The key facts

  • On October 8–14, Yandex data centers in the Ryazan, Kaluga and Vladimir regions and Sber's center in Balakovo were attacked; all three Yandex supercomputers were put out of action.
  • T-invariant experts: restoration may take from six months (soot cleaning) to one or two years if NVIDIA chips need replacing — chips that are under sanctions and in global shortage.
  • The attacks followed Russian strikes on Ukrainian data centers (September 23–28, at least six facilities); earlier, Russian authorities had discussed moving computing capacity out of Moscow and St. Petersburg.

A series of attacks on data centers

From October 8 to 14, 2026, several major Yandex and Sber data centers in Russia were hit by drone strikes, leading to disruptions in Yandex's cloud services. According to a number of outlets, including T-invariant, the attack was a retaliatory strike by the Ukrainian armed forces: prior to that, the Russian army had been striking Ukrainian data processing centers. The first strike occurred on the night of October 8 in the town of Sasovo, Ryazan region, where the supercomputers Chervonenkis (the most powerful among those publicly presented in Russia) and Lyapunov (third in performance) are located. The following night, Yandex's largest data center in Kaluga was attacked; on the night of October 11 — the center in Vladimir, home to the Galushkin supercomputer; and on the night of October 14 — Sber's large data center in Balakovo, Saratov region. After the attacks, all three Yandex supercomputers were out of service for an indefinite period.

Why the equipment is vulnerable

Four experts interviewed by T-invariant are confident of serious damage to computing capacity. The main reason for the vulnerability is direct free cooling (air cooling): even if fire did not touch the racks, air carrying soot, black carbon and fine particulates could have damaged the machines' internals. In the summer of 2026, Yandex announced the introduction of liquid cooling, which is more effective in case of fire, but, in the experts' opinion, it is unlikely to have managed to refit all its centers. One of those interviewed, who works at a global AI giant, spoke skeptically about Yandex's cooling technologies: 'Companies at the level of industry leaders abandoned free cooling long ago.' Cleaning the supercomputers of soot may take up to six months; if component replacement is required, including scarce NVIDIA chips or their analogues, full restoration could stretch to one or two years.

Risks for business and AI

The main problem for Yandex, according to those interviewed, is the partial or complete loss of computing capacity. After the first strikes on October 8–9, most investors maintained a moderately positive outlook — cloud technologies account for a small share of revenue and the load can be redistributed — however, the shares immediately fell by several percent, and the third strike intensified negative financial forecasts. Yandex does not disclose AI figures separately, but the 'Search Services and AI' segment, as observed by T-invariant, has the highest adjusted EBITDA margin among disclosed segments: 46.3% in the second quarter of 2026 and 44.8% for the first half. According to the company, nearly half of advertising revenue growth in the second quarter was driven by the ARGUS recommendation technology. Two supercomputers in two data centers served the development of neural networks and the AI ecosystem; if the capacity stands idle, the development of YandexGPT and Alice will be suspended indefinitely. Experts cite two reasons for the difficulty of restoration: sanctions restrictions on NVIDIA A100 deliveries to Russia and the global shortage of advanced chips — contracts for their supply are concluded two to three years in advance, which, according to the outlet, was confirmed by two independent sources.

Data centers not ready for war

According to T-invariant experts, Yandex built its centers and supercomputers 'in peacetime and for peacetime,' focusing on energy efficiency, power supply redundancy, simplicity and cheapness of cooling, and load redistribution. A source for the BBC Russian Service reported that Yandex regularly practiced shutting down one data center with load transfer, but a sequential failure of several large sites was probably not considered; two separate buildings in Vladimir may have been regarded as mutual backup. An expert in supercomputing technologies noted that in Russia, centers are built as simple 'boxes' without reinforced walls or sophisticated fire suppression systems, and 70–80% of all centers are concentrated around Moscow and St. Petersburg. Meduza, citing data from the company HAKA Moscow, which built Yandex's Vladimir center in 2016–2017, writes that the priorities were construction speed, scalability and reducing the need for active cooling; such metal structures are not resistant to drone strikes. The outlet also notes an error in location choice: air defense resources were concentrated around Moscow, but UAVs sometimes break through beyond the city, as in the case of the Moscow refinery. As a contrast, the expert cites the German center in Leibniz near Munich — a windowless cube surrounded by moats, with a 'tower within a tower' in the form of a Faraday cage — which ranks below Chervonenkis in the top-500 ranking but is protected by serious investment; underground centers in Norway and Finland and Alibaba's supercomputer in China are also mentioned.

Warnings beforehand and 'digital detox'

T-invariant writes that the first public calls to improve data center security came just days before the attacks. On October 6, Russian Deputy Minister of Digital Development Yevgeny Filatov, at the 'Digital Solutions' forum, raised the issue of moving computing capacity out of Moscow, St. Petersburg and the Moscow region to places 'not yet reachable by strikes.' On the evening of October 7, hours before the first strike, RBC reported that the IT business and the Ministry of Digital Development were discussing placing AI data centers in 'friendly countries.' By that moment, the BBC recalls, Russian troops had hit at least six Ukrainian data centers from September 23 to 28, leaving people without internet and interrupting TV broadcasts; the Russian Defense Ministry published on social media an image with the caption 'Digital detox… by forced order.' On October 4, Volodymyr Zelensky warned of retaliatory strikes: 'Do they really think we will not respond?' A T-invariant expert believes that restoring supercomputers of this level now and in the near future is impossible — it is not about money, but about the shortage of chips and components, global queues for them, and the high cost of grey and black imports. In November, the next update of the top-500 ranking will be released: Chervonenkis currently ranks 101st, Kazakhstan's Alem system — 104th (Kazakhstan's second machine, at 122nd place, is above all other Russian supercomputers). The outlet's experts have no doubt that the 'forced digital detox' this year will push Russia below Kazakhstan in the ranking, and competition with the US, China and Japan is already out of the question.