Researchers at the University of California have shown that five searches on ChatGPT can use up to half a liter of water. They come to this conclusion after a study by the Microsoft data center in Iowa, which largely specializes in GPT-4.

Researchers believe that the immense growth of AI models could have harmful effects on the environment. The Associated Press was able to view the report, in which researchers found that just five ChatGPT commands can use up to half a liter of water. This number also includes other factors such as indirect water consumption. It has long been known that data centers are large consumers of energy and water. At the end of June, we spoke to two top Google executives about the (water) sustainability of the data center in Saint-Ghislain. The tech giant’s Belgian data center uses 1.5 billion liters of water every year.

Environmental impact of AI

The aim of the study is to uncover the environmental impact of successful products such as ChatGPT. In this context, the researchers chose Microsoft’s data center in Des Moines, Iowa, dedicated to GPT-4. This is the fourth generation OpenAI model and the driving force behind various AI applications. Figures show data center water consumption increased by 34 percent to almost 1.7 billion liters between 2021 and 2022. This means it uses more than, for example, the Google data center in Saint-Ghislain.

According to the researchers, AI is the main culprit. The study includes direct (cooling of servers) and indirect water factors (consumption of power plants that supply energy). However, there is no report of a relationship between water and energy consumption and the complexity of the job. Microsoft admitted to AP that “training large models can be energy and water intensive” but that it was aiming for greater training efficiency. The tech giant also highlighted its efforts in renewable energy and sustainability in general.

Location is crucial

According to the researchers, the location of the data center in Iowa ensures that the situation is “not too bad.” So it could be much worse if the data center were in a warmer area. For the most part, the weather in Iowa is cool enough that ventilation is sufficient to cool servers. Only at 29 degrees does the system begin to draw water from the network to provide additional cooling. Logically, water consumption is highest in the summer months. Figures from last year showed the data center pumped up as much as 11.5 million liters of water over the summer months.

The study also takes a look at GPT’s biggest AI competitor, Google Bard. The researchers state that Google used thirty percent more water across all US data centers over the same period (2021-2022). Here too, researchers attribute the increase largely to the AI ​​explosion. What’s striking here is that Google’s facilities in Iowa were the thirstiest of all.