Yandex: a warning that AI can be bombed too

On the night of October 7–8, several Ukrainian drones hit Yandex’s data center in Sasovo, southeast of Moscow. Images circulated that day showed a column of smoke above the building. Yandex acknowledged it the following day: Sasovo’s operations were “completely halted.” That same morning, part of its Kaluga center went down. On Sunday the 11th, its Vladimir center stopped operating. [1][2]

Yandex is “Russia’s Google.” Search, maps, taxis, payments, e-commerce, cloud services and a voice assistant. For days, millions of Russians could not order a taxi, collect an order or pay with their phones. Some arbitration courts suspended hearings because they lost access to their databases. [3]

Then came a Sberbank artificial intelligence center in Balakovo, on October 14, and another serving the social network VKontakte in the St. Petersburg region, on the 18th. [3]

Nobody knows how much damage there is

Yandex has said little. Its data centers are, in its own words, the company’s “iron heart,” and on October 9 it admitted it was “very difficult” to say when the equipment would be running again. [1] Industry sources quoted in the press spoke of up to six months to replace what had been destroyed. [5]

On October 26, the company gave Moscow some good news: its three supercomputers — Chervonenkis, Lyapunov and Galushkin — had survived intact and were being installed at other sites. [4] Two of them were in Sasovo, the hardest-hit center. Chervonenkis ranked as the world’s nineteenth most powerful supercomputer in 2021.

That is the company’s account. Nobody has independently verified it. Satellite images show serious damage, and fires are unforgiving to electronics. “Soot is bad for equipment,” said Mikhail Klimarev of the Internet Protection Society. [3]

The honest conclusion is uncomfortable: the true extent of the damage is unknown. How many servers were lost, how many model training runs were interrupted, and how long Yandex will take to recover the capacity it had on October 7 are all unknown.

The problem is Nvidia’s chips

A building can be rebuilt. Replacing the processors inside it is harder.

Yandex’s three supercomputers run on Nvidia A100 accelerators. These are US-designed chips that make it possible to train language models like those behind AI assistants. Sanctions prohibit selling them to Russia. What Russia has, it obtained before sanctions or through indirect, expensive and slow channels.

That is why experts are looking at the chips rather than the concrete. Connor Martin of the Council on Foreign Relations warns that Russia will struggle to replace the A100s quickly. Klimarev goes further: with the global shortage of advanced chips and the sanctions, he doubts repairs can be completed “in half a year, a year, or even at all.” [3]

If Yandex is telling the truth and the supercomputers were saved, Russia is not starting from scratch. But the data centers housed more than those three machines. There were thousands of other servers. Every accelerator that may have burned among them is difficult to replace.

Digital sovereignty takes a hit

Russia has spent years building a “sovereign internet”: data kept inside the country, its own services, minimal dependence on the outside world. Yandex was the centerpiece of that strategy in artificial intelligence. Its models are Russia’s alternative to ChatGPT or Gemini.

The attacks have tested that strategy at its weakest point. Concentrating everything in a few centers makes political control easier, but also turns those centers into targets. “Because all the infrastructure and all the data is located inside Russia, it’s relatively easier to take down,” said Nick Reiners of Eurasia Group. [3]

The Kremlin’s response confirms this. On November 2, Putin signed a decree requiring private data centers to reserve 10% of their capacity for the state, available within 24 hours. The same decree allows the temporary use of data centers abroad, with authorization from the FSB, Russia’s security service. [6] The country of the sovereign internet is now considering hosting its infrastructure abroad.

The attacks may have done more damage to the strategy of AI sovereignty than to the machines. Without new chips, under sanctions and with the remaining capacity spread out and reserved, each month of delay weighs more heavily. Russia was already years behind the United States and China. Now the gap may grow.

The cost of defense

There is a second cost, less visible and more lasting: protection.

An industry working group linked to Russia’s Digital Development Ministry has approved recommendations for protecting data centers against drones. Strategic centers are advised to go underground or into bunkers. Existing centers should become less visible from the air and add reinforcement, electronic warfare and counter-drone systems. Experts quoted by Kommersant estimate that service prices will rise by around 10%. [5]

Then there is redundancy: duplicating facilities, spreading supercomputers across different sites, keeping spare capacity in case another center goes down. Yandex was already running drills for the loss of one of its six data centers. It lost three in four days. [3]

Every ruble spent on concrete, air defense and spare servers is a ruble not spent on training models. That is the damage satellite images do not show.

The lesson goes beyond Russia

Russia began by attacking data centers in Ukraine. In September, its attacks in the Kyiv region left around 100,000 households without internet access. [2] Ukraine responded by striking Yandex. Data centers are already military targets, on both sides.

For years, the AI race has been measured in chips, talent, energy and models. Since October, one more question has to be added: would the infrastructure survive a physical attack?

A data center is a huge building full of servers, cooling systems and electrical connections. A drone carrying a small explosive payload can put it out of service for months. [3] The same applies to the substations that power it and the cables that connect it.

Europe is also building large data centers for AI, including in Spain and Norway. The question is not whether someone will want to attack them, but whether they are prepared to keep operating when one fails. Having the most powerful technology is not enough. We need to build an architecture that can survive the loss of its main components.