The widespread belief that artificial intelligence consumes massive amounts of water per query is based on outdated research and misunderstands how data centers actually work. Experts now say current AI water consumption is roughly 99 percent smaller than claims circulating online, and the actual water used inside data centers is smaller still.

Why Did the Water Bottle Claim Go Viral?

In 2024, The Washington Post published a report claiming that a 100-word email generated by ChatGPT consumes an entire bottle of water, roughly 519 milliliters. The article then scaled that number dramatically: one ChatGPT-generated email per week for a year would consume 27 liters of water; if 10 percent of the U.S. population (about 16 million people) each sent one AI-generated email weekly, it would require more than 435 million liters annually, equivalent to Rhode Island households' water consumption for 1.5 days.

The math seemed alarming, and the story spread widely. But when Andy Masley, a former physics teacher turned writer, examined the methodology, he discovered a critical flaw. He reached out to Shaolei Ren, the associate professor of electrical and computer engineering at the University of California, Riverside who conducted the original calculation, and learned something surprising.

"The majority of that bottle of water, even in my own estimates, isn't actually used in the data center itself. It's used in offsite power plants to generate the electricity," Masley explained after speaking with Ren.

Andy Masley, Writer and Former Physics Teacher

About half of the water-bottle estimate comes from evaporation off lakes dammed by hydroelectric plants that generate power for data centers. This is a crucial distinction: the water isn't consumed by the AI system itself, but rather by the energy infrastructure that powers it.

What Do Current Experts Say About AI Water Use?

More recent estimates from organizations like EcoLogits find that individual AI prompts cost between 1 to 10 milliliters of water, roughly 99 percent smaller than the Washington Post's estimate. The amount of water actually used inside the data center itself is even smaller, between 0.2 to 2 milliliters.

Ren himself acknowledged that his 2024 estimates should not be used to describe today's AI systems. The problem is fundamental: AI efficiency changes rapidly, and generalizing from one model to all models is scientifically inaccurate.

"We cannot just use a number from two years ago to describe today's system. And we cannot use specific models' results to generalize other models. I could give you a super efficient AI model that uses almost zero resources, or I could give you a very large model that can. It's just never correct to say, 'AI uses this much water,'" Ren stated.

Shaolei Ren, Associate Professor of Electrical and Computer Engineering at University of California, Riverside

Ren added that ChatGPT itself is not a single model with universal resource needs. "ChatGPT is not a single model. Nobody knows what exactly they're doing under the hood. So I don't know the resource efficiency," he explained.

Ren

How Does Data Center Efficiency Actually Work?

A chatbot's water and energy efficiency depends on multiple interconnected factors that engineers can optimize independently. Understanding these variables helps explain why blanket statements about AI water use are misleading:

• Hardware Optimization: Engineers design and select processors and cooling systems that minimize energy consumption and water use for specific workloads.
• Algorithm Efficiency: The underlying AI algorithms themselves can be optimized to perform the same task with fewer computational steps, reducing power demands.
• Workload Scheduling: Data centers can schedule when and how AI models run, batching requests and timing computations to use resources more efficiently.
• Operating Mode Variations: The same AI model uses dramatically different amounts of resources depending on whether it's in thinking mode, reasoning mode, or generating plain text output.

Because these variables can change independently and frequently, efficiency can shift dramatically over time. A model that was resource-intensive two years ago might be optimized significantly today.

Why Does Computing Equipment Change So Fast?

Jonathan Koomey, an independent researcher formerly at Lawrence Berkeley National Laboratory who has studied data centers for more than 25 years, emphasized that computing technology evolves at a pace most people underestimate.

"One of the things about computing equipment that is hard for people to get their head around is that it changes fast. The problem is that three or four years is an eternity when it comes to computing equipment. Things have probably turned over a couple times in terms of the latest tech since then," Koomey noted.

Jonathan Koomey, Independent Researcher, Formerly at Lawrence Berkeley National Laboratory

This rapid evolution means that basing current AI efficiency estimates on research from several years ago is a recipe for misinformation. What was true in 2024 may no longer reflect reality in 2026.

How Much Water Do Different Data Centers Actually Use?

Data centers have extreme variability in water consumption, making averages nearly meaningless. The amount of water a facility uses depends heavily on its specific design, geographic location, and cooling methods.

For example, Meta's data center in Eagle Mountain, Utah used 35.1 million gallons of water in 2024. The facility spans 4.5 million square feet and consumed 1.1 million megawatt hours of energy that year. Meanwhile, Meta's Prineville, Oregon data center, which is roughly the same size at 4.6 million square feet, had significantly different water consumption patterns due to its location and cooling design.

Koomey stressed that "averages don't mean anything. It's very specific to data center designs and their locations." This specificity is crucial for anyone trying to understand the real environmental impact of AI infrastructure.

Koomey

What's the Takeaway for Public Concern?

The gap between viral claims and scientific reality has created genuine confusion among the public. In Utah, where the proposed Stratos data center project has sparked community concern, more than half of residents oppose the facility according to a Deseret News-Hinckley Institute poll conducted in mid-May. Many residents worry about water use, electricity costs, emissions, and heat in the region.

These concerns are not baseless, but they are often informed by outdated or misunderstood information. The real story is more nuanced: data center water and energy use varies dramatically depending on design and location, efficiency improvements happen constantly, and making sweeping claims about AI's resource consumption is scientifically inaccurate. Understanding these distinctions is essential for having informed public conversations about AI infrastructure and its environmental impact.