Yandex announced a lack of electricity to teach neural networks

“Not only are the computer capabilities not enough for us, but if you now want to build a large data center, it will not only have a year to sprinkle it with capabilities, but they are not. It will not be assigned as much as necessary, ”said Kraynov.

Then he answered the question of whether it is worth the companies to train their own AI models or use the existing ones. According to the expert, from the point of view of the economy and “humanity as a whole”, it is incorrect to constantly create and train new neural networks: it is better to work with those already released, “complete them in their data.”

Why is there not enough energy? The data centers for the teaching of AI really consume much more energy than ordinary data centers, confirmed by the director of the Selecttel Data Center, Ilya Mikhailov. According to him, this is due to the need to use more modern and powerful video cards.

There are no problems with the production of electricity, but there are difficulties with its delivery and distribution, he added. A punctual deficit can occur in Moscow and St. Petersburg, as well as in the nearby territories, but in the regions, on the contrary, a surplus, the expert said.

• As of April 2024, the Russian dates centers used approximately 1% of the Power of the Unified Energy System of Russia, said Maxim Khudalov, the main strategist of the vector capital investment company. The US experience shows that the growth of consumption in this segment can be 30-40% per year, he added.