Sovereignty, as a Service
Russia's first AI law crowns "sovereign" models and reserves the state for them. Most of the models it would crown run on Chinese weights, and all of them on Chinese chips. Read the Chinese and Russian press on the same law, and the two readings barely meet.
On 1 September, Russia's first law on artificial intelligence came into force. Federal Law №243-ФЗ is built around one word, суверенитет (suverenitet, "sovereignty"). It creates a protected class of "sovereign" models, reserves parts of the state for them, and from 2027 will let only them train on copyrighted Russian work. To qualify, a model must be Russian-built, Russian-hosted, and pass a state examination in "traditional spiritual and moral values."
Set the law beside the models it would most like to crown, and a gap opens. Some of Russia's leading models are, by their own makers' accounts, built on Chinese open weights. The rest train on an architecture borrowed from China, on chips Russia is buying from China because it can buy them almost nowhere else. The law does not hide this. It writes a second status whose definition is, in plain terms, a foreign open-weight model with its characteristics changed.
The facts are not much in dispute. What is strange is how differently they read in each country's press. In China, the same law is mostly a triumph, proof that a great power can answer Western sanctions by going open-source, with a strain of mockery about a neighbour that can no longer build its own models. In Russia, it is a sovereignty milestone at the official level and a quiet anxiety everywhere below it. This issue reads the law in the original, then reads the coverage on both sides, and asks what the distance between the two readings says about how these two states actually stand to one another.
Parallel Frontier is a biweekly intelligence report on China's and Russia's AI ecosystems, read in the original. Each issue works a single primary source in the language it was written in, for people who follow these systems professionally and need the source, not a recap. The name nods to 平行智能 (píngxíng zhìnéng, "parallel intelligence"), a real term in Chinese AI research.
On 26 July, President Putin signed Federal Law №243-ФЗ, "On support for the development of artificial intelligence technologies" (О поддержке развития технологий искусственного интеллекта в Российской Федерации), passed by the Duma on 8 July and approved by the Federation Council on 17 July. It came into force on 1 September, with its heaviest provisions held to 1 March 2027. It governs большие фундаментальные модели (bol'shiye fundamental'nyye modeli, "large foundation models"), any model of at least one billion parameters, and despite the encouraging title it is less a plan to build than a plan to sort. (full text, КонсультантПлюс.)
Everything turns on two labels. A суверенная модель (suverennaya, "sovereign") must be developed by a Russian legal entity that itself sets and changes the model's characteristics, be fully reproducible by that developer, and process and store its data in Russian data centres under Russian ownership. A национальная модель (natsional'naya, "national") carries the same requirements with one loosening. Third-party components are allowed if they are "distributed under an open licence" (должны распространяться на условиях открытой лицензии), so long as a Russian entity defines the model's essential characteristics. The first is meant to be Russian to the metal. The second is the permissive one, and, as the next section shows, the populated one.
Two privileges, both arriving on 1 March 2027. The first is copyright. Training on protected works is lawful only when done "exclusively for training a sovereign or national model," an opt-out regime with no compensation, available only to status-holders. The second is demand. The government may decree the cases in which only sovereign or national models may be used, including in state systems (for banking it must agree the terms with the Bank of Russia). There is no outright ban on foreign models, but there is a wall built around the state's own custom. (what the law changes, PPT.ru.)
To hold either status a model must be certified against "traditional Russian spiritual and moral values" (традиционные духовно-нравственные ценности), a procedure still being drafted. Russian coverage called it an "exam in spirituality" (экзамен по духовности). (Vedomosti, via Forbes.) What is not in the law is as telling as what is. There is no general duty to label AI content, no compute programme, no research money. For an ordinary business, the single new obligation from March 2027 is to tell users who owns the rights to a model's outputs. The statute allocates privilege and control. It does not build.
The law sorts models by how Russian they are, so it is worth asking how Russian the field actually is. The honest answer is a gradient, running from open borrowing to a single contested claim of independence.
At one end sit frank adaptations of Alibaba's Qwen. T-Bank's T-Pro and T-Lite are, in their own model cards' words, "built upon the Qwen 2.5 model family." MTS's Cotype is successive generations of the same base. (T-Tech model card, Hugging Face.) These are textbook "national" models, a foreign open base adjusted for Russian use, and their makers do not pretend otherwise.
Yandex is the awkward middle. By the company's own 2025 engineering write-up, YandexGPT 5 Pro was initialized "not with random weights, but with the weights of the Qwen-2.5-32B-base model." (Yandex, on Habr.) Its smaller sibling, YandexGPT 5 Lite, was trained from scratch. Yandex would later deny depending on Chinese models at all, a denial its own blog contradicts, and one we return to below.
Only Sberbank's GigaChat reaches convincingly for the "sovereign" bar. Sber says GigaChat 3 was trained from zero on its own data, "not the fine-tuning of a ready-made foreign model," while conceding that its design was "inspired by the architecture of DeepSeek V3." (Sber, on Habr.) Even here independence is partial. The blueprint is Chinese, and so is the silicon. Cut off from Western accelerators by export controls, Sberbank is now chasing Huawei's Ascend chips to train and serve the model. (Reuters, via Tom's Hardware.)
So the gradient runs from open Qwen fine-tunes, through a flagship initialized from Qwen that denies it, to a lone from-scratch model built on a Chinese architecture and Chinese chips. The law's two tiers sort that gradient into a near-empty "sovereign" summit and a "national" tier that legalises the Qwen fine-tuning most of the field relies on. The sovereignty the law secures was never going to live in the weights. It lives in the ownership, the hosting, and the control.
Cross the border and the picture inverts. In the Chinese press the Russian law is not an embarrassment to be managed but a validation to be welcomed, and the field's dependence is not a weakness but a result.
Chinese coverage of №243-ФЗ is close and, in the main, admiring. A widely syndicated commentary carried by Tencent and Sina calls the law "the world's first institutional practice to write 'sovereign AI models' into federal law" and "Russia's technological awakening under intense sanctions." (长安街知事, via Tencent.) The claim of a world first is the Chinese author's, not ours. What matters is the approval inside it. Russia's law confirms China's own argument, that open weights plus sovereignty are the answer to a Western technology blockade. Governance trade press (安全内参, Secrss) and Global Times covered the draft and its passage in the same key.
That argument travels under a slogan. Writing in Global Times, Tsinghua's Shi Anbin (史安斌) and Zheng En (郑恩) hold that "DeepSeek's rise offers the Global South new ways to reshape technological sovereignty," and that optimising open models for non-English data "strengthens digital-intelligence sovereignty" (数智主权, shùzhì zhǔquán) for the states that adopt them. (Shi Anbin & Zheng En, 环球时报, hosted at Peking University.) Shi is no neutral academic. He sits on the Propaganda Department's committee on international communication. The message is public diplomacy, by one of the people whose job is to craft it.
Turned toward home, the language is franker. People's Tribune, a People's Daily journal, casts open-sourcing as a strategy to "break the compute lockdown" (突破算力封锁, tūpò suànlì fēngsuǒ) and trigger a "herd effect" (羊群效应, yángqún xiàoyìng) that pulls the world's builders onto Chinese foundations. (人民论坛 / People's Tribune.) When Alibaba first opened Qwen, a finance column carried by the state-owned outlet Jiemian called it a "siphon effect" (虹吸效应, hóngxī xiàoyìng). The more models built on Qwen, the broader Alibaba Cloud's prospects, until it becomes "the weather vane" (风向标, fēngxiàngbiāo) the field follows. (子弹财经, via 界面新闻 / Jiemian.) Empowerment is the word for abroad. Capture is the word for home.
Where Russia specifically is concerned, a strain of condescension runs underneath. One widely reposted feature asks whether "the vanished great power" (消失的大国, xiāoshī de dàguó) has gone silent in the AI age, and answers that it has retreated from the map. Few can name a Russian model, its chips are generations behind, and Sberbank is reduced to chasing Huawei's accelerators. (心智观察所 / "Mind Observer," via 虎嗅网 / Huxiu.) China writes about Russia as a reference case and a customer, not a peer. That these attitudes coexist, approval and empowerment and capture and condescension, is not contradiction. It is 内外有别 (nèiwài yǒu bié, "inside and outside differ"), different messages for different audiences, all of them sanctioned.
Russia has noticed the dependence. What it engages is the fact of it, not the Chinese reading, and the reaction divides by who is speaking.
On 17 July, Sberbank's Herman Gref said aloud what the Chinese press only implies, that Yandex had stopped building its own foundation models and now merely fine-tunes Qwen, leaving Sber the last holder of a fully Russian model. Yandex rejected the charge within days, insisting it "continues to create its own fundamental models, maintains a complete development cycle, and does not depend on third-party solutions." (anti-malware.ru.) Inside Russia, "built on Qwen" is not a neutral description. It is an accusation, and the denial is proof it stings. The awkward part, noted above, is that Yandex's own engineers had already conceded it. The 2025 write-up states that YandexGPT 5 Pro was initialized "with the weights of the Qwen-2.5-32B-base model." The denial was contradicted by Yandex's own blog.
Gref has elsewhere put Russia six to eight months behind China's models, blaming compute access rather than talent. (Expert.) And the worry has reached the top. On 6 October, Prime Minister Mishustin told a digital-development forum that the leading AI powers restrict technology transfer precisely to keep others dependent, and that Russia could not accept such a fate for itself. The government, he said, would decide where critical infrastructure may use only domestic models, while for everyone else "the choice remains with them." (Rossiyskaya Gazeta.) It is the same gap the Chinese press mocks, admitted in Moscow, but recast as a hardware problem and a question of will.
The independent press, reporting from outside the country, is blunter still. Novaya Gazeta Europe headlined its survey of Russia's lag "We've got neural nets at home" (У нас есть нейросети дома), the phrase a parent uses to deny a child the brand-name thing, and observed that the "sovereign AI" slogan was turning, in practice, into a course of cooperation with BRICS, and above all with China. (Novaya Gazeta Europe.)
Put it together and the asymmetry is the point. China narrates Russia. Russia narrates itself. Moscow engages the dependence, defensively or anxiously or with a joke, but not the Chinese framing that files it under vanished powers and captured markets. As far as the open press shows, the two conversations run alongside each other and seldom meet. That is what a hierarchy looks like from the inside. And nowhere in the Russian mainstream is the conclusion to leave. It is to depend more safely.
The place the two narratives meet is a sales pitch, aimed at the same customers.
Having built the control layer at home, Russia is renting it out. Sberbank's First Deputy Chairman Alexander Vedyakhin has floated a BRICS "AI-sovereignty" platform for states that want their own AI but cannot afford to build it, naming partners across Africa, Asia and Latin America. Sber's AI Journey conference, with a BRICS+ forum attached, runs in Moscow from 25 to 27 November. (AiF.) The product is "sovereign AI," a national model a buyer can call its own.
Beijing is making the same offer, one scale up. At the BRICS summit in New Delhi in September, Xi Jinping proposed a China-led "open-source AI community" for members to jointly develop large language models, with training programmes and a shared cloud platform, and invited them into a Beijing-founded World AI Cooperation Organization. (Euronews.) This is the empowerment slogan institutionalised, open weights as the Global South's road to sovereignty.
The two offers describe one stack. Beneath a Russian "sovereign" model or a BRICS "open-source" one sits, in most cases, a Chinese base. Above it sits a control layer, national or sovereign, that certifies the model against a government's values and keeps its data at home. This is where sovereignty-as-control becomes a product. A buyer does not purchase independence. It buys the right to choose whose rules the model must obey, and a ready-made apparatus for enforcing them, on foundations it still does not own. Tellingly, BRICS declined to write Xi's open-source community into its New Delhi declaration. The members want the sovereignty. They are wary of whose foundation it would rest on.
Strip away the rhetoric and one order shows through both presses. China's narrates it as a win, Russia's as a risk to be managed, and neither quite answers the other. The sovereignty each advertises is real only at the control layer, and that layer sits on a Chinese base the stronger party is content to let the weaker keep badging as its own. The alignment is genuine and deepening. The equality is not. "Not One Bloc," this publication's standing read, sharpens here into something more exact, a bloc of unequals, each calling itself sovereign.
Sources are linked inline. The primary texts are Federal Law №243-ФЗ, «О поддержке развития технологий искусственного интеллекта в Российской Федерации» (signed 26 Jul 2026, in force 1 Sep 2026), via КонсультантПлюс and PPT.ru, with model lineages from the developers' own materials (T-Tech on Hugging Face, Yandex and Sber on Habr) and the Sberbank chip account via Reuters. The Chinese-language commentary is 长安街知事 via Tencent, Shi Anbin & Zheng En in 环球时报 / Global Times, 人民论坛 / People's Tribune, 子弹财经 via 界面新闻 / Jiemian, and 心智观察所 via 虎嗅网 / Huxiu. Russian reaction is from anti-malware.ru, Expert, Rossiyskaya Gazeta and Novaya Gazeta Europe. BRICS coverage is from AiF and Euronews.