centers use evaporative cooling, in which water absorbs heat and evaporates. The process works much like sweating: As water changes from liquid to vapor, it carries heat away from the surface. This approach can be effective, but it consumes water and may place additional pressure on supplies in drought-prone regions.

Researchers estimate that data center cooling systems in the United States consumed approximately 66 billion liters of water in 2023, or less than 1% of total national water consumption. On a national scale, that figure is relatively small compared with agriculture and some manufacturing industries.

Local impacts can be much more significant, however. A data center located in a water-stressed community may compete with households, farms and other businesses for limited supplies. As AI facilities grow larger and more numerous, their water demand could become a serious concern in parts of the Southwest and other regions vulnerable to drought.

AI water consumption could rise as data centers expand

Although current data center water use represents a small share of national consumption, rapid AI growth could change that balance. One analysis projected that U.S. data centers could consume between 731 billion and 1.125 trillion liters of water annually by 2030, depending on the pace of expansion and the technology used.

These estimates include more than the water used directly for cooling. Power plants that generate electricity for data centers may also consume water, particularly facilities that burn coal or natural gas. Those plants use water to cool steam after it passes through turbines that generate electricity.

As a result, switching data centers to renewable energy could reduce their overall water footprint. Solar and wind facilities generally require far less operational water than fossil-fuel power plants.

Better locations and renewable energy can lower water use

Where companies build data centers can be as important as how they cool them. Constructing facilities in drought-stricken areas can increase pressure on local water supplies, even when the data center represents a small percentage of statewide consumption.

Researchers have suggested locating new facilities in regions with lower water stress and abundant renewable energy. Potential locations include parts of Montana, Nebraska, Texas and South Dakota. Strategic siting, combined with more efficient processors and low-water electricity generation, could reduce future AI-related water use by as much as 86% in some scenarios.

Energy systems experts also emphasize the importance of restoring and replenishing natural water resources. With improved technology, responsible siting and water-reuse programs, some facilities may eventually operate with little or no water used for on-site cooling.

How liquid cooling reduces AI data center water consumption

Traditional data centers often use fans to move hot air away from computer racks. Cooling towers then remove heat from the facility, with some water lost through evaporation. This system can consume significant amounts of water, particularly during hot weather.

AI data centers produce more heat than many older facilities because they use powerful processors for machine learning and other computationally intensive tasks. Air cooling alone may not always be sufficient, which has led companies to adopt more efficient liquid-cooling systems.

In a liquid-cooling system, pipes carrying water or another heat-transfer fluid run close to the processors. This removes heat directly from the hardware rather than cooling the entire building. Many systems operate as closed loops, allowing the fluid to circulate repeatedly without being lost to evaporation.

After absorbing heat, the fluid can be cooled with outdoor air when weather conditions allow. In hotter or more humid climates, facilities may need mechanical refrigeration or limited water-based cooling to maintain safe operating temperatures.

Companies including Amazon, Microsoft and Google are developing or deploying liquid-cooling systems for high-performance computing. Microsoft has also described data center designs intended to use zero water for cooling in certain applications.

New cooling technologies are being developed

Several emerging technologies could further reduce the water requirements of AI infrastructure. One approach uses carefully engineered pipe systems to remove heat more efficiently from individual processors. Another, known as immersion cooling, places computer hardware in a tank filled with a specialized nonconductive coolant.

Immersion cooling can transfer heat efficiently without relying on conventional evaporative cooling towers. However, it may make maintenance more difficult because technicians must remove hardware from the cooling fluid when components need to be replaced or upgraded.

Researchers and companies are also exploring underwater data centers. Placing servers in the ocean could provide access to a naturally cool environment, but maintenance and hardware replacement would be challenging.

Processor design is another part of the solution. Some newer AI processors can operate at temperatures of approximately 113 degrees Fahrenheit, potentially allowing large fans and ambient air to remove heat under many weather conditions. However, running processors at higher temperatures may affect performance, and some facilities also house older servers that require cooler conditions.

During extreme heat or high humidity, even advanced systems may need additional cooling. In Texas, for example, some AI data centers reportedly use water-based cooling only during the hottest part of the summer.

How much water do Texas data centers use?

Data centers currently account for a small share of Texas’ total water demand, according to researchers at the University of Texas at Austin. However, continued construction could increase that proportion substantially by 2040, with estimates ranging from approximately 3% to 9% under some scenarios.

Improving water efficiency could reduce that impact. As cooling systems become more advanced, facilities may be able to support larger computing workloads without using proportionally more water.

The exact amount of water used by an AI data center depends on several factors, including the processor type, cooling design, weather, electricity source, facility size and operating schedule. For that reason, a single water-use figure cannot accurately describe every AI query or data center.

Companies are investing in water-reuse programs

Cooling technology is only one part of the sustainability challenge. Technology companies are also investing in water-replenishment projects, wastewater treatment and systems designed to return more water to local communities than their facilities consume.

Amazon Web Services has said it aims to become water positive by 2030. Google has announced projects intended to replenish more than 120% of the freshwater it consumes at data centers by the same year. The effectiveness of these programs will depend on where the projects are located and whether they benefit the same watersheds affected by data center operations.

Some companies are also reusing waste heat. In parts of Europe, heat captured from data center cooling systems is transferred to district heating networks and used to warm homes and other buildings. Reusing this heat can reduce reliance on fossil fuels while improving the overall efficiency of data center operations.

The future of AI data center sustainability

The environmental impact of AI data centers will depend on decisions made by technology companies, engineers and policymakers over the next several years. Key factors include where facilities are built, how they are cooled, what power sources supply them and whether companies invest in local water restoration.

In water-rich regions with renewable electricity, an AI data center may have a relatively limited effect on local resources. In drought-prone areas powered by water-intensive fossil-fuel plants, the combined impact can be much greater.

Experts say the most effective strategy is to combine multiple solutions: build in suitable locations, use renewable energy, improve processor efficiency, adopt closed-loop or air-based cooling, reuse wastewater and make water-use data more transparent.

AI does not have a single, universal water footprint. Its environmental cost depends on infrastructure choices. With careful planning and continued innovation, the industry could reduce water consumption even as demand for artificial intelligence continues to grow.

This article was first published by Knowable Magazine, a nonprofit publication dedicated to making scientific knowledge accessible. Sign up for the Knowable Magazine newsletter.