The AI water scare is built on misleading numbers and missing context, one writer argues

AI data centers consume roughly 0.04 percent of all freshwater used in the United States each year. That figure, including both water used inside data centers and water used by power plants to generate their electricity, comes from Andy Masley, writing for his own publication. He argues that fears about AI draining water supplies are based on missing context, misleading statistics, and a misunderstanding of how water is used across the American economy.

Masley is explicit about the scope of his argument: he is not addressing AI’s electricity demands, which he calls a “very real problem.” His focus is water alone, and on that question he concludes the numbers do not support the alarm.

What the numbers actually show

In 2023, all US data centers combined used between 200 and 250 million gallons of freshwater per day. The country as a whole consumed around 132 billion gallons daily. That puts the entire data center industry, which supports not just AI but the whole internet, at roughly 0.2 percent of national freshwater consumption.

Of that total, only 50 million gallons per day were used onsite inside the data centers themselves. The rest was consumed offsite at power plants generating the electricity those data centers draw. AI accounts for approximately 20 percent of data center power use, meaning AI-specific water consumption sits at around 0.008 percent of US freshwater when measured at the data center level.

On a personal level, Masley points to research suggesting each AI prompt uses roughly 2 milliliters of water when onsite and offsite consumption are combined. The average American, he writes, uses enough water every day for 800,000 such prompts. A pair of jeans requires the water equivalent of 5.4 million prompts to manufacture.

Why the coverage looks so alarming

Masley identifies several patterns in media coverage that he says systematically distort the picture.

• Articles compare data center water use to the number of households supplied, without noting that household use represents less than 8 percent of an average American’s total water footprint.
• Reports cite maximum permitted water volumes rather than actual usage. Permits are set for worst-case scenarios and rarely reflect real consumption.
• Stories about construction-related water problems near data centers are framed as examples of data centers draining local supplies during normal operation.
• Large raw numbers are cited without comparison to other industries such as agriculture, golf courses, or steel production.

One widely shared Washington Post article claimed that writing an email with ChatGPT uses a full bottle of water. Masley examines the underlying study and argues the figure is only reachable by stacking multiple unrealistic worst-case assumptions simultaneously, including the idea that users query the model ten or more times per email and that efficiency has not improved since 2020.

A separate case frequently cited as evidence of data center harm involved residents near a Meta facility in Georgia whose wells ran dry. Masley notes that the data center had not yet begun operating when the problem occurred. Construction sediment had contaminated nearby groundwater. The facility was not even designed to draw from local groundwater at all.

Local concerns and the case of desert cities

Masley does not dismiss local planning concerns entirely. He acknowledges that data centers, like any large industrial facility, require careful siting. He found one partial exception to his broader argument: Newton County, Georgia, where county documents cite a Meta campus expansion as one factor among several driving water infrastructure costs upward.

For high-stress regions like Maricopa County in Arizona, home to Phoenix, he argues the relevant comparison is not between data centers and individual households but between data centers and other industries competing for the same water. In Maricopa County, golf courses use roughly 30 times as much water as all local data centers combined, while generating far less tax revenue per gallon. Masley calculates that data centers produce approximately 50 times more tax revenue per unit of water than golf courses in the county.

He also points out that water cooling is often environmentally preferable to air cooling. One study he cites found that switching from air to liquid cooling reduces a data center’s total power consumption by around 10 percent, lowering carbon emissions in the process.

Finally, Masley argues that AI tools have in documented cases saved far more water than data centers consume. A single irrigation optimization tool saved South American farmers 19 billion gallons of water over two years, roughly twice the amount all American AI data centers used in the same period.

His conclusion is pointed: the water footprint of AI is so small relative to everyday activities and other industries that individual users changing their AI habits to protect water supplies would have no measurable effect. The debate, he argues, should focus on planning decisions made by ecologists, economists, and city officials, not on individual consumption choices driven by misleading statistics.