While ChatGPT may appear to be an ethereal entity designed to respond to our inquiries and assist with tasks including image creation, it incurs a tangible environmental cost beyond the financial expenses associated with its premium services. The facilities that host ChatGPT and similar AI models, such as the Digital Realty Innovation Lab in Washington D.C., not only require significant physical space but also exert considerable pressure on power distribution networks. Additionally, these chatbots, including ChatGPT, consume a substantial amount of water, and this demand is on the rise.
It’s important to clarify that ChatGPT does not literally consume water; rather, the data centers that support its operations utilize water to regulate the temperature of their servers, primarily through cooling towers or water-cooled evaporative systems. Annually, this water consumption can range from approximately 100 million to 1.8 billion gallons, not accounting for indirect usage linked to electricity production and chip manufacturing for the servers.
However, estimating the specific water consumption of ChatGPT presents its own challenges. Due to the variability of user prompts, arriving at a precise figure for ChatGPT’s water use is quite complex.
That said, various research efforts in recent years have endeavored to gauge the water consumption associated with ChatGPT. One preprint study on arXiv suggests that GPT-4 could consume between 350 and 417 million gallons (1,334,991 to 1,579,680 kiloliters) of water annually, assuming 700 million queries are processed daily. This suggests a wide range of potential estimates regarding the chatbot’s water usage.
Additional Assessments of ChatGPT’s Water Consumption
An early investigation into the water consumption of ChatGPT appeared on arXiv in 2023 and was later published in the Communications of the ACM in 2025. Entitled “Making AI Less ‘Thirsty’,” the study assessed GPT-3 using Microsoft’s infrastructure and estimated that the training process alone could demand around 5.4 million gallons of water. The projected water usage for ChatGPT, when responding to user queries, was estimated at approximately 500 ml (16.9 fluid ounces) for generating 10 to 50 average-length replies.
Another relevant study, “A Water Efficiency Dataset for African Data Centers,” was released in Communications of the ACM in 2025, having initially emerged on arXiv the previous year. This research concentrated on AI chatbots’ water consumption in Africa and sought to predict the water needs of Llama-3-70B and GPT-4 when tasked with composing a 10-page report and a medium-length email of 120-200 words. Alarmingly, GPT-4 could utilize as much as 14 gallons (53 liters) for the former task and up to 0.68 gallons (2.6 liters) for the latter, whereas Llama-3-70B consumed 0.16 gallons (0.6 liters) and 4 fluid ounces (0.12 liters), respectively.
Google has reported that each interaction with its Gemini model utilizes about 0.26 ml (around five drops) of water on average. While this figure appears relatively low, the company has not disclosed the volume of daily prompts handled by Gemini; thus, the significance of the 0.26 ml figure cannot be fully assessed without additional context.
