How much water does AI really use?

(CBS)– Google says a typical AI query uses five drops of water.

OpenAI’s Sam Altman describes a similar amount — about one-fifteenth of a teaspoon.

But another viral estimate says a short email written with AI’s help uses a half-liter bottle of water.

The difference is enormous: across those three widely shared claims, the largest amount is about 2,000 times the smallest.

None of them is fully right. They’re not measuring the same thing. Some estimates count only the water used to cool data centers. Others add the water used to generate the electricity that powers them. Still others rely on outdated assumptions about fast-changing technology.

As figures ricochet across blogs, social media, protest signs and public meetings, nuance is lost.

That leaves a basic question: How much water does AI use?

The answer can depend on the AI model, the data center’s location and the mix of sources in the local power grid. In some cases, amounts can differ by hundreds of times over.

To sort through the claims, we reviewed the scientific research behind some widely shared figures about AI’s thirst for water.

We also compared estimates with water used for other products and services people eat, buy and use every day, including the viral statistic that producing a single hamburger uses more water than thousands of AI prompts. That one is true, no matter how you count.

In the end, while data centers consume a lot of water, the data show the biggest demands on the country’s water are activities that make up everyday life, from growing food to watering lawns to bathing.

Here are five charts that put AI’s water consumption in context.

The total water consumption of data centers in the US and future projections

Researchers cannot directly measure all of AI’s water consumption, but they can estimate how much is used by the data centers that power it.

Scientists at Lawrence Berkeley National Laboratory estimate U.S. data centers used about 228 billion gallons of water in 2023. About 17 billion gallons went to cooling servers. Another 211 billion gallons were tied to generating the electricity that powered them.

The researchers project the total could grow to between 469 billion and 844 billion gallons by 2028, depending on factors ranging from data-center technology to the mix of energy sources feeding the electric grid.

But AI accounts for only part of that demand. The International Energy Agency estimates AI accounts for 15 to 20 percent of data center electricity demand, and by extension, its water use. The rest is accounted for in supporting other digital activity, from satellite data that helps predict the weather and protect troops to billions of us streaming movies, television and TikTok videos.

What are the other ways we use similar amounts or much more water in the US?

For people trying to understand AI’s impact on water supplies, scale matters.

Compared with many other demands on water nationwide, today’s data centers consume relatively modest amounts. Americans use more water to irrigate lawns, flush toilets and produce food — especially meat — than all U.S. data centers consume combined.

The comparison becomes even more striking when looking only at the share of data-center water usage attributed to AI.

The portion of data centers’ cooling water attributed to just AI processing is less than the water used for one year’s worth of golf-course irrigation, vehicle washing, residential swimming pools and restaurant dishwashing, based on estimates derived by government and industry.

Those comparisons do not include the additional water associated with generating electricity. Nor do they mean AI’s water use is insignificant. They just help place data center water usage within the many ways Americans consume the resource every day.

A lot of water goes into growing food — for people and animals

Scientists estimate U.S. crop production used about 183 trillion gallons of water in 2019.

Coffee is among the thirstiest of crops grown in the U.S., consuming more than 1,200 gallons of water per pound produced. However, almost all of that comes from rainwater.

Water taken from rivers and wells might be a better comparison with water-cooled data centers, and nut trees are among the least efficient users. Certain nuts in the U.S., such as macadamia and pine nuts, can consume more than 300 gallons of irrigation water per pound.

Scale matters a lot and the U.S. grows very little of either crop when compared to the entire industry.

Combined, nuts make up less than 1% of crops grown and less than 4% of irrigation water used. Yet, that 4% still adds up to 538 billion gallons of water – more than twice the water AI data centers are using.

In aggregate, the thirstiest food grown in the U.S. isn’t food directly consumed by humans.

Grasses and grains, bound for animal feed, make up a staggering amount of the agricultural product of the U.S. The country grows 1.7 trillion pounds of fodder crops a year, consuming 5.5 trillion gallons of irrigation water, according to researchers at the University of Twente in 2024.

Agriculture uses far more water than data centers

Even counting only water consumed from irrigation, agriculture uses many times more water than data centers. Fodder crops (food grown for animal feed) alone use 5.5 trillion gallons of irrigation water per year.

Other

Sugar

Fruits &

Vegetables

Oil & Nuts

Fibres

Cereal

Fodder

Data centers:

228B gallons

So does this mean we should be more concerned about the water used by the livestock industry than by data centers or AI? Economist David Zetland said comparing different activities that consume large amounts of water is the wrong way to think about it.

“On the margin,” he said, “the straw that breaks the camel’s back is only gonna break the back because all the other straw is on the back first.”

He said people try to connect water use to a range of consumer products because they want a villain, but the real problem is water for all purposes is far too cheap — or free — and doesn’t include the cost of sustainable management.

“The value of water is radically higher than the price we pay for it,” said Zetland. “If you don’t put a price on it, then every economist will tell you it’s gonna be over soon. And that’s what we’re seeing right now.”

Where data centers are built matters

Where an AI data center is located can matter a lot when it comes to water.

The impact on water supplies can vary dramatically from one region to another because data centers depend on local electricity grids, water resources and climate conditions.

Researchers have identified several states as particularly favorable for data-center development from a water perspective. But many existing facilities are concentrated elsewhere. A CBS News analysis of locations tracked by Data Center Map found large concentrations of data centers in water-stressed states including California, Arizona and New Mexico.

In one recent study, researchers identified Texas, Nebraska, South Dakota, Louisiana and Idaho among the most favorable states for future development. A common factor was access to large amounts of wind and solar power, which require little water compared with many other sources of electricity.

But even the most favorable locations come with tradeoffs. Texas, Nebraska and South Dakota all face challenges expanding the capacity of their electric grids to support more data centers.

Water use is only part of AI’s environmental footprint

When it comes to AI and data centers, some researchers argue electricity may be the more important environmental concern.

Data centers accounted for an estimated 4% of U.S. electricity consumption in 2023. Studies estimate that could reach 12% by 2028 as demand for AI computing grows.

In many regions, utilities are already struggling to connect new data centers because of a shortage of additional transmission lines and generation capacity.

That growth carries consequences beyond grid strain and higher utility bills. Producing electricity emits fine particulate matter and other pollutants from power plants and backup generators.

The increased pollution falls disproportionately on people living near power plants and data centers, said Shaolei Ren, a professor of electrical and computer engineering at the University of California, Riverside. Ren and his colleagues estimate that data centers, driven by AI’s growing computing workload, could contribute to hundreds of thousands of asthma symptom cases and more than 1,000 premature deaths annually by 2028.

Other concerns tend to be highly local. In northern Virginia, the nation’s largest data-center hub, facilities already occupy 6,200 acres — with another 21,000 acres of data centers proposed. That’s the size of nearly 16,000 football fields or one-and-a-half times the size of Manhattan.

Researchers have also raised questions about cooling water discharge causing pollution, though one recent study found “limited to non-existent data” on the impact so far. Nearby residents have raised concerns ranging from impacts on water wells to air, noise and light pollution.

Some technologists have proposed more radical solutions. Elon Musk has floated the idea of placing data centers in orbit, though experts see significant technical and economic hurdles. Another tech firm is experimenting with floating data centers in the world’s oceans.

No single number can capture AI’s water footprint

Many viral statistics about AI and water are rooted in real science. But as that research gets sliced into social media posts, signs and political talking points, complex findings become over simplified slogans.

The next time you see a claim about AI and water, the most important question may not be, “Is it true?”

It may be what’s being counted and how it fits into the bigger picture of U.S. water use.