The Iron Heart

The strikes acquired a second life online within days. The claim was that hitting Yandex had silenced Russia's bot army and made Western social media calmer. A viral chart showed automated traffic from Yandex's network collapsing to a fraction of its peak, offered as proof, and Ukraine's president repeated the gist from a security-forum stage. It is worth clearing this away first, because it is neither what the data shows nor this report's subject. The chart counted automated requests from Yandex's network, the overwhelming majority of them Yandex's own search crawler indexing the web, not social accounts. Cloudflare, whose data it was, said the fall was "not necessarily bot behaviour on social media platforms," and its own post-mortem found the malicious automation did not disappear but moved, roughly 170 million requests shifting to other hosts inside a week. More basic still, network traffic is not attribution. Influence operations route through intermediary hosts and jump boxes so that origin cannot be read off a graph, and Russian bulletproof hosting is in any case a nationality-agnostic commodity, rented by ransomware crews and influence operations alike. A dip at one provider says nothing about who was behind the requests. No platform data was ever published for the Truth Social or "MAGA" version of the claim at all. (Cloudflare and RFE/RL.)

None of this proves a negative. A few operations that did sit on Yandex, the "Pravda" content network among them, dipped for a matter of hours before recovering, and isolating a change in propaganda accounts alone is, as one researcher put it, exceedingly difficult. But the headline claim rests on a misread chart, the Ukrainian president's own office later clarified he had meant the "atmosphere" online rather than a measured fall in bots, and attribution from traffic volume is not something the method supports. We note it, and turn to what the strikes actually hit, which is both verifiable and, as it happens, the more consequential story.

When the first drone reached Sasovo on the night of 7 October, Yandex answered with a statement it headed, in its own words, "The Iron Heart." The data centres, it said, are the железное сердце (zheleznoye serdtse, "iron heart") of the company, the physical plant beneath services used by millions of people and hundreds of thousands of businesses. Three of those centres were struck inside five days, at Sasovo in the Ryazan region, at Kaluga, and at Vladimir, and two were shut down completely. For a company that had spent years describing its infrastructure as a heart, the aim was precise. (Yandex, investor relations.)

Sasovo and Vladimir matter more than the outage counts because of what they hold. By Yandex's own account, the company trains its neural networks on three supercomputers, named for the mathematicians Chervonenkis, Galushkin and Lyapunov, built on AMD EPYC processors and Nvidia A100 accelerators and used, in Yandex's phrase, to "train neural-network models with billions of parameters." Chervonenkis and Lyapunov sit in Sasovo, and Galushkin is in Vladimir. When they entered the TOP500 list of the world's fastest machines in 2021, they placed 19th, 36th and 40th, the three most powerful computers in Russia. The drones reached the buildings holding all three. (Yandex newsroom.)

The visible damage was to services. Yandex Cloud, which serves more than sixty thousand customers, dropped into an emergency mode and told clients to fail over to their own disaster recovery. Roughly eighty services degraded, from retailers and internet providers to an online-therapy platform and the Russian segment of TikTok, with knock-on outages as far as Central Asia, and the share price fell about three percent. Yandex said it could give no timeline and "cannot confirm whether the equipment can be restored." (RFE/RL.)

Here the record shifts from what Yandex has confirmed to what analysts assess, and the line matters. Specialists cited by Meduza and the science outlet T-invariant judge that development of YandexGPT and the Alice assistant has been suspended indefinitely, that clearing the soot damage could take up to six months, and that fully restoring the machines could take one to two years if components have to be replaced. None of this is company-confirmed, Yandex has said nothing about the machines' condition, and the widely shared figure of 2,700 destroyed GPUs is an unverified open-source claim, not a fact. But one constraint needs no insider. The A100 is export-controlled and globally back-ordered, with supply booked two to three years out, so whatever the drones broke, Russia cannot simply buy again. That is the difference between an outage and a setback. (Meduza.)

Set the two halves of Russia's digital power beside each other and they fail in opposite ways. The AI layer is concentrated in three named supercomputers in two buildings, a physical target a drone can reach, as demonstrated by the strikes. The influence layer is the reverse, distributed and deliberately offshore, as the opening section laid out, rented from bulletproof hosts, routed through intermediaries, and, when one provider went dark, migrated to others inside a week. One is a fixed address. The other is a habit of changing address.

This is why the viral story got the casualty backwards. It mourned the influence layer, which was barely touched, and looked past the AI layer, which was not. The automated traffic "silenced" at Yandex was the search crawler, not the influence operations it was mistaken for. The supercomputers that train Russia's frontier models are precisely the thing a precision strike can take off line and, under sanctions, keep off line. The internet held a wake for the resilient half and missed the fragile one.

A troll network survives a strike because dispersal is cheap. It stands up new hosts, re-announces its address ranges, and carries on. A frontier-model cluster cannot be dispersed the same way. You cannot smear three A100 supercomputers across borrowed cloud regions, and you cannot reorder the chips. Even the cooling works against it. By those same assessments, Yandex's halls draw outside air directly, so smoke and soot from a single hit can settle on the boards and short them. The layer that looks most sovereign, the home-grown models and the machines that train them, turns out to be the one most exposed to a few kilograms of explosive and an export-control list.

On 24 August, six weeks before the first drone reached Sasovo, President Putin signed Decree №604, "On measures to ensure the security of critical-infrastructure objects." It gives the government, acting on the president's instruction, the power to place a critical-infrastructure owner under временное управление (vremennoye upravleniye, "temporary management"), with the Federal Agency for State Property Management, Rosimushchestvo, as the default manager. The reach is not small. The decree extends temporary management not only to the physical site but to the owner's movable and immovable property, its securities, and its shares in Russian legal entities. (Rossiyskaya Gazeta, official text.)

What brings an owner under the decree is the part that reads, after October, like a description of Yandex. Temporary management can follow when an owner is judged to have acted ineffectively on security, in the decree's words "the ineffective implementation of measures to counter the threat of attack using unmanned aerial vehicles," or to have failed to restore an object's functioning, or restored it too slowly. Yandex has just had three data centres struck by drones it could not stop, with analysts assessing that full restoration could take one to two years. The trigger language and the company's situation are close to a match.

The honest qualifier is that this is exposure, not a seizure, and perhaps not even an applicable one. The decree's list of covered objects names communications facilities (объекты связи, obyekty svyazi) and a catch-all for sites "of special importance," but it does not name data centres, and whether Yandex's halls qualify is untested. Officials have framed the instrument narrowly. First Deputy Prime Minister Manturov called it "not nationalisation," and the Kremlin's spokesman cast it as protection against drone strikes. The point is not that Moscow will take Yandex. It is that the company's own state wrote itself a lever over critical-infrastructure owners who cannot defend or restore their sites, six weeks before Yandex became the clearest example of one. The sovereign-AI champion now carries a second-order risk, and it does not come from Kyiv.

Put the findings together and "sovereign AI" turns out to mean something more brittle than it sounds. The models Russia calls its own are trained on three machines in two buildings that a drone can reach, run on chips that cannot be reordered, and owned by a company that, having failed to defend its sites, now sits within reach of a decree letting its own government take them into temporary management. Sovereignty, on this evidence, is not resilience. It is a flag planted on infrastructure that is centralised, un-replaceable and legally exposed at the same time.

This publication argued in its last issue that Russia's "sovereign AI" is a control layer rented on Chinese foundations. The strikes supply the physical companion to that reading. A sovereignty you can bomb, cannot rebuild under sanctions, and can lose to your own state is sovereignty in name. The drones did not disprove the claim. They showed what it was resting on.