How AI is bringing the power grid to its knees

This is the hidden and underestimated face of generative AI: it is extremely energy-intensive. The operation of ChatGPT and all its rivals does not come from the sky, and according to a study by researchers from the company Hugging Face cited by the BBCthese technologies could consume around 33 times more energy than machines using task-specific software.

In fact, when working on these tools, most of the energy consumption takes place within the giant data center. According to estimates from the International Energy Agency (IEA), these data centers will use nearly 1,000 terawatt hours per year globally, or the electricity consumption of Japan.

Is AI sustainable in the long term?

In some countries, the problem already arises. Like Ireland where almost a fifth of electricity is consumed by data centers. A figure that is expected to increase over the coming years.

Aware that this is not sustainable in the long term, the world of Tech is mobilizing. Progress will notably come on the chips used to train AI. Nvidia, the number 1 in the sector, explains that it took 8,000 chips and a 15 megawatt power supply to train an AI in 90 days. Thanks to Blackwell, its latest addition, only 2000 are needed, and a consumption of 4 megawatts.

At the same time, it is very likely that data centers will soon be built with their own energy production facilities, particularly based on solar. Enough to relieve national electricity networks.

Quoted by our colleagues, Dale Sartor, consultant and affiliated with the Lawrence Berkeley National Laboratory in the United States, ventures a prognosis for the future of generative AI: “If the old way is cheaper and easier, there won’t be many markets for the new way”.

This problem is in any case more and more discussed. A few weeks ago, we mentioned the work of Alex de Vries, employee of the central bank of the Netherlands, who created the Digiconomist site in order to measure and alert the general public on the energy impact of new technologies .

He was particularly interested in Google’s chatbot, and he estimates that if it were used during every web search, its electricity consumption would be equivalent to 29 billion kilowatt hours per year, the consumption of countries such as Kenya, Guatemala and Croatia over the same period.

📍 To not miss any news from Presse-citron, follow us on Google News And WhatsApp.