TL;DR
Russia's first law governing large AI foundation models took effect September 1, 2026, trading state support and market protection for Russian ownership, data localization, and ideological conformity.
What happened
- Signed July 26, 2026 by President Putin after a broader draft was scrapped in June following heavy criticism from lawyers, businesses, and government bodies.
- The final law, titled "On Support for the Development of AI Technologies in the Russian Federation," narrowed its focus to foundation models with at least 1 billion parameters.
- Two official statuses, "sovereign" and "national," are available to qualifying developers; both require Russian legal entity status, Russian-owned data centers, and conformity with "traditional Russian spiritual and moral values."
- The Ministry of Digital Development, FSB, and FSTEC are drafting the assessment procedure; developers would disclose model architecture and filtering mechanisms to accredited labs.
- The law also covers foreign developers serving Russian users, potentially pulling non-Russian companies into its compliance scope.
Why it matters
- Russia's four dominant foundation model developers, Sber (GigaChat), Yandex (YandexGPT 5 Pro), T-Bank (T-Pro, T-Lite), and MTS AI (Cotype), are already state-linked, making the law less a constraint than a formalization of existing control.
- Hardware dependency is acute: Western export controls block advanced accelerators; Russian models run on Chinese chips, smuggled GPUs, and pre-sanction Western hardware. YandexGPT 5 Pro was built on Qwen-2.5-32B weights; T-Bank and MTS AI models are built on successive Qwen generations.
- The 1-billion-parameter threshold is crude compared to the EU AI Act's compute-based criteria (10^23 FLOPs for GPAI, 10^25 FLOPs for systemic-risk models), meaning the law captures small open models and frontier models alike with no meaningful risk differentiation.
- The "spiritual and moral values" conformity test, tied to a presidential decree, gives the FSB a structural role in certifying which AI systems can operate at scale in Russia, extending decades of digital-space control into the foundation-model layer.
- Russian models are absent from international leaderboards and rank outside the top ten even on domestic benchmarks like MERA and LLM Arena, signaling the law protects a strategically important but globally uncompetitive sector.
What to watch next
- Government resolution on assessment procedures: the values-conformity and architecture-disclosure rules are not yet published; their specifics will determine how invasive the law becomes in practice.
- Foreign developer response: whether global AI companies serving Russian users comply, geo-block, or ignore the law will test its extraterritorial reach.
- Hardware trajectory: Russia's own chip designs are described as years away; any acceleration there, or tightening of Chinese accelerator supply, would materially shift the competitive picture.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.