How Much Water Does ChatGPT Use? It Depends
Primakov/Shutterstock
ChatGPT may seem like this magical, disembodied oracle that exists to answer all our questions and help us with tasks like image generation, but it has a real, physical cost beyond the money we have to pay for its higher tiers. The data centers that power ChatGPT and other AI models, such as the Digital Realty Innovation Lab in D.C., take up a ton of space and also put a heavy strain on our power grids, but that’s not all. Chatbots like ChatGPT also use a lot of water, and that use is only growing.
Now, we’re not saying that ChatGPT needs to drink, but the data centers that house ChatGPT servers use water to keep the hardware cool, primarily in the form of cooling towers or water-assisted evaporative cooling. This can consume anywhere between 100 million and 1.8 billion gallons of water annually, and that’s not even counting indirect usage, such as the water used to generate electricity and to manufacture the chips used in the servers.
That’s for data centers as a whole, though; figuring out how much water ChatGPT uses exactly is a whole other kettle of fish. Given how variable ChatGPT prompts are, there’s likely no way to definitively know how much water ChatGPT uses.
However, there have been several studies over recent years that have tried to quantify the popular chatbot’s water use. One pre-print study available on arXiv, for example, projects that GPT-4o could use anywhere between 350 and 417 million gallons (1,334,991 to 1,579,680 kiloliters) of water annually, based on 700 million queries a day. But that’s not the only figure that researchers have proposed for the ChatGPT’s water consumption.
Other estimates of ChatGPT’s water usage
Bangla press/Shutterstock
One of the earliest attempts to analyze and estimate how much water ChatGPT used was uploaded to arXiv in 2023, before being published in Communications of the ACM in 2025. The study, titled “Making AI Less ‘Thirsty’,” looked at the water usage of GPT-3 as implemented on Microsoft’s servers. The authors estimated that training GPT-3 alone could consume a total of 5.4 million gallons of water. ChatGPT’s consumer-facing water use, meanwhile, was projected to be around 500 ml (16.9 fluid ounces) for “roughly 10–50 medium-length responses.”
Another study, titled “A Water Efficiency Dataset for African Data Centers,” was also published in Communications of the ACM in 2025 (and initially posted to arXiv the year before). It focused specifically on AI chatbots’ water use in Africa and tried to approximate the water consumption of Llama-3-70B and GPT-4 when tasked with writing a 10-page report and “a medium-length email of 120-200 words.” Somewhat disturbingly, GPT-4’s water use for the former task could be as high as 14 gallons (53 liters), while the latter maxed out at 0.68 gallons (2.6 liters). Llama-3-70B, in comparison, used 0.16 gallons (0.6 liters) and 4 fluid ounces (0.12 liters), respectively.
Google, for its part, has revealed that each Gemini prompt uses 0.26 ml (roughly five drops) of water on average, which doesn’t seem too bad. However, it did not disclose how many prompts Gemini handled daily; without that information, the 0.26 ml number doesn’t really allow us to get a full picture of Gemini’s water use.
