How much water does it take to write 100 words with ChatGPT?

When you write a short email message of about 100 words using ChatGPT, the technology behind it uses about half a liter of water. This figure comes from a scholarly study published in the journal Communications of the ACM by Pengfei Li, Shaolei Ren, and colleagues from the University of California, Riverside. The authors of the study, titled “Reducing the Thirst of Artificial Intelligence,” emphasize that this figure includes both direct water consumption for cooling servers and indirect water used to generate the electricity that these systems consume.

Since most users do not just stick to one “question”, but develop a longer conversation, consumption quickly escalates. A conversation that consists of between 10 and 50 exchanges uses about half a liter of water, and with each extension of the conversation this value is multiplied proportionally. The reason for this much consumption lies in excess heat. Modern Nvidia chips, which are key to learning and operating artificial intelligence, emit between 300 and 700 watts of heat per unit. Since tens of thousands of such chips are used simultaneously when learning models, enormous amounts of heat are generated. The most common cooling methods are based on evaporation, where water flows right next to the equipment, absorbs excess heat and then evaporates into the atmosphere. In this process, as much as 80 percent of the pumped water is “lost”.

A single large center now uses more water than a town of 10,000 people for all its needs. Google used about 30.6 billion liters of water in 2024, an 8 percent increase from the previous year. Similarly, Microsoft reports steep growth. Its Iowa complex used 43.5 million liters of water during GPT-4 training in July 2022, and another 50.7 million liters in August. Meta used 3.07 billion liters of water in 2023, while Amazon does not disclose data.

According to a report by Lawrence Berkeley Laboratory, industry in the US will directly use 65.8 billion liters of water in 2023, and indirectly use another 798.7 billion liters through electricity. By 2027, global artificial intelligence is expected to require between 4.2 and 6.6 billion cubic meters of water annually, which is close to half of the UK's total annual water consumption.

This is already causing controversy. Google halted construction of a $177 million data center in Chile in 2024 because it failed to properly assess the impact on local water resources amid a 15-year drought. Similar problems are brewing in Mexico, Uruguay, Spain, and Arizona, where local residents have halted a $12.4 billion project.

Companies often hide data. There are three major reporting gaps. The difference between pumped and actual evaporated water, the omission of indirect electricity consumption (which is 12 times larger), and the concealment of data at the level of individual locations due to competitive and reputational risks. Whether artificial intelligence will bring more solutions than harm through smarter management remains an open question.