It must be understood that this figure is not a fixed number like the price of a soda. It depends on factors such as the model you use, where the server that handles it is located, and even the time of day. It is estimated that each interaction can consume between 10 and 25 milliliters of water. That is, for every 20 to 50 questions, around half a liter of drinking water could be used.
But these figures are controversial and have been changing. A 2023 study mentioned that between 10 and 50 queries on GPT-3 used about 500 ml. However, the same researchers reviewed their calculations. They discovered that GPT-3’s actual energy consumption was at least four times higher than they had estimated. This brought the figure to approximately 2 liters of water for every set of between 10 and 50 questions.
The complexity of the request also alters the expenditure. Generating a 100-word email with GPT-4 can consume about 519 milliliters of water. To put it in perspective, that is more than a small water bottle for a simple email.
Where all that water really comes from
This water is not used to “think,” but to keep the servers from melting. Data centers are gigantic warehouses full of computers that are always on. Processing billions of calculations for AI generates extreme heat.
To cool them, most use systems that rely on evaporative cooling. Imagine a tower where water circulates and evaporates to carry the heat outside. The problem is that the evaporated water is not recovered. It is lost into the atmosphere. Furthermore, to protect sensitive hardware, this water almost always must be drinking water. Dirty or untreated rainwater will not do.
The impact does not end at the data center. A significant part of the water expenditure is indirect. It comes from the power plants that burn coal or gas to generate the electricity that powers these servers. Producing that electricity also requires evaporating enormous amounts of water in their own cooling towers.
Why the figures fluctuate so much: location, model, and more
The server’s location is key. Asking ChatGPT from Texas is not the same as asking from Washington. In Texas, generating that 100-word email used about 235 ml of water. The same request, handled from a data center in Washington, consumed almost 1.5 liters. The difference is abysmal and depends on the local climate and the efficiency of the infrastructure.
The model version makes another big difference. GPT-4, being more complex, was expected to consume more than GPT-3. But some analyses argue the opposite. A technical blog points out that newer models, such as GPT-3.5 and GPT-4, are much more efficient per parameter than GPT-3. They suggest that a typical conversation might use only about 5 milliliters, not 500. Even Sam Altman, of OpenAI, said in 2025 that an average query used 0.3 ml. The discrepancy is enormous and reflects how quickly technology evolves and how opaque companies are with their real data.
The training phase of the model is where most of the water is used at once. Training GPT-3 consumed around 700,000 liters of fresh water in data centers in the United States. If that training had been done in Asia, where the infrastructure may be less efficient, consumption would have tripled. For larger models like GPT-4, the water footprint of training is estimated to be ten times greater.
The real problem: the scale becomes inhumane
The expenditure per question seems small, but multiplied by hundreds of millions of daily users, the figure is colossal. A recent calculation estimated that ChatGPT could be consuming around 148 million liters of water per day. It is as if the entire population of Taiwan flushed the toilet at the same time. Per year, it would be enough water to fill the Central Park reservoir in New York seven times.
This consumption competes directly with people’s consumption. In the United States, one in five data centers draws water from areas already suffering from water stress. In The Dalles, Oregon, Google’s servers consume almost a quarter of all the water available in the city. In Chile, a Google data center in Santiago used 7.6 million liters of drinking water per day in a region ravaged by a megadrought.
Big tech companies report double-digit increases in their water consumption year after year. Microsoft increased by 22.5%, Google and Meta by 17%. Amazon does not even publish its total figures, which some researchers interpret as a way to evade scrutiny.
The solutions are technical, but also political
The industry is looking for alternatives, although many are patches. Some companies, such as Digital Realty, use non-drinking water, such as rainwater, for cooling. Others promote precision liquid cooling, where a special fluid cools the chips directly, reducing or eliminating the need for water. Moving data centers to cold countries in northern Europe also helps, but it slows down responses.
The most effective approach would be to use renewable energies such as solar or wind, which do not require water to generate electricity, unlike coal and gas plants. There is also the option of training models at night or during cold seasons, to minimize evaporation in cooling towers.
In the end, it is a problem of transparency and regulation. There is no international law that requires reporting or limiting this consumption. Meanwhile, the projected growth is alarming: it is estimated that by 2027, the AI industry could need between 4.2 and 6.6 billion cubic meters of water. That is the annual consumption of a country like Denmark.
The next time you ask ChatGPT to summarize an article for you or write you a joke, remember that behind those words there is a sip of an increasingly scarce resource. The real cost is not in your monthly subscription, but in that glass of water that, somewhere in the world, someone else will not be able to drink.
