Training GPT-3 took an estimated 5.4 million liters of water, and AI's water use is still climbing

Study title: Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models

Training OpenAI's GPT-3 in Microsoft's U.S. data centers consumed an estimated 5.4 million liters of water, according to a peer-reviewed analysis by researchers led by Shaolei Ren of UC Riverside, published in Communications of the ACM. About 700,000 liters of that evaporated on site in data-center cooling; the rest was consumed by the power plants that supplied the electricity. None of it appears in AI model cards, which routinely report carbon emissions but not water.

Read the study abstract

The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonetheless, the equally important water (withdrawal and consumption) footprint of AI has largely remained under the radar. For example, training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand is projected to account for 4.2-6.6 billion cubic meters…

By the Truvace deskNovember 4, 20264 min read1 sourceEvidence: Emerging · 30/100 — Commentary, opinion, or narrative review · 1 independent sourcePeer-reviewedTRV-2026-2359· cite & verify

Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models — "Data center cooling tower" by Rsparks3 is marked with CC0 1.0. To view the terms, visit https://creativecommons.org/publicdomain/zero/1.0/deed.en/.

In brief

Running the model costs water too. The authors estimate GPT-3 consumes roughly one 500 ml bottle of water for every 10 to 50 medium-length responses (about 800 words in, 150 to 300 words out), depending on where it runs: about 10 responses per bottle in Microsoft's Washington state data centers, about 70 in Ireland. They call this a conservative figure, because newer models and less efficient third-party data centers can use several times more energy and water per request.

At global scale, the paper projects that AI's electricity demand of 85 to 134 TWh in 2027 implies 4.2 to 6.6 billion cubic meters of water withdrawn, more than Denmark withdraws in a year four to six times over, or about half of the United Kingdom's total. Of that, 0.38 to 0.60 billion cubic meters would be evaporated and lost to the local watershed. For context, Google's own data centers withdrew 29 billion liters and evaporated more than 23 billion liters in 2023, nearly 80% of it drinkable water, with consumption up 20% in 2022 and another 17% in 2023.

These are modeled estimates, not metered measurements. They combine published energy figures with Microsoft's annual average water efficiency for each site and national averages for power-plant water use, and GPT-3's actual training location has never been disclosed. Water used to manufacture chips and servers is left out for lack of public data, so the totals undercount AI's full footprint.

Main points

Training GPT-3 (175 billion parameters, about 1,287 MWh of electricity) is estimated to consume 5.4 million liters of water in an average U.S. Microsoft data center, including about 700,000 liters evaporated on site for cooling.
Location changes the bill by more than sixfold: the same training run is estimated at about 2.3 million liters in Ireland, 2.4 million in Texas and 15.3 million in Washington state.
Using the model consumes roughly one 500 ml bottle of water per 10 to 50 medium-length responses, an estimate the authors call conservative.
Global AI demand in 2027 is projected to withdraw 4.2 to 6.6 billion cubic meters of water, about half the United Kingdom's annual withdrawal, and to evaporate 0.38 to 0.60 billion cubic meters.
Data-center cooling evaporates about 1 to 9 liters of water per kWh of server energy (highest in hot weather, such as Arizona summers), on top of about 3.1 liters per kWh consumed by the average U.S. power plant.
Cutting carbon and cutting water can pull in opposite directions: sunny midday hours are cleaner for carbon but hotter, which raises cooling water use; in Virginia, hourly carbon and water efficiency were almost uncorrelated (0.06).
The authors recommend reporting water use in AI model cards and cloud dashboards, counting power-plant water as standard, and scheduling training and inference at water-efficient times and places.

The problem

Training and running large AI models consumes millions of liters of mostly drinkable freshwater that companies do not disclose: an estimated 5.4 million liters to train GPT-3 alone, with global AI projected to withdraw 4.2 to 6.6 billion cubic meters of water in 2027.

The rundown

AI uses water in three places. Data centers cool their servers with cooling towers or water-assisted air cooling, and both evaporate water. Cooling-tower water can only be recycled about 3 to 10 times before minerals build up, so operators keep topping it up with clean freshwater, often drinking-quality, and roughly 80% of what they draw evaporates. Power plants evaporate more water generating the electricity. A third, largely hidden share goes into manufacturing chips, which needs ultrapure water: Apple reports that its supply chain accounts for 99% of its total water footprint.

The estimate works like a carbon footprint. Energy used is multiplied by two factors: the data center's on-site water efficiency, and the water intensity of the local grid adjusted for cooling and power overhead. Manufacturing water is spread over a server's lifetime. Because OpenAI never said where GPT-3 was trained, the authors worked out the figure for 18 Microsoft data-center locations using Microsoft's published efficiency numbers, from about 2.3 million liters in Ireland to 15.3 million in Washington state.

The authors argue the real numbers are likely higher. They assume 0.004 kWh per response, but measurements on comparable modern systems run about 0.010 kWh per request for the smaller Llama-3-70B and about 0.016 kWh for Falcon-180B. Microsoft's data centers are among the industry's most water-efficient, so the same model in a typical third-party facility would use several times more water on site. Their U.S. power-plant figure (3.14 liters per kWh) is also below the 4.35 reported by the 2024 U.S. data center energy report, which projects that on-site water use by U.S. data centers could double or quadruple from 2023 to 150 to 280 billion liters a year by 2028.

The paper's fixes are about disclosure and timing. It asks AI developers to add water to model cards and cloud dashboards, to report power-plant water use as standard (Meta has started), and to fund research on chip-manufacturing water. Because water efficiency changes with weather and the grid's fuel mix, training could be shifted to cooler hours or more water-efficient sites. Doing that well means balancing "follow the sun" for low-carbon solar power against "unfollow the sun" to avoid the hottest, thirstiest hours.