Yandex Sasovo Data Center Drone Strike: 2,700 GPUs Claimed, What Is Verified

On the night of October 7-8, 2026, drones hit Yandex's data center in Sasovo, in Russia's Ryazan region. A fire broke out and the site stopped operating. A viral version of the story says the strike destroyed 2,700 Nvidia GPUs. That number is not confirmed. This post separates what is claimed from what is verified, because the difference matters for anyone who tracks AI compute.

Here is the short version. The attack is confirmed by Yandex itself. The shutdown is confirmed. The presence of two of Yandex's three AI supercomputers at the site is confirmed by Reuters. What is not confirmed is the count of destroyed GPUs, and what is still unknown is whether any of the equipment can be recovered.

TL;DR: claimed versus verified

What happened at Sasovo?

According to CNBC, Yandex suspended operations at one of its key data centers on Thursday, October 8, after a drone strike. The attack caused a fire at the company's data center in Sasovo, Ryazan, in western Russia, and affected part of its infrastructure. Yandex told CNBC that the data center had stopped operating entirely and that some services may be unavailable. It added, "There are no casualties, and emergency services are on site dealing with the aftermath."

Reuters describes the Sasovo hub as one of Yandex's five large data centres, hosting tens of thousands of servers. Later on October 8, Yandex said in a statement carried by Reuters: "The scale of the damage to the infrastructure is still being assessed. For the moment, we cannot confirm whether it is possible to restore the data centre equipment in Sasovo."

So the official record is a fire, a suspended site and an open damage assessment. Nothing in it counts GPUs.

Where does the 2,700 GPU number come from?

The figure traces to an open-source intelligence account on X. In the feed summary that circulated, the account said the 35+ MW facility "reportedly housed nearly 2,700 Nvidia A100 GPUs by 2021." The phrase "by 2021" is the key. It is a count of hardware that was reported years ago, before Western sanctions on Russia hardened, not a post-strike damage assessment.

Aggregator headlines then compressed that into "destroys 2,700 Nvidia GPUs." That is two leaps in one: from "reportedly housed" to "housed today," and from "fire at the site" to "all destroyed." We have not found a primary or reputable source for either leap. If you saw the number on a timeline, treat it as a reported historic estimate of what the site held, not a loss figure.

A second caution applies to the "35+ MW" figure. It also came from the same OSINT source. We did not find Yandex or Reuters confirming a power rating.

What do we know about the supercomputers?

Reuters reports that two of Yandex's three supercomputers, used to develop its AI model, are housed at Sasovo, and that Yandex has not said whether they were damaged.

Tom's Hardware fills in the names and specs. It says the Sasovo site accommodates two Nvidia A100-based supercomputers named Chervonenkis and Lyapunov. Chervonenkis, the company's most powerful system, ranked 19th in the Top500 list in 2021. Its 21.53 FP64 PFLOPS would place it 101st today. Both machines have been used to train AI models including YandexGPT, Russia's leading large language model. The third system, Galushkin, is in Vladimir.

Tom's Hardware also offers a judgment: given the size of the fire, "the systems' chances of survival are low." That is a reasoned guess by the outlet, not a damage report. Yandex has not confirmed it.

Why replacing A100-class hardware is hard for Russia

The A100 is an older Nvidia data center GPU. Yandex ran its cluster years ago, and the Tom's Hardware piece notes that, because of Western sanctions, Yandex can no longer upgrade these machines. Advanced accelerators and servers face export restrictions into Russia. We have covered how those controls work and where they leak in Taiwan's B300 smuggling indictments and the 2026 export control overhaul.

That is why even a moderate loss matters more to Yandex than the same loss would to a Western cloud. A US lab that loses a cluster can order more. Yandex would have to find gray-market hardware or lean on other suppliers. Tom's Hardware suggests Yandex may be forced to use Chinese infrastructure for future model training, which is an outlet's projection and not a company statement.

What happened to Yandex services?

Reuters says some Yandex services were unavailable. Tom's Hardware reports that many services using Yandex Cloud went down and that the company recommended customers use alternative zones. The OSINT feed also flagged a power failure in the ru-central1-b zone. Yandex told CNBC that its core services were unaffected, and Tom's Hardware says Yandex claims key consumer services remain largely available.

Those statements are consistent if you read them carefully. A cloud zone can fail while search and maps keep running from other sites. Training clusters live in one place. User-facing inference is replicated across several. So the likely pattern is a bigger blow to Yandex's AI development capacity than to its consumer products. That is our reading, not a statement by Yandex.

Who did it, and what does "mirroring" mean?

Ukraine has not directly claimed the strike, according to CNBC. Speaking to reporters in Kyiv, President Volodymyr Zelenskiy did not name Yandex but said, in Reuters' words, that Ukraine was "mirroring" Russian attacks on data centres. His quote: "We always respond in mirror-like fashion. You all know, they have been hitting and continue to hit our data centres. We are responding. I can't share all the details."

The "mirror" framing has a factual base. The Record reported in early September that a Russian strike in July damaged a major data center in Kyiv, forcing several local internet providers to suspend operations. Techmeme's headline, which cites Reuters, calls the Sasovo hit the first major attack on a Russian data hub since the war started. We could not open that Reuters article text directly, so we treat the "first major" label as a wire-service characterization.

Russia was already preparing for this

The Record also reported that a decree signed by President Putin in late August lets the government temporarily take control of critical infrastructure if operators fail to protect it, including from drones. Russian media said the rules affect data centers. Russia had 181 data centers as of February, according to brokerage BCS data cited by the outlet. Much of that capacity sits around Moscow and St. Petersburg, which are increasingly exposed.

The RUVDS chief executive told Kommersant that tighter security would raise capital costs and, in time, prices for end users. That cost logic is not unique to Russia. It echoes what we see in the West, where siting and permitting rules are also lifting the cost of capacity. See our reporting on the San Francisco data center moratorium and the AI data center backlash map.

What this means for people who build with AI

Most readers do not run compute in Ryazan. The lessons still carry.

Concentration is a risk, not just a cost. Training runs and large inference fleets often sit in one site or one cloud zone. Sasovo shows that a single event can remove two of three flagship supercomputers from a company's lineup. Ask where your checkpoints live, how often you sync them, and how fast you can restart in a second region.

Availability zones are not geopolitical insurance. A cloud zone failure is a normal engineering event. A conflict that targets a whole facility is a different class. If you serve customers in a conflict-adjacent region, plan failover across countries, not just zones.

Hardware age and supply matter. Yandex ran A100-class machines because it could not buy newer ones. That is a reminder that your fleet is only as good as your ability to replace it. For local builders, the same logic shows up in the RAM price spike and in what a DGX Spark can or cannot replace.

Treat viral numbers as leads. A count that circulates with a source like "by 2021" is not a loss figure. Before you cite it, find the line in a wire report or company filing. This is the habit we teach in our workshops.

What to watch next

• Yandex's damage assessment. Whether the two supercomputers are recoverable.
• Cloud status updates. When the affected zone returns, and whether customers move zones.
• More strikes. Whether other Russian data centers, such as Yandex's other four large sites, are hit.
• Russian hardware moves. Any purchase of replacement accelerators or a shift to Chinese infrastructure.
• Ukrainian statements. Whether Kyiv claims the strike and what it says it targeted.
• Retaliation. Russian strikes on Ukrainian data centers, which The Record said had already begun in July.

What we could not confirm

We read the Reuters, CNBC and Tom's Hardware reports and The Record's September coverage. We could not read Yandex's own statements directly, so we cite them as carried by those outlets. We found no reputable source that confirms the 2,700 A100 count as a post-strike loss, the 35+ MW rating, or the claim that the site is a "critical hub" for the whole Russian internet. A Hacker News submission of the Reuters story had no comments at the time of writing. One further caution: some posts frame the strike as aimed at a machine-tool plant next door. That detail also comes from the OSINT feed, and we have not found a reputable outlet that states it as fact.

Details on damage, outages and hardware counts were accurate as of October 9, 2026, and may change as Yandex completes its assessment.

Related reading

• San Francisco data center moratorium: 45 days, up to 22 months
• The AI data center backlash, mapped
• Taiwan indicts 9 over smuggled Nvidia B300 servers
• Trump replaces Biden AI chip rules to block remote GPU access
• UK and Ukraine: battlefield data for AI training
• RAM prices and local AI builds
• Primary reporting: Reuters, CNBC, Tom's Hardware