How Much Water Do Data Centres Actually Use?

Data centres hold and process the digital information used for online services and artificial intelligence systems. They consume large quantities of water to keep equipment cool.

The University of California, Riverside has estimated that a single 100-word prompt to an AI system can use the equivalent of a small bottle of water. A session of 10 to 50 prompts can use about half a litre. Training large AI models requires hundreds of thousands of litres of water.

The United States has 5,426 data centres and together they use billions of gallons of water each year. A large site can use up to 5 million gallons per day, almost equivalent to the daily use of a town of 50,000 people. A medium-sized site can use about 110 million gallons per year. In 2021, US data centres used an estimated 449 million gallons per day.

 

Why Do Data Centres Use Water?

 

Servers produce heat while operating. High temperatures can damage the equipment. Cooling systems are used to control this heat.

Many facilities use evaporative cooling towers where clean water is evaporated to remove heat. Others use liquid cooling systems that bring water or synthetic liquid into direct contact with components. Immersion cooling places servers into a non-conductive liquid. These systems can lower water use but cost more to install.

Water is also used indirectly through electricity generation. Power plants that supply energy to data centres consume large amounts of water. In 2023, indirect water use from electricity for US data centres reached around 211 billion gallons and coal-fired plants can use more than 19,000 gallons of water per megawatt-hour. Gas plants use about 2,800 gallons.

 

 

Where Is This Happening Most?

 

About 20% of data centres are in areas with water shortages. In Phoenix, Arizona, daily cooling demand can reach 170 million gallons. Northern Virginia, which has over 300 data centres, used nearly 2 billion gallons in 2023, a 63% increase since 2019. Loudoun County, which has 200 facilities, used about 900 million gallons last year.

Water withdrawals can lower aquifer levels and reduce supply for homes and agriculture. Wastewater from cooling can also raise the temperature of rivers and harm aquatic life. Higher temperatures and drought can increase cooling demand while limiting water supply.

In 2023, residents in Uruguay protested against a proposed large facility during a drought. Similar opposition has been reported in Holland and Chile.

 

How Does AI Affect Water Use?

 

Artificial intelligence requires intensive computing which increases both electricity use and cooling needs. Goldman Sachs projects that power capacity for data centres will increase by 165% by 2030. Water use for AI worldwide could reach between 1.1 trillion and 1.7 trillion gallons per year by 2027. This is close to the household water use of the state of California.

The training of ChatGPT used about 185,000 gallons of water. This accounted for 6% of a local utility’s supply during peak months. Larger AI models will require more resources to train.

 

What Can Reduce Water Use

 

Closed-loop cooling systems can reuse water and cut freshwater use by up to 70%. Some facilities use treated wastewater or low-quality water from agriculture instead of drinking water.

In cooler climates, free cooling systems can use outside air to lower temperatures. Direct-to-chip cooling and immersion cooling can also lower evaporation losses. Switching to renewable power sources such as wind or solar can lower indirect water use compared to coal or gas plants.

Fewer than 1/3 of operators track water use. Without measurement, it is difficult to determine whether use is going down.