Rahul SomvanshiUsing ChatGPT can have a water footprint, depending on how water use is accounted for across the computing and electricity systems involved. OpenAI CEO Sam Altman said—while acknowledging he was recalling the figure from memory—that roughly 38,000 ChatGPT queries use as much water as producing one California almond. A California almond requires between 1.1 and 3.2 gallons of water to grow, which calculates to roughly 0.1 to 0.3 milliliters of water per query under Altman’s estimate.
One difficulty in comparing AI water-use estimates is that studies don’t always use the same definition of “water use.”
Water-footprint studies can include direct cooling-water consumption, water associated with electricity generation, and, depending on the study boundary, water used elsewhere in the technology supply chain. These aren’t interchangeable numbers.
There’s the water that evaporates directly from cooling towers on the facility’s roof. There’s water consumed at distant power plants that generate electricity for the data center. And then there’s water used when the computer chips themselves were manufactured.
Modern data centers use a range of cooling systems. Evaporative cooling remains common, while closed-loop, dry and hybrid systems are also being deployed. Closed-loop systems recirculate water instead of constantly dumping it and pulling in new supplies. But in evaporative and recirculating systems, water minerals build up over time, requiring periodic flushing and fresh water input. The Department of Energy specifically discusses cooling towers as sources of water consumption and describes direct-liquid cooling and closed-loop approaches as alternatives being explored in federal facilities.
The more significant question isn’t whether a single ChatGPT query uses much water. It’s whether the total water footprint of billions of AI infrastructure operations across the world strains local water supplies—especially in areas already experiencing drought. A 2023 study estimated that training GPT-3 could have consumed about 700,000 liters of freshwater through on-site evaporation under the study’s assumptions.
There’s a meaningful difference between what one query uses and what global AI deployment requires. Altman also argued that discussions about AI water consumption have become detached from actual measurements.
AI infrastructure uses some water, and across billions of daily operations at scale, the combined effect matters. Especially in water-stressed regions like California, where tech infrastructure is concentrated and water resources are limited. And in places like Texas and Arizona, where data centers are rapidly expanding, regional water availability is a growing consideration.
The conversation needs to move past individual query estimates toward understanding the cumulative water impact of AI infrastructure expansion in water-stressed regions. Researchers are actively studying this, though the full future scale and regional distribution of demand remains uncertain.
