← All Writing

From the Desk  ·  Water  ·  August 31, 2026

Is AI Actually Wasting Water?

Last time, we gave the data-center debate a ruler.

87 million gallons. Compared with what? An airport. A third of a farm. Four golf courses.

That answered the first question. Now for the one that actually matters:

So what?

Is all that water wasted? It evaporates, right? Doesn't it just come back as rain?

The answer fits in one line: globally, yes. Locally, no.

And water policy is local.

Where the water goes

Most data-center cooling works by evaporation. Water absorbs the heat, turns to vapor, and leaves.

That is called consumptive use, and the name is doing real work. The water is not destroyed - it rejoins the planet's water cycle and rains somewhere. Just mostly not back into the river or aquifer it came from. (An aquifer is the underground rock a county's wells drink from. It refills slowly, and only from local rain.)

Your county's aquifer does not care about the planet's water cycle.

Withdrawn is not consumed, and the whole debate keeps slipping on the difference. Most water a facility withdraws goes back into the local supply. The best estimate is that 45 to 60 percent of withdrawals end up consumed. A headline that says "used" could mean either, and the difference is roughly two-fold.

The national number is a shrug

In 2023, U.S. data centers directly consumed about 17.5 billion gallons of water - around 0.3 percent of the public water supply. Even counting the water power plants use to make their electricity, data centers come in under 1 percent of the country's consumptive water use.

Nationally, this is not a crisis. Anyone who tells you the country is running dry because of chatbots is selling something.

The headlines earn about half their panic

Overstated, with receipts:

• The viral claim that a ChatGPT conversation drinks a 500ml bottle of water traces to one modeled scenario for an older model. It was never measured. A careful recheck lands around 5 milliliters per prompt - a hundred times smaller. That is not a bottle. That is a sip.
• About 80 percent of the "AI water use" in scary headlines is indirect - water consumed at power plants. Every industry on the grid shares that number. AI just got the headline.
• The Georgia story about a data center drying up a family's taps? The taps ran dry from construction sediment before the center ever operated - and it doesn't draw local groundwater. The real story was bad construction practice. The water story was wrong.

So yes, some of this is sensationalism. A number without a benchmark is a press release, and panic without a measurement is a headline.

"A number without a benchmark is a press release, and panic without a measurement is a headline."

The local number is not

The other direction does not flinch.

Two-thirds of the data centers built since 2022 sit in places already classified as high water-stress - meaning local demand already runs close to local supply. The Memphis Sand Aquifer is being pumped faster than it recharges. And in Newton County, Georgia - Georgia, where some of you are reading this - proposed data centers have requested more water per day than the entire county currently uses.

Read that again. Not more than the county's data centers. More than the county.

Timing makes it worse. Data centers consume the most water on the hottest days - exactly when a drought has the least to spare. The peak of the demand lands on the peak of the shortage, every summer, on schedule.

And there is a number nobody headlines: construction. Meta's own environmental index excludes 1,019 megaliters of water withdrawn for construction in 2024 - about 269 million gallons. That is roughly twelve times what its Prineville facility used all year, sitting outside the headline figure.

Then there is the data void. Meta is the outlier because Meta publishes. Most operators publish nothing verifiable, and the gap between press release and reality can be absurd: one Stargate figure in circulation says a building will use 12,625 gallons a year, while the local mayor says 10 million. Those numbers are a hundred times apart. Both cannot be the ruler.

So: overstated or understated?

Both. Nationally overstated, locally understated, and the average of the two is a lie.

"0.3 percent of national supply" is a true sentence that hides the county where the number is 30 percent. "AI is draining our water" is a false sentence that happens to be true in specific watersheds with specific names.

"Is AI wasting water?" has no national answer. It only has a list of local ones.

"Nationally overstated, locally understated, and the average of the two is a lie."

That is the ruler argument extended one level. Compared with what - and now: where, when, and on whose aquifer?

The space option

Elon Musk has an answer to the water question, and it is orbit.

In January, SpaceX filed with the Federal Communications Commission - the agency that licenses what goes into orbit - for a constellation of up to one million data-center satellites. The pitch writes itself: in space, cooling uses zero water - because vacuum. No air, no evaporation, no aquifer.

The catch writes itself too. In a vacuum there is no air and no water to carry heat away. The only option is radiators - giant panels that glow the heat off into space - and they are enormous. The International Space Station needs six metric tons of them to shed 70 kilowatts; each proposed satellite must shed 120. The all-in cost of the constellation pencils out somewhere between $1.45 trillion and $9.8 trillion - and only if launch costs fall to a number nobody has ever achieved. The chips last about five years in orbit, so you re-launch thousands of times a year forever just to stay even.

It is the most expensive way imaginable to avoid the water question.

File it where it belongs: not a plan, a receipt. When the water fight gets loud enough, trillion-dollar radiators start sounding reasonable to somebody.

The move

Same as before. You do not need to love data centers, and you do not need to panic about them.

You need the local numbers.

Is there an aquifer under the proposal, and is it recharging faster than it's pumped? What does the facility withdraw, what does it consume, and what does it put back? What does it use on the hottest day of the year, not the average day? And what did the construction phase drink before anyone turned the servers on?

If the answers exist, they exist at your county meeting, not in a national headline.

Ask for them there.

More From the Desk