Russia adopts AI regulations, lags behind in the race
Russia has adopted a law aimed at regulating artificial intelligence in the country, inadvertently admitting that the country has next to nothing in the area of "sovereign AI" stipulated by the law, with almost all of its major AIs based on foreign models. Despite senior officials' claims that Russia is an important participant in the global AI race, a brief reality check shows that this is far from true.
AI law passed
Earlier this month, the State Duma, Russia's lower chamber of parliament, passed the law On Supporting the Development of Artificial Intelligence Technologies in the Russian Federation.
The law introduces two types of AI in Russia: "sovereign" and "national." In accordance with the law, a sovereign neural network is one that has been developed and deployed entirely by a Russian company that controls all stages of the process and stores data on servers within the country. Meanwhile, national neural networks are permitted to use foreign components, although only those available as open-source code.
In terms of the regulatory framework, the law doesn't offer anything groundbreaking. The ministry of digital development was appointed as the sole regulator. The labelling of AI-generated content in consumer products and the media is supposed to be voluntary. Training AI models on data from publicly available sources is not considered to be copyright infringement.
Tech industry insiders unimpressed
The Russian tech industry was largely unimpressed with the law. Natalya Kasperskaya, president of the software development company InfoWatch Group and a major lobbyist for the country's IT sector, published a scathing review of the law on her Telegram account.
"The bill suffers from being overly declarative, lacking a consistent terminology framework, delegating excessive authority to the Russian Federation Government, and fundamentally ignoring security issues," she said, adding that the legislation is effectively aimed at cementing the monopoly of two AI developers, Sberbank and Yandex.
According to Kasperskaya, the law contains no safeguards regarding AI trustworthiness and security, while it "almost entirely lacks protections for citizens' rights, including the right to privacy, the right to be informed, the right to protection from autonomous decisions made by technical systems regarding their rights and legal status, and the right to opt out of AI use."
"Adopting the bill in its current form will cement our country’s lag behind others and lead to the collapse of the domestic AI industry," Kasperskaya concludes.
Big expectations
For years, top Russian officials have been speaking about the importance of fostering the AI segment, promising support for companies using AI-based systems and those developing their own AI models.
Back in 2023, the Russian government said that the use of AI would be mandatory for all companies that plan to receive any subsidies from the federal budget, although that stipulation was never enacted. Apparently, Russian officials hoped that AI could contribute to the growth of the country's economy, which has been struggling since Russia’s invasion of Ukraine in February 2022 and the ensuing Western sanctions.
But the development of AI required substantial investment, which the Russian economy, burdened by the war in Ukraine and sanctions, could hardly afford.
Still, the government and lawmakers seem to be under the illusion that Russia is indeed a major player in the global AI race.
While presenting the recently adopted bill, Sergey Boyarsky, head of the State Duma committee on information policy, said: "There is a race not only between services, but between countries. There are three of them: China, the United States of America, and the Russian Federation."
Reality check
However, a brief glance at Russian AI models shows that most of them could hardly qualify as "sovereign" under the recently adopted law, as they rely heavily on Chinese models.
Currently, Russia has several workable AI models: GigaChat, developed by the country's largest lender, state-run Sber; Alisa AI from Yandex, "the Russian Google"; T-Pro from lender T-Bank (formerly Tinkoff); Cotype from mobile phone operator Mobile TeleSystems (MTS) and BerryLM from e-commerce giant Wildberries.
Novaya Gazeta Europe recently ran tests of Russian AIs, concluding that all of the above solutions, except for Sber's GigaChat, are built on top of Qwen, a model developed by China's Alibaba.
Incidentally, Yandex admitted the use of a Chinese AI model in its documentation, saying: "We initialized our training pipeline not with random weights, but with the weights of the Qwen-2.5-32B-base model (in AI model design, weights are numerical coefficients that determine the importance of information and in which the model’s knowledge is stored). A full training cycle for our Alice AI LLM model, initialized with Qwen3-235B weights, takes an order of magnitude less time than if we had initialized it with random weights."
The tests found that T-Bank's T-Pro 2.0 model is based on Qwen3-32B and MTS' Cotype on Qwen-2.5-32B.
Sber’s GigaChat, the only technically "sovereign" Russian AI, was "inspired" by China's DeepSeek V3, as its developers admitted, according to the Novaya Gazeta report.
However, the performance of GigaChat turned out to be inferior. Novaya Gazeta's tests showed that GigaChat was unable to correctly solve rather basic tasks.
Lagging behind in capacity and investment
The main reason why Russia has been unable to develop competitive AI models is a substantial gap in computing capacity, a shortage of AI-compatible chips, and significantly lower investment.
Capital expenditures by the four US companies involved in the AI race - Microsoft, Amazon, Alphabet and Meta - for 2026 are projected at $635bn to $665bn and could still be increased to $725bn. By comparison, the combined budget of all Russian players for purchasing AI hardware is $1.5bn to $2.5bn.
Data centre capacity is also a major factor. Just one US data centre, the xAI campus outside Memphis, is already operating at 1.3 gigawatts and is rated for 2 gigawatts. The combined capacity of all Russia’s data centres ranges from 1.2 to 1.7 gigawatts.
Meanwhile, potential investors in Russian data centres would also face electricity shortages. Even if they are able to purchase hardware abroad and bring it to Russia, they would not be able to connect new data centres to the electricity grid - at least in the Moscow region, where most companies developing AI models are based. Currently, utility companies are simply refusing to connect new data centre facilities due to a lack of capacity.
Sanctions, local production delays
Finally, sanctions imposed on Russia over the war in Ukraine are blocking local companies' access to AI-compatible chips. The A100, H100, H200, and B200 chips have been under an embargo since 2022. Gray-market imports, on which companies had relied for years, have collapsed - from thousands of servers in 2024 to just dozens in 2026, according to Novaya Gazeta.
Nvidia, a major chip manufacturer, can now determine, based on network latency, which country a chip is operating in and can block those used in Russia.
When it comes to Russia's own AI-compatible chip production, the picture is grim. Baikal AI accelerators are expected to be launched around 2029-2030, and they already lag behind Nvidia's latest products.
Technically, Russia could import some AI-compatible chips from China, but Chinese companies prioritise domestic customers, leaving Russian firms waiting in line for years.
In this situation, Russia's chances of closing the gap with the leaders in the AI race appear very slim.
No comments:
Post a Comment