Is AI really costing us all this water?

The water panic around AI is the first environmental argument I’ve watched go viral before anyone agreed on the units.

I was with friends the other day, mostly non-techies. The second AI came up, the talk went straight to water. Not the models. Not what it can do. Water. Like we’re all going to run dry because somebody asked ChatGPT to rewrite an email.

This is the plastic-straw fight of AI. Easy to picture. Easy to shame. Mostly the wrong unit. People dunk on a ChatGPT reply the way we used to dunk on a straw in an iced coffee, while the real water story sits in the same cup: the almond milk. One California almond is about 1.5 gallons of fresh water. A normal text prompt is teaspoons and milliliters. An AI image is closer to a sip. The almond in the latte still wins most days.

Sam Altman put an average ChatGPT query at about 0.000085 gallons, roughly a fifteenth of a teaspoon. On a podcast he floated about 38,000 queries per almond. PolitiFact called that Mostly False, and fair enough. Fresh-water almond math plus independent prompt ranges lands closer to about 1,000 to 10,000 conversational prompts per almond. GPT-4o work that counted cooling and power water put one prompt somewhere between 0.6 and 17 milliliters. EcoLogits has GPT-5.5 writing an email around 6.11 milliliters. Longer agent jobs can run more than ten times a normal chat. The shower-for-six-hours story was never serious. I’m not here to defend Sam’s spreadsheet either.

That still leaves the part people are actually pointing at when they say “AI is using all this water.” Not your one reply. The stack at scale.

Per query being small doesn’t make the industry small. Hundreds of millions of people querying every day is a different problem than one dinner chat. Training still uses a lot of water. Data centers still take treated water in places that don’t have it to spare. Campuses keep getting bigger. CapEx, power, land, and local politics travel with that buildout.

Farms still dwarf data centers nationally. Crop irrigation alone was nearly 30 trillion gallons in 2020. Data centers were about 228 billion in 2023. That gap doesn’t erase local fights over a new cooling plant next to a dry town. It does mean treating your Tuesday prompt like it emptied a reservoir is the straw again. The buildout is the cup.

The bill for AI is going up, and part of that bill is water. The other side is what we could do with the same stack if we aim it: farming models, supply chains that waste less food, health, climate work, the boring logistics of getting food and medicine where they’re needed. I’m optimistic about that. Potential still isn’t a guarantee. Bad choices can waste a good tool. A conversation that stops at the straw and never asks what we build with this is only half a conversation.

I left that dinner stuck trying to say that other half. Instead we stayed on water and almonds. Fine. Keep the units honest. Then put more of the energy on what we build with it.

Cheers,