AI data centers consume as much as 12 times more water indirectly than directly, lab finds

Indirect water consumption 12 times direct use, lab finds

No U.S. law requires technology companies to report the total water used to power their data centers, including the water consumed at electricity-generating plants. Among the largest tech firms, only Meta includes this indirect water use in its sustainability accounting.

The gap matters because indirect water use far exceeds what data centers draw on-site. U.S. data centers’ indirect water consumption has historically been about 12 times as great as the amount they directly consume, the 2024 Lawrence Berkeley analysis found. The multiplier varies by energy source: coal and nuclear plants require large volumes of water for cooling, natural gas needs less, and solar and wind use almost none.

Google’s just-released 2025 sustainability report provides a case in point. The company said it consumed 10.9 billion gallons of water that year — a 34% increase from 2024 — almost entirely for direct data-center cooling. A paper published earlier this year by Alex de Vries-Gao, a researcher at VU Amsterdam, found that Google’s indirect water consumption is around three times its direct use. The company pays for enough renewable energy to match 100% of its consumption, which enables it to claim low-carbon power, but critics including 16 state attorneys general who signed a 2025 letter have argued that purchasing renewable-energy credits does not offset water consumed in a specific watershed.

Meta reported 19 billion gallons of indirect water use in 2024 — more than 20 times its direct consumption. The company has a plan to be “water positive” by 2030 through restoration projects, but that pledge does not address indirect water use.

Amazon said its data centers use water seven times as efficiently as the industry average and that the company is 75% of the way toward replenishing a gallon of water for every one it draws, though that goal also does not include indirect use. An Amazon spokesman said the company “recognizes the connection between energy and water” and pointed to its more than 700 wind and solar projects globally.

The water demands of AI data centers are concentrated in regions already under water stress. Recent analyses by the Guardian and Bloomberg found that about two-thirds of new U.S. data-center construction is in water-stressed areas, including Phoenix. A 2025 analysis by the nonprofit sustainability organization Ceres found that data centers’ combined direct and indirect water demands in Phoenix today amount to about 3% of the city’s annual water use. By 2031, they could exceed 20% — a quantity approaching what Phoenix residents use to maintain lawns and landscaping. The analysis also said water demand could grow rapidly in neighboring New Mexico, where some of Phoenix’s power is generated.

“Cheap land and cheap power put data centers in the high water-stress areas,” said Matthew Pine, chief executive of Xylem, a U.S. water-technology firm. Pine added that water-intensive coal-fired power plants are being kept running past their intended retirement dates to meet AI demand.

Many new data-center complexes come with dedicated “behind the meter” natural-gas power plants. Microsoft has a planned collaboration with Chevron in West Texas; Meta, Amazon, SpaceX and other builders have similar off-grid plans. While such arrangements can reduce competition for grid power, local water must still be shared.

Kenneth Gillingham, a professor of environmental economics at Yale, said one reason companies have not historically reported indirect water use is that their data centers can be far from the power stations. In areas where water is not stressed, indirect use matters less, he added.

Cooling technology is a key variable. Most existing data centers use evaporative cooling systems that are energy-efficient but water-hungry, according to the Lawrence Berkeley report. Retrofitting those systems would be prohibitively expensive, experts said.

Nvidia recently showed off a closed-loop cooling system that does not require additional water once filled. Gillingham said the design appeared to be a win on two levels: it zeroes out direct water use while reducing energy needed for cooling. Josh Parker, head of sustainability at Nvidia, said the water footprint for AI is “potentially driven by electricity generation” and noted that other U.S. industries are not similarly scrutinized. He also suggested AI could help other sectors reduce water consumption.

Microsoft committed in 2024 to equipping all new data centers with closed-loop cooling starting in 2027. The company has also said it aims to “replenish more water than we use” as part of a “community-first AI infrastructure” pledge.

The lack of transparency around data-center water use has fueled suspicion. Alex de Vries-Gao said that in many cases, “the only information you’re getting is an extremely tiny part. You’re just getting the tip of the iceberg.”

Since 2024, an estimated $170 billion in AI data-center capacity has been blocked, stalled or canceled, according to climate consulting firm Carbon Direct, which said much of that stems from transparency and nondisclosure issues that obscure water and power requirements.