An October 8 drone strike shut down Yandex's data center in Sasovo, Russia. The company has identified two supercomputers at the site, together containing 2,688 NVIDIA A100 GPUs. Their condition remains undisclosed. We examine the potential cost of replacing the computing systems and what a prolonged outage could mean for Russian AI development.
In the early hours of October 8, 2026, drones struck a Yandex data center in Sasovo, a town in Russia's Ryazan region. A fire broke out at the facility. The company said the site had been taken completely offline and warned that some services might be disrupted. It reported no injuries. RBC, Vedomosti and Reuters independently reported the incident.
Later that day, Yandex said it was still assessing the damage and could not yet establish whether the equipment could be restored. Kommersant and Reuters carried the statement. According to Yandex's previously published information, the Sasovo site houses two supercomputers, Chervonenkis and Lyapunov. The company has not disclosed whether either system was physically damaged.
What happened at the data center
According to Kommersant's timeline, Yandex Cloud Alerts reported a power disruption in the ru-central-b availability zone at 1:44 a.m. Moscow time. Yandex subsequently confirmed a fire affecting part of the infrastructure and said the Sasovo data center had been fully shut down. Technical teams continued working at the facility throughout the day.
Ryazan regional governor Pavel Malkov reported that two drones had been shot down over the region and that the roof of a business facility in Sasovo district had caught fire. The number of drones that actually struck the data center's equipment has not been publicly established. Andriy Kovalenko, the head of Ukraine's Center for Countering Disinformation, described the attack as retaliatory, according to AFP.
The disruption affected more than Yandex's own products. Businesses relying on Russian cloud infrastructure also experienced problems. During the day, reports emerged of technical failures at a range of websites and online services. Kommersant cited difficulties purchasing bus tickets online in the Moscow region. Yandex, however, told Reuters that its main consumer-facing services were operating normally. These service disruptions should not be confused with evidence of equipment loss: an unavailable cloud zone does not, by itself, mean its servers have been destroyed.
FACTUM reported earlier on the strike and fire in Russian. This analysis addresses a different issue: the computing capacity potentially at risk and the cost of rebuilding it.
The two supercomputers based in Sasovo
On November 15, 2021, Yandex announced that two of its three supercomputers, Chervonenkis and Lyapunov, were located in Sasovo, while the third, Galushkin, was in Vladimir. The systems were built for machine learning and large-scale data processing. On October 8, 2026, Reuters reported that Yandex still associated two of those three systems with Sasovo, but declined to say whether they had been damaged.
Yandex's published technical specifications provide a breakdown of the hardware:
Chervonenkis: 199 compute nodes, 1,592 NVIDIA A100 GPUs with 80 GB of memory each, 199 TB of system RAM, 21.53 petaflops of Linpack performance and power consumption of 583 kW.
Lyapunov: 137 compute nodes, 1,096 NVIDIA A100 GPUs with 40 GB of memory each, 68.5 TB of system RAM, 12.81 petaflops of Linpack performance and power consumption of 323 kW.
Combined, the two systems account for 336 compute nodes, 2,688 GPUs and 34.34 petaflops of Linpack performance. These published specifications indicate the scale of the computing installation potentially affected. They do not establish that every node was still in the original configuration in October 2026 or that every node was exposed to the fire.
In November 2021, Chervonenkis ranked 19th and Lyapunov 40th on the global TOP500 supercomputer list. Both remained on the June 2026 list, at 101st and 161st respectively. They represent substantial computing resources, although A100-generation hardware has since been overtaken by newer GPUs designed for AI training.
Linpack measures a system's speed in solving a particular class of mathematical problems. It is not a direct measure of how quickly a large language model can be trained. That depends on GPU memory, interconnect bandwidth, model architecture and software as well as raw computing performance. The two machines' petaflop figures therefore cannot be treated as an equivalent share of Russia's total AI capacity.
How much would replacing the systems cost?
Yandex has not published a financial estimate of the damage. A rough replacement cost can nevertheless be derived from the number of compute nodes and published prices for comparable server systems. The following calculation assumes the replacement of all 336 nodes. It is not an estimate of losses already confirmed after the strike.
One historical benchmark is NVIDIA's DGX A100 platform, introduced in May 2020. NVIDIA announced a starting price of $199,000 for a server containing eight A100 GPUs. At that price, 336 servers would cost approximately $66.9 million. This is a reference point based on the entry-level price of a product from that generation, not documentation of what Yandex paid or an appraisal of the equipment's present value.
A more current benchmark comes from Lenovo's 2026 cost analysis. It listed a typical selling price of $397,801.60 as of June 15 for a ThinkSystem server fitted with eight NVIDIA H200 GPUs. Buying 336 such servers at that unit price would cost approximately $133.7 million. The same analysis priced a system with eight newer B200 GPUs at $550,475.10, or about $185 million for 336 servers.
Those figures do not mean an exact replacement of the old supercomputers would require 336 H200- or B200-based servers. Newer processors have different performance characteristics, and some workloads could reach comparable throughput with fewer machines. This model prices a modern installation with the same number of eight-GPU nodes. It is not a proposed or technically optimized purchase specification.
Server prices also exclude separate high-speed networking, additional storage, installation, electrical connections, cooling upgrades and other engineering work. Allowing an illustrative 15–30% on top of the H200 server cost produces an estimated $154 million to $174 million for the computing installation, rounded to $150 million to $180 million. The allowance is an assumption made for this FACTUM calculation, not a vendor quotation.
In other words, building a modern replacement cluster of similar scale would be a project costing well over $100 million, with the final figure depending on the hardware selected. That is not the established financial damage from October 8. It excludes the other servers at Sasovo and the building's physical infrastructure. Reuters reported that the data center contained tens of thousands of servers, but their inventory and the extent of any damage have not been disclosed. The total loss to the facility cannot be calculated from the available information.
Restoration involves more than buying GPUs
Both supercomputers use NVIDIA A100-based compute nodes connected by high-speed InfiniBand networks. In a large computing cluster, the servers function as a coordinated system. Training a single model can involve many GPUs working together. Failures in the network, storage or power supply can interrupt the entire training job even when the processors themselves are intact.
After a fire, operators need to inspect server racks, electrical distribution equipment, cabling, network switches and cooling systems. Water or fire-suppression agents may also render components unusable without leaving visible burn damage. Surviving servers could be moved to another facility, but that would require spare power and cooling capacity, suitable networking and safe transport. Yandex has released no information about available fallback infrastructure or the condition of the individual GPUs.
Obtaining replacement hardware presents another obstacle. NVIDIA's 2026 annual report says US export controls cover shipments of A100 and H100 accelerators and systems containing them to Russia. The US Bureau of Industry and Security likewise describes extensive export restrictions on technology supplied to Russia. These restrictions do not make repairs or procurement physically impossible, but they constrain direct lawful supply channels for compatible accelerators and complicate plans for a new installation. Published US server prices should not be mistaken for prices available to a Russian buyer.
What an extended outage could mean for Yandex AI
Yandex's supercomputers were built primarily for demanding machine-learning workloads: training and fine-tuning models and running large numbers of experiments. Losing them for an extended period could disrupt the schedule of new research and development work without necessarily bringing down AI features that have already been deployed.
In September 2026, Yandex released the open models Alice AI Search Pretrain and AliceAI-Foundation-80B-A3B-Base. The company continues to develop language, multimodal and search models. Access to large GPU clusters affects how quickly teams can train successive generations, test competing approaches and experiment with new architectures. If available capacity shrinks, researchers must reschedule training runs and reduce the number of large jobs they can execute simultaneously.
In their original published configurations, Yandex's three named supercomputers contained a combined 3,776 GPUs. The 2,688 based in Sasovo represented roughly 71% of the GPU count across those three systems and approximately 68% of their combined Linpack performance. These are shares of three publicly documented machines, not of Yandex's entire AI infrastructure in 2026. The company has expanded its cloud capacity, and the size of its other computing clusters is not sufficiently documented to determine the overall proportion at risk.
A trained neural network can continue to serve users on other hardware if the model files are available there and enough processing capacity has been provisioned. Models do not disappear when one facility stops operating, provided copies exist elsewhere. There is no public evidence that the strike destroyed Yandex's trained models, training datasets or backups. An assessment of the longer-term impact requires information not just about the GPUs but also about checkpoints, storage systems and the location of backup copies.
The implications for Russia's AI infrastructure
The immediate effect was visible in cloud services: the loss of one physical site can disrupt businesses that have no direct connection to the building in Sasovo. Moving workloads to another availability zone requires spare capacity. The fact that an application runs online does not mean it is independent of a third-party cloud provider or a particular data center.
A second potential consequence is reduced computing capacity at one of Russia's largest developers of AI models. If Chervonenkis and Lyapunov were lost in full, the costs would go beyond the purchase of new equipment. They would include interrupted training jobs, workload migration, rebuilding software environments and time spent checking the results of unfinished experiments. No defensible monetary estimate is possible without internal information on utilization, ongoing projects and redundancy.
A third consequence concerns the resilience of the industry as a whole. The attack demonstrated that a major data center supporting both commercial services and AI computing could be taken offline by a physical strike. For other operators, the incident underscores the importance of geographically separated clusters, backup sites, independent power supplies and tested disaster-recovery procedures. Those measures require additional capital investment and ongoing expenditure.
The loss of two Yandex systems would not, however, mean the end of artificial intelligence development in Russia. Other companies operate computing infrastructure of their own. Many AI tasks do not require a supercomputer-scale cluster, while trained models and copies already deployed on other servers do not depend on the continued operation of the machines that trained them. Public sources do not provide a basis for calculating a nationwide percentage loss of AI capacity or a reliable restoration timetable.
What was established by the evening of October 8
The drone attack, fire, full shutdown of the Sasovo data center and disruptions to some services have been reported. Yandex's own published records identify the site as the location of two of its largest historical computing systems. The company has not disclosed damage to the supercomputers themselves or issued a financial damage estimate.
Using published 2026 server prices, a modern installation containing 336 eight-GPU nodes would cost roughly $134 million before additional infrastructure, and an estimated $150 million to $180 million with the illustrative engineering allowance used here. Those figures describe the potential budget for a replacement computing facility. They are not an invoice for losses incurred in the attack.
The technical question after the fire is which components have survived: the servers, networking, power systems and storage. The answer will determine whether restoration is primarily a repair and relocation effort or the construction of a new computing facility. For Russia's AI industry, the consequences depend not only on the cost of the equipment but also on how long the computing resources remain unavailable for model training.