On 16 July, Natalya Kaspersky told RTVI that Russia can no longer catch up with the United States and China in artificial intelligence. She compared the country to someone walking along railway tracks used by trains: there is, she said, "no chance of catching that train." The remark came a week after the State Duma passed a law on supporting AI. FACTUM's assessment: the InfoWatch president is offering a judgement, not a fact. But the figures that can actually be counted point to something less comfortable than "falling behind."
The bottom line
Arguments about Russian AI are usually fought on ground where nobody can win: whose neural networks are smarter. Each side produces its own tests, each is right on its own terms, and the conversation collapses into a matter of belief.
Some things are not settled by belief: how many machines for training neural networks physically stand in the country, how much money goes into them, and what is happening to supply. Those numbers are public, and they describe a picture in which the word "lag" turns out to be too gentle.
A lag is when you move in the same direction, only slower. Russia's capacity is not lagging. It is moving the other way: global computing power is growing while Russian supply collapses. The pedestrian in Kaspersky's metaphor is at least walking forward.
How much hardware Russia has
Neural networks are trained on specialised graphics cards, known as accelerators. Their number is the ceiling on what a country can do.
Nobody knows precisely how many are in Russia, and that is the first diagnosis. ServerICT, a technology distributor, put the figure at roughly 15,000 cards measured in equivalents of NVIDIA's top model, drawing on data from TAdviser and the T1 holding — and called it an upper bound. The independent analyst Alexei Boyko gives a number several times lower. A four- to fivefold spread in estimates of the same stock is not a disagreement among experts; it is evidence that nobody is counting.
Take the generous version. FACTUM's calculation: if every one of those cards were gathered into a single hall and run flat out for three months, the total work would amount to roughly one tenth of what went into training the largest known neural network, Elon Musk's Grok 4 (training-scale data from the research institute Epoch AI, February 2026). One tenth. Of one model. From last year.
And that is deliberately generous: there is no single hall, the cards are scattered across competing companies, and much of the stock is running finished services rather than training anything.
The most powerful machine Sber has confirmed publicly is the Christofari Neo supercomputer: more than 700 cards, launched in 2021. The company has published nothing newer about its own stock. For comparison, Musk's cluster in Memphis runs on the order of 100,000 to 200,000 cards.
What is happening to supply
The point here is not the stock. It is the direction of the curve.
After 2022, accelerators were brought in around the sanctions. Bloomberg, working from customs records, traced one such channel: more than a thousand servers carrying NVIDIA chips entered Russia in 2024 through a single Indian pharmaceutical company. There were many channels, and the count ran into thousands of servers a year.
Then, according to ServerICT: hundreds in 2025, dozens in 2026. Two reasons. From late 2024 NVIDIA imposed strict buyer vetting and serial-number tracking. And China, itself under restrictions, is buying up everything available on the grey market at a two- to threefold premium — Russian buyers simply cannot outbid it.
Meanwhile the global race is accelerating. By Epoch AI's calculations, the computing power poured into training frontier models is growing fivefold a year. One curve multiplies by five annually. The other falls from thousands to dozens.
The severing runs the other way too: in July, Russian apps disappeared from Google Play after App Store. The difference is that a messenger can be installed from another store. There is no other store for a graphics card.
Money: a different order of magnitude
Russia's entire market for AI accelerators in 2025 came to 62.7 billion rubles, or roughly $700 million (estimate by the T1 holding). ServerICT puts the combined 2026 hardware budget of every Russian player — Sber, Yandex, T-Technologies, MTS and VK — at $1.5–2.5 billion.
A single cluster, the one used to train Musk's previous model, cost around $4 billion (Epoch AI estimate). Everything Russia will spend on AI hardware in a year, in other words, is a third to a half of the price of one hall built for one neural network.
An independent yardstick points the same way. According to the annual report of Stanford's HAI institute, published on 13 April 2026, private investment in AI in 2025 reached $285.9 billion in the United States and $12.4 billion in China. Russia does not appear in that comparison at all. In March, Sber and Yandex asked the state for 400–450 billion rubles a year — about two percent of American private investment at current rates — and the money has not been allocated. The federal budget is already running over plan by more than a trillion rubles as it is.
What the law actually does
The State Duma passed the law on supporting AI technologies on 8 July. Seven days elapsed between introduction and passage — a timeline Kaspersky herself flagged as without precedent in the industry.
The law creates two categories, a "sovereign" model (entirely Russian) and a "national" one (foreign components permitted), and applies only to large neural networks. It contains no restrictions on dangerous applications of AI. Labelling of generated content is voluntary.
The law regulates not the building of models but their classification: who counts as ours. Kaspersky, together with Igor Ashmanov, described where this probably leads — a Russian legal entity will develop a model with a Russian name, and the entire substance underneath will be foreign.
The impression is that this is the content of current policy. The race is not being funded; statuses are being handed out in it. The mechanism is familiar: when changing reality is expensive, you change the words that describe it.
Russian models, honestly
The claim that Russia has "a model and a half" is rhetoric, and testing it takes more than press releases.
Sber has its flagship, GigaChat, released in open access; Yandex has YandexGPT. These are working products used by millions. But everything published about their quality comes from the developers themselves, and it repays reading literally.
Sber's own text on its flagship says the model "confidently beats" one Chinese network and "almost catches" another — where the second has its reasoning mode deliberately switched off, is three times smaller and was released a year earlier. The judge in that comparison is an American model. How much computing went into training GigaChat, Sber does not disclose for any version.
Russian models do not appear in the independent international counts — Epoch AI, the Stanford report. That does not mean they are bad. It means there is no way to verify claims about them from outside.
What could disprove this assessment
Efficiency may outrun scale. By Epoch AI's own data, models learn roughly three times more efficiently on the same hardware each year, and the cost of running them halves every two months. If that curve outpaces the race for size, the barrier to entry collapses — and fifteen thousand cards stop being a death sentence. Sber's latest release, as it happens, is half the size of its predecessor.
Perhaps the chase is the wrong goal. Kaspersky says as much herself: she is not convinced the country's focus should be on large models at all. Applied AI is a different race with different hardware demands, and Russia's position there is fundamentally different.
Open weights devalue the lead. As long as China publishes its models under permissive licences, a "domestic" network means someone else's base plus fine-tuning. Cheap and fast. The paradox is that this is precisely the outcome Kaspersky and Ashmanov fear: in that scenario the law works, just not the way it is advertised.
Energy. At Sber's conference in November 2025, Putin pointed to nuclear power as the foundation for expanding computing capacity. Electricity is indeed becoming a bottleneck in the United States. But Russia's constraint today is chips, not kilowatts, and reactors do not solve that.
Conclusions
Established. Estimates of Russia's accelerator stock differ by a factor of several, and the highest of them yields a fraction of one frontier training run. Supply has fallen from thousands of servers a year to dozens. The whole industry's annual hardware budget is smaller than the cost of one foreign cluster. Global spending on training is growing fivefold a year. The law was passed in seven days and regulates statuses rather than development.
Still a hypothesis. That the race is definitively lost. The technology is young, the efficiency curve is steep, and "no chance" is one person's judgement rather than a measurement. Kaspersky is entitled to her forecast; we record that it is a forecast.
If current trends hold, the question will stop being a technical one. It is already closer to a question of what the state considers achievable for itself — and what it is prepared to call success instead.
Sources. Epoch AI (Trends in AI dashboard, updated 5 February 2026); Stanford HAI, AI Index Report 2026 (13 April 2026); ComNews (ServerICT, T1 and TAdviser estimates, May 2026); Bloomberg (customs-records investigation, October 2024); GARANT (text of bill No. 1271570-8); RTVI; The Moscow Times; Vedomosti; Kommersant; RBC; publications by Sber and SberCloud.
Caveats. The calculation of Russia's aggregate computing capacity was performed by FACTUM on the basis of published ServerICT estimates and is deliberately generous: it assumes every card is pooled into a single cluster and devoted solely to training. Estimates of the number of accelerators in Russia differ by a factor of four to five, and no unified counting methodology exists. Data on the quality of Russian models is published by their developers and cannot be independently verified. Several figures on foreign clusters are estimates by research institutes rather than company reporting. Kaspersky's remarks are translated from Russian.
Current as of 16 July 2026.