Russia's AI Ambitions Under Fire

Independent outlet T-invariant asked leading Russian supercomputing experts to assess how Ukrainian drone strikes on Yandex data centers had affected the company's artificial intelligence operations. T-invariant concluded the damage could set Russia's AI development back years and push it behind Kazakhstan in the global supercomputing race.

All three Yandex supercomputers were reportedly offline indefinitely after a series of attacks. The first strike came on October 8 in Sasovo, Ryazan Oblast, home to Chervonenkis, Russia's most powerful publicly ranked supercomputer, and Lyapunov, its third most powerful. The following night, Ukrainian forces struck Yandex's largest data center, in Kaluga. On October 11 they hit a facility in Vladimir housing the Galushkin supercomputer.

All four specialists interviewed by T-invariant said they believed Yandex's computing infrastructure had suffered extensive damage. The main vulnerability, they said, was the company's direct free-cooling system, which pulls outside air in to cool the servers. Even if flames never reached the equipment, smoke and soot could have contaminated circuit boards. Carbon particles conduct electricity and can bridge tightly packed connections, causing widespread short circuits and motherboard failures.

The experts estimated that cleaning the systems could take six months. Replacing damaged components, particularly scarce and expensive Nvidia chips or their equivalents, could stretch recovery to one or two years.

The disruption threatens YandexGPT and the Alice voice assistant just as demand was surging. According to Yandex B2B Tech figures cited by financial news outlet Finam, token consumption at Yandex AI Studio rose eighteenfold in six months. As of the date of this post, Yandex's service-status page indicated that the platform was unavailable.

The experts argued that the facilities were designed for peacetime, not war. "Everything was built for normal civilian life. Nobody imagined drones hitting the roof," one supercomputing expert told T-invariant, noting that natural disasters are rare in the regions where the facilities operate.

A source told the BBC that Yandex regularly practiced shifting workloads between data centers after individual outages. Simultaneous failures across several major sites apparently were not part of the plan.

Russia's remaining supercomputing infrastructure may also be vulnerable. Sber's Christofari and Christofari Neo systems are housed in Skolkovo, near Moscow, where air defenses are stronger. MTS operates its GROM supercomputer in Lytkarino, while Yandex maintains additional facilities in Mytishchi and Ivanteyevka. All three Moscow Oblast cities have faced recurring Ukrainian drone attacks.

The consequences could soon show up in the Top500 list of the world's fastest supercomputers. On the June list, Chervonenkis ranked 101st, narrowly ahead of Kazakhstan's Alem at No. 104. Another Kazakh system ranked 122nd, above every other Russian machine. T-invariant's experts expected the November list to place Kazakhstan ahead of Russia in top-end supercomputing performance.

Russia already trailed the United States, China, and Japan by roughly two orders of magnitude. The gap looks worse next to Nebius, the Netherlands-based company led by former Yandex chief Arkady Volozh. Its two supercomputers, ISEG and ISEG2, have a combined peak performance of approximately 425 petaflops. ISEG ranked 12th globally when first listed.

According to the Russian government's mathematical sciences development strategy through 2036, Russia's five Top500 systems collectively deliver less than 100 petaflops of peak computing power. In other words, Volozh's one company commands more peak supercomputing power than all of Russia's Top500-listed systems combined.