How a drone strike disrupted Yandex services but left AI-compute losses unclear

Reports described access problems with Yandex Cloud, Disk and Documents, as well as websites and services that use Yandex infrastructure. The reporting does not give a total number of affected users or organizations, or establish that every reported problem resulted directly from the Sasovo shutdown.

Yandex said its main consumer services were operating normally, while warning that some services might be unavailable. The company was working to restore affected services, and The Record reported that Yandex was moving workloads to other facilities. The report did not specify which workloads were moved or whether they included AI-training systems.

The Record also cited network-monitoring group NetBlocks’ estimate that Yandex connectivity was about 70% of normal. That figure describes network connectivity, not a measured share of servers or processing capacity lost. It cannot establish how much computing work stopped or whether the company’s AI systems were affected.

Sasovo was one of Yandex’s five large data centers in Russia and that the site housed tens of thousands of servers. Neither figure indicates how much equipment was damaged or unavailable after the fire. The available reports do not provide an equipment-by-equipment assessment.

What is known about Yandex’s AI computers?

Sasovo houses two of Yandex’s three supercomputers, Lyapunov and Chervonenkis. Tom’s Hardware reported that the systems are used to train AI models. Yandex had not said whether either machine was damaged, and the cited reports do not establish whether training stopped or continued elsewhere.

The distinction matters: a facility shutdown and reports of service disruption show that the incident affected operations, but do not show that the two supercomputers were damaged or that a specific amount of AI-computing capacity was lost. The reported network-connectivity estimate cannot fill that gap, and there is no published figure in the reporting for lost computing work.

The Record reported that Yandex was shifting workloads, but without details on which tasks moved, when they moved or whether the supercomputers were involved. That account indicates a response to the outage; it does not show that all operations at Sasovo could be transferred or that AI training resumed at another location.

Sarkis Darbinyan, a digital-rights lawyer and RKS.global expert, told Meduza that a strike on one regional data center would not necessarily take a major online service completely offline, citing distributed infrastructure and the rerouting of traffic. That is an assessment of how services may withstand a regional outage, not confirmation of the status of Yandex’s equipment or a list of workloads that remained available.

Meduza also reported that some files uploaded to Yandex Cloud may be stored across multiple data centers, while customers using virtual machines are responsible for backing up their own data. Those arrangements describe possible protections for customer files; they do not establish whether particular services, computing tasks or AI training could continue after the Sasovo shutdown.

Why is the strike’s target still unclear?

The data center occupies part of the grounds of the Sasta machine-tool plant. Meduza reported that Yandex’s facility has been on part of the plant’s former premises since 2014, and that local authorities described a fire at an unnamed enterprise. The reporting did not establish whether the intended target was the data center, the plant or both.

The immediate confirmed effects are narrower than some of the wider claims the incident could invite: the fire, damage to part of the infrastructure, the site shutdown and disruptions affecting some services. The condition of the two AI-training systems, the scale of any equipment loss, and the duration of service problems remained unresolved in the reports available Thursday.

Yandex was still assessing whether equipment could be restored. The reporting did not say when Sasovo would resume operations or identify which workloads, if any, had been moved from the site.