Every time you type a prompt into ChatGPT, a little water evaporates somewhere. And no, OpenAI isn’t going to put that on a dashboard for you. The ChatGPT water footprint is one of those inconvenient numbers that gets buried under press releases about new model launches.
Let’s cut through the noise. Depending on the model, the server location and even the time of day, a single interaction can gulp down 10 to 25 milliliters of drinking water. Run 20 to 50 questions and you’ve basically thrown away half a liter. That’s a bottle of water gone before you finish your coffee.
The Numbers Keep Moving (And Upward)
A 2023 study claimed that 10 to 50 queries on GPT-3 used around 500 ml. Then the same researchers went back, checked their math, and realized GPT-3’s real energy consumption was at least four times higher than they’d estimated. The corrected figure jumped to roughly 2 liters of water for that same batch of questions.
The complexity of your request matters too. Generating a simple 100-word email with GPT-4 can consume around 519 milliliters of water. That’s more than a small bottle, for an email you could’ve written yourself in 90 seconds.
But here’s the catch: newer models like GPT-3.5 and GPT-4 are actually more efficient per parameter than GPT-3. Some technical analyses argue a typical conversation might use only 5 milliliters. Sam Altman himself said in 2025 that an average query used just 0.3 ml. The gap between 500 ml and 0.3 ml is absurd, and it tells you exactly how opaque these companies are about their real AI data center water consumption.
Where the Water Actually Goes
None of this water is used to «think.» It’s used to stop servers from melting. Data centers are giant warehouses full of machines running 24/7, processing billions of calculations that generate brutal heat.
Most facilities rely on evaporative cooling. Picture a tower where water circulates and evaporates to carry heat outside. The problem? That evaporated water is gone forever. It doesn’t come back. And to protect sensitive hardware, it almost always has to be drinking water. Untreated rainwater won’t cut it.
Then there’s the indirect cost. A big chunk of the water footprint comes from the power plants burning coal or gas to generate the electricity those servers need. Those plants have their own cooling towers, evaporating their own water. So every prompt hits the water supply twice.
Location Changes Everything
Where the server sits is a huge variable. Asking ChatGPT from Texas is not the same as asking from Washington state. That same 100-word email consumed about 235 ml from a Texas data center. From Washington, it sucked up nearly 1.5 liters. Same prompt, wildly different damage.
Local climate and infrastructure efficiency decide the bill. Training is where the real carnage happens: GPT-3’s training consumed around 700,000 liters of fresh water in US data centers. Had that training run in Asia, with less efficient infrastructure, consumption would have tripled. For GPT-4, the training water footprint is estimated to be ten times larger.
The Scale Is What Should Scare You
Per-question costs look tiny until you multiply them by hundreds of millions of daily users. One recent estimate put ChatGPT’s daily water consumption at around 148 million liters. That’s like every person in Taiwan flushing a toilet at the same time. Over a year, it’d fill New York’s Central Park reservoir seven times.
And this is where it gets ugly: that water competes directly with people. In the US, one in five data centers draws water from areas already under water stress. In The Dalles, Oregon, Google’s servers consume almost a quarter of the city’s entire water supply. In Chile, a Google data center in Santiago used 7.6 million liters of drinking water per day in a region gutted by megadrought.
Big tech reports double-digit annual increases in water use. Microsoft up 22.5%, Google and Meta up 17%. Amazon doesn’t even publish total figures, which researchers read as a way to dodge scrutiny.
Fixes Exist, But They’re Mostly Band-Aids
Some companies, like Digital Realty, use non-drinking water such as rainwater for cooling. Others push precision liquid cooling, where a special fluid cools chips directly and slashes water needs. Moving data centers to cold northern Europe helps, but it slows response times.
The real fix is renewable energy — solar and wind don’t need water to make electricity, unlike coal and gas. Training models at night or during cold seasons also cuts evaporation in cooling towers.
But none of this matters without transparency and regulation. No international law requires companies to report or cap this consumption. Meanwhile, projections say the AI industry could need between 4.2 and 6.6 billion cubic meters of water by 2027 — the annual consumption of a country like Denmark.
Next time you ask ChatGPT to summarize an article or crack a joke, remember: behind those words there’s a sip of a shrinking resource. The real cost isn’t your subscription. It’s the glass of water someone else won’t get to drink.
