AFK Sistema outlined the scale of investments required to create and maintain a competitive world-class large language model. According to Evgeny Chereshnev, Senior Vice President of the corporation, as stated to TASS, the minimum costs will exceed $10 billion, and the total volume of investments in artificial intelligence and related infrastructure could reach $100 billion.
The cost per cycle grows exponentially. Today, training a model with 1.7–2 trillion parameters costs about $100 million. With an increase in complexity to 100 trillion parameters, the price soars to $10 billion per cycle, and two to four such cycles are required per year. Simultaneously, energy consumption sharply increases: one cycle of a hypothetical 100-trillion-parameter model would be comparable to the annual consumption of Estonia or Slovenia.
Chereshnev called international cooperation under the terms of an honest consortium, where technologies available to one participant are available to all, the only way out. According to him, Russia should unite with other countries that are in a similar situation of potential dependence on US and Chinese AI technologies.
In fact, it is a strategic choice: either build sovereign AI in a coalition, or fall into digital dependence on two superpowers.
Read more on the topic:
- Russian neural networks may be allowed to train on state data
- New AI in Russia trained to warn about dangerous reactions to medications and dietary supplements
- Russian neural network GigaChat successfully operates on the ISS without connection to ground servers