How Does AI Use Water? Data Centers Explained

How does AI use water? Almost all of it goes to cooling. The computers that run a chatbot turn electricity into heat, and the buildings that hold them, data centers, have to get rid of that heat. Some do it by blowing air across the equipment, which uses a lot of electricity and very little water. Many of the large ones instead pump warm water through a cooling tower, where part of it evaporates and carries the heat off into the air. That evaporated water is the direct use. A second, larger use hides behind the plug, because the power plant generating the electricity also evaporates water to cool itself.

A single prompt uses a tiny amount, somewhere between a few drops and a few teaspoons depending on who's counting and what they count. The aggregate is large, and the strain is local. I'll walk one prompt through both pipes, then set the published numbers next to the "gallon per prompt" claim you've probably seen.

One Prompt, Two Pipes

Start inside the building.

When you send a message, a rack of specialised chips does the arithmetic that produces the reply. If you want the mechanics of that step, the guide to how AI models work covers it without maths. For this article the only thing that matters is that every watt of electricity the chips draw leaves as heat, and heat has to go somewhere.

The cheapest place to send it, in energy terms, is into water. This is the step I'd want you to picture, because everything else follows from it. A cooling tower sprays warm water through moving air, some of that water evaporates, the evaporation carries a great deal of heat away with it, and the water that remains comes back cooler and goes round again. It can't go round forever. Each pass leaves the minerals behind, and the Li et al. paper that first put numbers on AI's water use notes that tower water is typically recycled only 3 to 10 times before it has to be discharged to prevent mineral and salt buildup. Fresh water is added constantly to replace what evaporated and what was flushed. That make-up water has to be clean. In many cases it's drinking water, because dirty water clogs pipes and grows bacteria, and a clogged tower is a dead data center.

Now the plug.

Much of the electricity in the United States comes from plants that boil water into steam to spin a turbine, and those plants need cooling too. The US Geological Survey defines thermoelectric power water use as water used in generating electricity with steam-driven turbine generators. Li et al. sort AI's water into three scopes, and the plant is the second one.

The number I'd remember is that the second row is usually bigger than the first. A data center that runs efficiently on site can still be a heavy water user through its electricity, which is why the Berkeley Lab report found the indirect footprint of US data centers to be nearly 800 billion litres in 2023 against 66 billion litres directly.

Withdrawal Is Not Consumption

Two words do most of the work in any honest water number, and headlines mix them up constantly.

The USGS defines withdrawal as water removed from the ground or diverted from a surface source for use. It defines consumptive use as the part of withdrawn water that's evaporated, transpired, put into products, or otherwise not available for immediate reuse. A cooling tower withdraws the make-up water it adds and consumes the part that evaporates. Li et al. put the split at roughly 80 percent evaporated for a tower with good water quality, with the rest discharged.

Power plants sit at the other extreme. Li et al. cite US averages of about 43.8 litres withdrawn per kWh but only 3.1 litres consumed, because most plant cooling water goes straight back to the river or bay, warmer than it left. So a plant "uses" fourteen times more water than it "uses." Both figures are correct. When a number crosses my screen, that's the first thing I check.

I'd add one more distinction. Evaporated water isn't destroyed. It falls as rain somewhere, which is cold comfort for the aquifer it came from, and the reason water stress is a local problem rather than a global one.

How Much Water Does One Prompt Use?

Here are the published figures I could trace to a primary source, with what each one counted.

The calculation underneath is short, and I'd redo it yourself, because it's the part that makes the meme fall apart. Google's 0.26 ml divided by 0.24 Wh is 1.08 litres per kWh, which matches the roughly 1 litre per kWh that Li et al. cite as Google's fleet-wide on-site figure. Add the plant at Berkeley Lab's 4.52 litres per kWh and the same prompt costs about 1.3 ml. Altman's figure works out to 0.95 litres per kWh, so it looks like an on-site number, though his post doesn't say. Li et al. land at 10 to 50 ml for GPT-3, a model published in 2020, and they counted the plant. Efficiency moves quickly here. Google reports that the energy per median prompt fell 33 times over a single year, so a 2023 estimate for a 2020 model and a 2025 estimate for a current one were never going to agree.

So the "gallon per prompt" claim is off by a factor of 76 to 379 against the highest published figure, and by a factor of about 14,500 against the lowest. It isn't a rounding error. It's a different unit.

Two hedges belong here. Google's number is a median for a text prompt, and it's Google's own estimate. A long document summary or an image request will cost more, and the tokens explainer shows why output length is the lever. Altman's number carries no method at all, so I'd treat it as a claim rather than a measurement.

Why Is ChatGPT Wasting Water?

Per prompt, it isn't, by any figure in the table. The reason the question exists is that prompts are counted in the billions and data centers cluster in particular places.

The Berkeley Lab report puts direct consumption by all US data centers, AI and otherwise, at 66 billion litres in 2023, up from 21.2 billion in 2014, and expects hyperscale sites alone to consume 60 to 124 billion litres in 2028. Li et al., citing one technology company's own sustainability report, say its self-owned data centers evaporated more than 23 billion litres in 2023, nearly 80 percent of it drinking water, with consumption rising 17 to 20 percent a year.

What makes that a problem in one town and a non-event in another is timing and place, because a tower evaporates hardest on the hot, dry afternoons when the local utility is already near its limit, and a campus that draws steadily all year still hits the aquifer hardest in the one week it can least afford. Li et al.'s range of 1 to 9 litres per kWh has the Arizona summer at the top end for exactly that reason. The Environmental Law Institute's blog on data center cooling, which ranks for this search, makes the same point in plain words. For many communities the challenge isn't the total, it's when and where the water is used and who pays when supplies run short.

My concern is less the chatbot and more the disclosure. Li et al. point out that carbon figures now appear on model cards while water figures, on site or off, mostly don't. You can't judge a number nobody publishes.

How Much Water Does AI Use per Day?

Nobody has an AI-only daily figure. The same buildings serve video streaming, email, payroll systems and everything else on the internet, and none of the public reports separate the AI share out, so the best available number is for all US data centers together.

I want to be careful with this table. The two USGS rows are withdrawals, the two Berkeley Lab rows are consumption, and the years differ, so this is a scale comparison and nothing finer. Even so, the scale is clear. US power plants withdraw in one day about 7.6 times what every US data center evaporates in a year. Thermoelectric power was 41 percent of all US withdrawals in the last national USGS compilation, and power, irrigation and public supply together were 90 percent.

That's the finding behind the CBS headline in the same search results, that the heaviest demand on American water is farming, lawns and flushing rather than data centers. Nationally that's true. It also isn't the question a county with a new campus and a falling water table is asking.

Why Can't AI Use Ocean Water?

Sometimes it does. Google's data center in Hamina, Finland, built inside a former paper mill and opened in 2011, cools with seawater from the Bay of Finland and feeds the waste heat to a heat recovery facility. So the honest answer is that seawater works when the building is next to the sea and designed for it.

Most aren't. Data centers get sited near cheap power, fibre and customers, which usually means inland, and my guess is that the economics of the site win over the cooling design nearly every time. And an evaporative tower is the wrong machine for salt water. Remember that even fresh water has to be flushed after 3 to 10 cycles because minerals concentrate as the pure water evaporates. Salt concentrates the same way, faster, and it corrodes what it touches. A seawater design has to move heat without evaporating the water, which is a different and more expensive build. I couldn't find a public per-site count of how many facilities do it, so treat Hamina as proof it's possible rather than evidence it's common.

Does AI Pollute the Water It Uses?

Mostly not at the data center, with two exceptions worth knowing.

The evaporated share leaves as vapour and comes back as rain. The discharged share, the flush after those 3 to 10 cycles, carries the concentrated minerals and whatever was added to keep the tower clean, and goes to treatment or a sewer. Li et al. describe the discharge in terms of mineral and salt buildup rather than toxicity.

The exceptions are upstream. Power plants return most of their cooling water warmer than they took it, which is a separate question for that river even though the water is otherwise unchanged. And chip manufacturing, the third scope, is where Li et al. say discharged water may contain toxic chemicals or hazardous waste, with recycling rates at many plants still low. If pollution is your worry, I'd look at the fab before the server hall.

How Bad Is AI for the Environment Overall?

Water is one line of a longer bill, and the biggest line is electricity. Berkeley Lab estimates US data centers used 176 TWh in 2023, or 4.4 percent of national electricity, and projects 6.7 to 12 percent by 2028. That electricity is where the off-site water comes from, and it's also where the carbon comes from.

There's a trap in trying to fix one line at a time. A data center can switch to dry coolers and use no water on site at all, and Li et al. note that this typically raises cooling electricity, which raises the water consumed at the plant. Optimising for water can cost energy and optimising for energy can cost water. Which trade is right depends on whether the site sits in a desert or next to a hydro dam. I don't think a national rule can answer that.

The broader harms, from jobs to misinformation, are a separate piece. The article on why people say AI is bad sets the environmental cost beside the others so you can weigh them.

What You Can Reasonably Do

I wouldn't ration prompts out of guilt. At five drops or even fifty millilitres, one person's chatbot use isn't where the water goes, and shorter prompts and shorter answers already use less of everything.

Do care about siting and disclosure, because that's where the numbers move. If a provider you pay publishes water usage effectiveness for both the site and the source, that's a company you can hold to a figure. If a data center is proposed near you, the questions that matter are the cooling type, the water source, the projected litres per kWh, and what happens on the hottest week of the year. Those four answers say more than any national average.

And when the next "one prompt equals a bottle of water" post comes round, you have the table. Use it. Ask which model, which year, and whether the plant was counted. The honest figure has a scope attached. The viral one never does.