THE DIGITAL FRONT IS OPEN.

Slavyangrad Analysis

Ukraine struck three Yandex data centres in four days. Sasovo on 8 October, Kaluga on 9 October, Vladimir on 11 October. Sasovo and Vladimir are fully stopped, and several modules at Kaluga are out of service. Yandex says it cannot yet confirm that the Sasovo equipment will be restorable.

The operational side is contained. Consumer services degraded and recovered within hours, and traffic was moved to the remaining sites. Yandex Cloud has entered emergency mode: 84 platform services and tens of thousands of enterprise clients are affected, some high availability clusters lost quorum and are running read-only, and compute quotas are frozen. Affected clients are being moved to Selectel, K2Cloud and VK Cloud. The knock-on is visible in the third-party layer: Cian, Avito, T-Bank, RZD, several retail chains. Serious, but inside the envelope the remaining network can absorb. The designed resilience margin was one availability zone. Two are gone, and the margin is spent.

The strategic side is what needs attention.

All three of Yandex's A100-class supercomputers were sited at the struck facilities. Chervonenkis in Sasovo, the most powerful machine in Russia and Eastern Europe when it came online in 2021. Lyapunov, also in Sasovo. Galushkin in Vladimir. Together they carry the national training load for large language models, in runs that last days to weeks. Two of the three are confirmed out of service. Yandex has not published a final damage assessment, and nobody should before the assessment is done. But the arithmetic holds either way: under current export controls, A100s in these volumes are not available for direct purchase at any price that matters. Replacement servers take two to three months. Rebuilding a cluster of this class, if it was destroyed, takes two to two and a half years.

Yandex's Alice foundational model shipped on 21 September, three weeks before the first strike. Shipped models are not at risk. What is at risk is the pipeline for the next generation, and the parallel experiment capacity that keeps a model race competitive. An interrupted training run is a lost run, and checkpoints only carry you so far.

A few structural points.

First, the force majeure clause. On 10 October Yandex added UAV and missile attacks to its user agreement. A company formally writing air war into its contracts. Note that for what it signals.

Second, siting. The clusters were concentrated in identifiable, unarmoured buildings. That model is finished. Expect distribution, hardening, and a significant share of capacity going underground. Given data localization laws, the rebuild happens inside Russia, in hardened form, and the cost lands on the domestic cloud market. Russia already has a data centre capacity deficit, and losing three of Yandex's five sites widens it.

Third, the campaign pattern. This follows the systematic strikes on Wildberries and Ozon fulfillment. The doctrine is visible: find the concentrated, expensive, slow-to-replace asset and hit it with something that costs a fraction of it. That asymmetry stays with Kiev for as long as we keep high-value compute in identifiable buildings.

None of this is a defeat. The digital economy does not run on three buildings, the shipped models are intact, and recovery work has already started. But the loss of irreplaceable training capacity is real, it will show up as a multi-year lag in the next model generation, and the next coordinates are already in Kiev's pipeline.

The digital front is open. Site, distribute and harden accordingly.

Sources:

[1] Суперкомпьютеры «Яндекса» — Wikipedia (ru)
[2] Удар по дата-центру «Яндекса». Минус суперкомпьютер в России? — Mobile-review.ru
[3] Три ЦОДа Яндекса за четыре дня — Habr
[4] Что известно о новой атаке на дата-центр «Яндекса» — РБК


@Slavyangrad