- raw material used for chip production;
- electricity used to power all that computing power;
- water to cool the heat that millions of processors generate.
To get an idea of energy consumption, a recent report published this year by the IEA (International Energy Agency) points out that, driven by the advancement of AI, the global demand for data centers could double by 2026. In total, there are now more than 8 thousand data centers around the world, 33% of which are located in the United States, 16% in Europe and almost 10% in China.
The prospect is that AI models will become increasingly complex, which will increase the demand for energy resources. Estimated global consumption in 2022 is 460 terawatt hours (TWh), but is expected to reach more than 1000 TWh in 2026. This demand is roughly equivalent to Japan’s electricity consumption.
Although algorithms and hardware are becoming more efficient, the environmental cost of training and inferring AI models is an element that cannot be left out of the equation when we discuss the future of AI, especially because there may be other factors hidden in the discussion.
This is exactly what professor Shaolei Ben, from the University of California, draws attention to. In an article published by the OECD (Organization for Economic Co-operation and Development), he argues that water consumption is an often ignored factor in discussions about AI and the environment.
According to their research, air pollution and carbon emissions are the best-known environmental costs of AI, but it is also quite thirsty for water, consuming a significant amount of the resource in two main ways:
