Skip to content
MaungaAI news, explained

Is it true?

Is it true that every ChatGPT question uses up a bottle of water?

By the end, you'll know where the famous bottle of water comes from, what figures exist on the water and energy each chatbot question uses, and why those figures contradict each other so much.

Published · Spanish original

Translated with AI from the Spanish original. Nobody has reviewed this translation; if anything seems off, the Spanish version is the reference.

Partly. The bottle comes from a real estimate for long answers from a 2023 model. For a short question, the available figures are much lower, but they come from the companies themselves and no one has checked them independently.

Where the claim comes from

The idea of "a bottle of water per question" has a specific origin. In April 2023, four researchers led by Shaolei Ren, of the University of California, Riverside, published a study called "Making AI Less Thirsty". In it, they estimated how much water it takes to run an AI model like the ones behind ChatGPT.

Their most quoted calculation said that GPT-3, an OpenAI model of the time, used about 500 millilitres of water, roughly a small bottle, for every 10 to 50 medium-length answers. The exact figure depended on where the data centre was and on the time of year. In other words, according to that study, the bottle wasn't per question but per several dozen.

The "one bottle per answer" version caught on in September 2024, with an article in The Washington Post produced with the same researchers. It calculated that writing a 100-word email with GPT-4 could use 519 millilitres of water. But that is one specific case (a long text, a large model and a particular set of assumptions), and on social media it ended up summed up as "one bottle for every question".

Why AI uses water

When you ask ChatGPT something, the answer doesn't come from your phone. It is worked out in a data centre: a building full of very powerful computers that run nonstop. Those computers use a lot of electricity and get hot, and a common way to cool them is with water that evaporates, much like sweat cools us down.

There is also a second, hidden cost: the water used to produce the electricity those centres consume, for example at thermal or nuclear power stations, which are also cooled with water. This is often called "indirect" water. The Riverside study adds the two together: the water used inside the data centre and the water used at the power stations.

That's why the amount varies so much from place to place. A data centre in a hot, dry area, or connected to a power grid that relies on thermal power stations, uses far more water than one in a cold climate or powered by solar or wind energy.

What the data says

Since 2025, some companies have published their own figures per question, and they are much lower. In June 2025, Sam Altman, head of OpenAI, wrote on his blog that an average ChatGPT query uses about 0.34 watt-hours of electricity and about 0.32 millilitres of water, "roughly one fifteenth of a teaspoon". Altman didn't explain how the calculation was done or what it includes.

In August 2025, Google published a technical study on its Gemini assistant, with a detailed method. According to Google, the typical text question (the median, meaning the one in the middle: half use more and half use less) used about 0.24 watt-hours and 0.26 millilitres of water in May 2025, around five drops. This is what Google says about itself: it is not an independent measurement.

That study was criticised by researchers, including Shaolei Ren himself. The main criticism: Google only counts the water used inside its data centres and leaves out the water used at power stations, which the Riverside study does include. On top of that, the median hides the most costly questions, such as those that require a lot of reasoning, analysing long documents or creating images.

To put the figures in context, the International Energy Agency, an international public body, estimated in April 2025 that data centres used around 1.5% of the world's electricity in 2024. That covers the whole internet, not just AI, although the agency expects that consumption to grow sharply in the coming years, mainly because of AI.

In short: the bottle of water comes from a real estimate, but for long answers, from a 2023 model and also counting the water used at power stations. The figures companies give for a short question are hundreds or thousands of times lower, but no outsider has checked them and they don't count the same things.

What we don't know

There is no independent measurement of how much water a ChatGPT question uses. OpenAI hasn't published its method, and companies aren't required to report this data in a way that allows comparison between them. Each study decides what to count (direct water, indirect water, the water used to train the model), and that explains much of the gap between one set of figures and another.

It's also unclear how much the heaviest questions use. Models that "think" before answering, ones that generate images or video, and very long conversations use a lot more than a short question, but there is hardly any public data on them.

And there's something the per-question figure misses: the real impact of water use is local. A drop per question multiplied by hundreds of millions of questions a day is concentrated in a handful of data centres. If they are in an area hit by drought, the problem for local residents can be serious, even if each user's share seems tiny.

Watch out: Any figure for water per question depends on what is counted and where the data centre is. Be wary of both the bottle and the five drops if no one tells you how the number was calculated.

Sources

More guides

All guides