In the world of AI supercalculerculators, a trend has existed for a few years. Indeed, these calculation monsters double performance approximately once a year. According to a recent study, the largest supermarket used for artificial intelligence in 2030 could consume up to 9 GW of electricity.
Ever more massive supercomputers
In February 2025, the border supercomputer from the Oak Ridge national laboratory in the United States executed the simulation of the most complex universe ever made. This feat obviously implies immense costs but also, a significant consumption of energy. Now, these same financial and energy needs double each year For this kind of machines, as revealed by a pre-published study on the Arxiv platform on April 23, 2025.
Selon l’expert en IA Konstantin Pilz du think tank Center for Security and Emerging Technology (CSET), la performance of calculation IA superordinators doubled every nine months between 2018 and 2025. This is combined with a multiplication by two each year the cost of acquiring equipment and electricity consumption. The study authors analyzed the data concerning 500 superordinators dedicated to AI.
-Up to two million chips in 2030
In detail, it is a question of an annual increase of 1.6x in the number of chips but also, an improvement of 1.6x of performance per chip. It must be said that today, the Democratization of generative AI Make that the systems containing more than 10,000 chips have become a kind of standard, while these were rare five years ago. In addition, some tech giants now have monstrous superordinators, such as Meta 100k (Meta Ai) or Colossus (XAI). The latter has cost $ 7 billion in equipment and requires No less than 300 MW of electricity.
According to the estimate of the Konstantin Pilz team, the most efficient supercomputer in June 2030 could contain up to two million chipsat a total cost of around 200 billion dollars. Regarding the energy needs, the experts evoke an incredible Electricity consumption of 9 GW. However, let us remember that in France, 1 GW corresponds roughly to the needs of a million people.
Also, it is important to emphasize Certain inequalities in the area of supercalculators. Today, the share of the private sector with regard to the computing power worldwide is 80%, while it was only 40% in 2019. The rest, namely the computing power of the public sector is owned by the United States and China, respectively 75% and 15%. However, Cloud services somewhat distort the situation, since the physical location of supercomputer can differ from that of users.
