AI Water Usage: Addressing Water Stiff in Data Centers

AI is hot, attracting headlines, investments and users. It also literally works hot. Data Center (AI) models that run artificial intelligence use a large amount of electricity and generate huge heat. To prevent servers from overheating, many facilities rely on cooling systems that use water.

There are two ways to use water in AI data centers. Beyond the water that cools the server, the data center indirectly contributes to the use of water through the power generation required to power the operation. Its indirect use often accounts for more than 80% of the overall water use.

Reducing the AI's water footprint means tackling two very different issues. What happens inside the walls of a data center, and what happens beyond them in the power grid.

Direct water use: localized and sometimes stressful

Just as the human body cools itself by sweating, data centers are often cooled by the evaporation of water. This is a process that dissipates heat, causes water to be lost to the atmosphere, and therefore counts as “consumption.” Water is often drawn from the same local government system that supplies homes and businesses.

While most major tech companies now disclose direct water use, not all data centers follow suit, the big picture is unclear. Recent reports estimate that businesses will consume between 45% and 60% of the withdrawn water.

A recent report by the Lawrence Berkeley National Laboratory estimates that US data centers' direct water consumption (about 40% to about 40% of global data centers) is approximately 17.5 billion gallons in 2023. Assuming a consumption rate of 50%, that means 35 billion gallons of water withdrawal, or about 0.3% of the adjacent US total public water supply. The same report predicts that direct water consumption in US data centers could double or quadruple the 2023 level by 2028.

At the national level, data center water use is relatively modest. However, in some areas where data centers are concentrated, especially those already facing shortages, the burden on local water systems can be increased. Bloomberg News Approximately two-thirds of US data centers built since 2022 are reported to be in high water stress areas.

In Newton County, Georgia, some of the proposed data centers reportedly requested water per day rather than the entire county uses daily. Currently, staff there are faced with tough choices. There is a risk of rejecting new projects, requiring alternative water-efficient cooling systems, investing in costly infrastructure upgrades or imposing water allocations on residents.

The biggest stress is timing, not full use. On hot days when residents and businesses need the most water, data center water also demands spikes. In Arizona, data centers' monthly water usage during the summer can be nearly twice the average level.

Indirect water use: thirsty electricity

Another part of the equation is electricity that runs the data center. In many places, whether you train AI models in your data center or turn on your home light bulb, electricity is produced by fossil fuel-based power plants that require their own cooling water. The US electricity sector draws out approximately 11.6 gallons of water, consumes 1.2 gallons per kilowatt-hour produced, placing it at the nation's largest water user. Water used to produce the electricity that runs a data center is considered indirect water use.

The water used by the power plant is usually free of drinks and is not drawn from the urban water system. Still, it can stress rivers, aquifers and ecosystems, especially in water surface areas.

In most US data centers, this indirect use is significantly higher than direct on-site water use. One paper was estimated to have used GPT-3 in 2023 to generate a single text output of 150-300 words, consuming a total of 16.9 milliliters of water in an average US data center. Although these numbers may have been reduced due to increased efficiency in later models, indirect water use remains dominant.

How to minimize the impact on data centers on water

Unlike electricity, data center cooling systems are a design choice. Evaporative cooling is low cost and efficient, but water is most needed and can bear local supply during the available summer heat waves. To manage that peak demand, data centers can build on-site water storage and install thermal energy storage. Upgrading water infrastructure, such as widening distribution and securing leaks, can also help local systems better handle spikes in demand.

Evaporative cooling alternatives include using air-based water and liquid immersion cooling, recycled water to reduce drinking water use and discarding reuse of cooling to reduce cooling demand. Some advanced designs recycle the cooling water in a closed loop, so no water is consumed. These “zero water” designs eliminate the need to utilize local drinking water supplies. However, many of these designs can increase electricity demand and increase indirect water use. Water-cooled data centers consume about 10% less energy than air-cooled data centers.

In an immersion cooling system, the server is immersed in a liquid that carries heat without evaporating water. Jason Alden/Bloomberg/Getty Images

In areas with water stress, the number one priority must be low to zero water cooling systems to reduce direct use. They also need to invest in reducing indirect water use and minimizing carbon footprint from increasing electricity demand to add renewable energy to the local grid. In wet areas with carbon-intensive grids, even when onsite water consumption implies ongoing use of high evaporative cooling, it should be prioritized to reduce power usage to reduce overall water consumption.

The reality of intertwined water and electrical systems forces data center operators to navigate the tough trade-offs between global climate goals and local water needs. These options are often not simple, but they may be inevitable until renewables dominate the power grid.

The views expressed in this article are those of the author and do not necessarily reflect the views of the employer or related institutions.