Russia’s State Duma has passed a landmark artificial intelligence law establishing the legal definition of AI for the first time and imposing strict requirements on ”national” and ”sovereign” AI models. The most significant rule mandates that these systems must process data within Russian borders and involve Russian legal entities. After presidential approval, the government will finalize details such as which sectors are sensitive and how data access for AI training will be regulated.

The law adopts a broad definition of AI, covering technological solutions that simulate human cognitive functions, learn independently, solve problems without pre-set algorithms, and deliver results matching or exceeding human intellectual performance. This encompasses not only the AI models themselves but also the supporting infrastructure, software, data processing, and services involved.

A new legal term introduced is ”large foundational AI model.” This refers to programs trained or fine-tuned on datasets to identify patterns, generate information, make decisions, or forecast outcomes aligned with human-set goals. Such models must have at least 1 billion parameters and solve a diverse range of tasks rather than a single specialized function.

The legislation applies separate rules for ”sovereign” and ”national” AI models, essentially extending Russia’s existing data localization mandates to these systems. Data processing must occur domestically with participation from Russian entities. At the same time, the law pledges support for developers, including access to necessary training data.

How Russia plans to regulate AI models with data localization

The law sets a legal framework but leaves many practical details open for the government to determine. Officials will decide which AI applications count as ”sensitive,” the conditions for data access, and what forms of state support developers receive. These regulations will be important in defining which AI qualifies as ”national” and when foreign infrastructure is prohibited.

This move aligns with Russia’s broader push for digital sovereignty. The country already requires personal data on Russian citizens to be stored domestically and imposes sector-specific rules for protecting information and critical infrastructure. Now, a similar logic extends to generative AI-technology that has rapidly evolved from experimental tech to mainstream tools for development, search, marketing, and government services.

The 1 billion-parameter threshold aims to separate truly universal AI models from narrower, application-specific neural networks. For context, Meta’s LLaMA family-which is banned in Russia due to Meta’s extremist designation-boasts billions to tens of billions of parameters. Meanwhile, Russian tech giants have been developing their own large-scale models for corporate and everyday use for several years.

Sberbank offers GigaChat, Yandex has introduced its YandexGPT lineup, and companies like MTS are rolling out their own generative AI solutions for business clients. These efforts show a growing domestic AI ecosystem that competes with leading global offerings.

Compared internationally, Russia is moving faster toward specific AI regulation than the United States, which lacks a unified federal AI law. Its approach bears more resemblance to the European Union’s AI Act, enacted in 2024, which uses risk classification and supplier obligations. However, Russia’s law places greater emphasis on definitions, data localization, and state control over data access and developer support.

There is a practical downside: stricter infrastructure and participant requirements raise the entry bar for smaller AI teams. Building and running large language models is already costly-industry analysts estimate that cutting-edge LLMs require dozens or hundreds of AI accelerators. The AI infrastructure market is consolidating around major cloud and telecom providers. If data access and government backing are limited or selective, the new law could reinforce existing dominant ecosystems, making innovation harder for newcomers.

Once the law is signed and secondary regulations are published, clearer categories for AI systems and sensitive areas will emerge. This will directly impact whether Russian companies can develop and deploy large models on mixed domestic and foreign infrastructure, or if they must operate entirely within the country’s borders for many projects.

* Meta-owner of the LLaMA model family-is designated an extremist organization in Russia and banned.

Russia’s AI law sets the stage for stronger government oversight and localization in a sector that’s rapidly expanding worldwide. The key will be whether the implementation balances protecting sovereignty without stifling startups and innovation. The coming months will reveal how authorities translate broad mandates into rules that could reshape the country’s AI development landscape.