/Technology

One thing has been nagging me since Thursday, and it wasn't the smoke in the Sasovo footage. It was Yandex's statement telling customers to "activate their disaster recovery plans." I read it twice. Anyone who has ever written one of those plans knows what it usually is: an eleven-page PDF someone put together in a bad week before an audit, and that nobody opened again. Now half a country's digital life was being told to open it. Let's go through the facts, which are few, and which it's best not to inflate. On Thursday, October 8, drones hit a Yandex data centre in Sasovo, in the Ryazan region. Fire, operations suspended, reports of problems in Yandex Cloud, Disk and Documents. A day later, on Friday, it was Kaluga's turn, the largest of the five, with several modules knocked out. Then Vladimir, the third in under a week, with the company writing on Telegram that operations were "completely suspended" and that the platform was running in emergency mode. Yandex has five data centres, four in Russia and one in Finland. Three of the four Russian ones have been touched. Do the maths. I should say right away that Ukraine has not confirmed the Sasovo attack, which is the usual pattern. And there is an annoying ambiguity: next to Sasovo there is a plant under British sanctions since 2024, and the press couldn't say whether the target was the data centre, the plant, or both. I'm not going to settle that. Anyone who says they know is selling something. Now, what really interests me. I spent a rotten part of my working life explaining to clients the difference between "having redundancy" and "having redundancy that works." They are different animals. In 2017, on a Wednesday that still gives me heartburn, a disk failed on a volume that was supposedly mirrored, and the mirror had stopped syncing eleven weeks earlier. Nobody noticed, because the alert was going to a mailbox that no longer existed. That's what comes to mind when I read that Yandex "can redistribute part of the load to its other data centres." It can. In theory it can. The question is how much, for how long, at what latency, and whether those other centres were already at 70% utilisation before the load of a site with tens of thousands of servers landed on them. Tens of thousands of servers. Stop on that number for a second. That's not a server room, it's a small town made of fans. There is a comfortable idea that circulates a lot among people who write about infrastructure, that the cloud is an ethereal, distributed concept with no single point of failure. Honestly, that always sounded like advertising to me. The cloud is a building. It has an address, a power substation, cooling pipes, a fibre line that goes through a hole in a wall. Everything "distributed" about it is distributed across a finite number of buildings, and Yandex has four in Russia. Four. When someone manages to hit three of them in a week, the word "distributed" loses almost all practical meaning. And here I hesitate, because the easy reading is "this is the end of Yandex's services" and I don't believe that. The company is still running. Services degraded, they didn't vanish. The shares fell 3.75% in Moscow, the biggest drop on the exchange that day but far from a collapse. Big systems degrade gracefully, or at least they degrade while dragging their feet, and Yandex has very good engineers, who built a search engine capable of holding off Google in a market where Google barely managed to get in. Underestimating them would be stupidity on my part. But I'm also not convinced by the opposite narrative, the "resilient Russia that absorbs the blow," because absorbing a blow with one data centre in emergency mode and another with dead modules is not the same as absorbing it without consequences. What I really want to discuss is what the dead modules cost, and the answer has little to do with the building. A building gets rebuilt. Concrete, cables, even cooling can be bought. What can't be bought overnight is accelerators. The Institute for the Study of War, in Washington, wrote that export controls and sanctions will make it harder for Yandex to quickly replace advanced semiconductors that were destroyed. That sounds like a technical footnote and is, in my opinion, the centre of the whole story. If these centres hold the capacity to train and serve models, and Yandex runs YandexGPT, one of Russia's main families of generative models, then every lost module is GPU that doesn't come back through the normal channel. It comes back through the parallel one: triangulated imports, intermediaries in the Emirates, in Kazakhstan, in Armenia, in places where the invoices look pretty and the deadlines look ugly. It worked like that before. Now there is sudden demand for replacements, and those who supply through the grey channel know how to charge for urgency. Second-order effect: the price goes up. Third-order, the one nobody does well: Russia has to decide who gets the hardware it can find. Yandex Cloud, which sells to businesses? AI projects with strategic value? State bodies? A scarce resource forces political decisions about who gets served, and political decisions in a system like the Russian one have an annoying tendency to reward whoever is closest to power rather than whoever produces the most value. I've seen this film on a microscopic scale, in a six-person team, when there was one good test machine and three people wanted it. The one who got it wasn't the most important project. It was whoever shouted first. At the level of a country the scale changes, but the mechanism is the same. (An aside. When I started running sites with some real traffic, around 2012, I put the backups on the same server as the data, in another folder, with a cron job running at 03:15. I thought it was prudence. A disk died and took everything with it, data and backup together, because they were physically on the same spinning metal platter. Almost everything I know about infrastructure came from embarrassments like that.) Back to the question of reciprocity, because that's where things get uncomfortable. Zelensky said, about Sasovo, that they always respond "in mirror-like fashion." He is right in the factual sense: Russia hit at least eight major data centres in Kyiv, causing internet blackouts across Ukraine, and Moscow justified it by saying they processed information for the Ukrainian military. Now Russia's Defence Ministry says the same about Yandex, in reverse, attributing to it the processing of intelligence for Ukraine. The symmetry of the argument is almost comic. Each side accuses the other of using civilian infrastructure for military purposes, and each side uses that accusation as a licence to hit infrastructure that serves millions of people who never asked for anything. I'm not going to pretend to moral equidistance here, because I don't have it: it was Russia that started the war, and it was Russia that, in this specific episode, opened the round of bombardments on data centres in September. But there is another thing, and it is both cynical and technical. Nobody in the engineering world should be surprised. A Russian military blogger, Artjockey, wrote on Telegram that Russia was ill-prepared for a "data centre war," and that Ukraine had prepared to be a target while Russia had not. You can disagree with the tone, but the underlying point is impeccable: whoever started the escalation didn't notice that their own ground was as fragile as the adversary's. There is an asymmetry of preparation. Ukraine has lived through four years of bombing of its power grid and learned to decentralise, to keep generators, to spread services across several countries, to move data out of the territory very early. Russia made the opposite bet, that of digital sovereignty, of "everything at home." One of the hidden costs of that sovereignty is precisely this: you concentrate everything and then you have fewer places to hide. The strategy follows a pattern we could already see. Nova Poshta's warehouses, in Ukraine, were hit in the summer; the answer came on Wildberries' warehouses, and Ozon's too. Now, data centres against data centres. Each rung climbs one step of "economic value" for every step the other side climbed. First refineries, then ports, then e-commerce logistics, now computing. If this continues, the next rung isn't hard to guess, and I don't like writing it. I don't like it, but I'll write it: the next rung is submarine cables, internet exchange points, or the links that cross borders. It has already happened to Ukraine, in a limited way. And that's where the rest of us start to have a personal interest in the story, because the internet doesn't respect war borders as well as generals would like. This connects me to something else. A lot of time is spent talking about AI as if it were software. Pure, abstract software, which runs "somewhere." It doesn't run somewhere. It runs on very specific hardware, in a small number of places, fed by electricity in quantities that still surprise me when I see the bills of people who operate real clusters. This story is, at bottom, the first time someone deliberately hits that hardware, in a sustained way, because of what it does and not just where it sits. Or at least it's the first time it has been confirmed, according to Newsweek, against Russian data infrastructure since the start of the full-scale invasion. There may have been earlier things that were not confirmed. I'm not claiming that certainty. That changes the conversation about risk, and it changes it even for people who have nothing to do with Russia or Ukraine. Anyone building products on top of APIs from large providers should have asked the boring question this week: where is the data centre that serves my traffic? How many are there? What happens if the primary disappears and the secondary is at 94% occupancy? I ask myself the question about the platforms I run, and the honest answer is that I depend on more things than I like to admit. A single region from a single provider for half of what matters. I know I should change it. I haven't yet. Laziness, cost, and that little voice that says "it's never happened." The "it's never happened" voice is the most expensive one in the business. (Another detour, a short one: a few years ago, on an August weekend, a DNS TTL set to 86400 seconds made me wait a whole day for a migration that should have taken forty minutes. I had forgotten to lower it the night before. Nobody died, but I kept the rule forever.) Back to Yandex, there is a question that is rarely discussed and that intrigues me: what the company actually loses, in terms of trust. Yandex lives on something engineering rarely measures: habit. Millions of Russians open Yandex in the morning without thinking, for the map, the taxi, the email, the search. If those services fail for an afternoon, the company recovers. If they fail intermittently for weeks, people start having alternatives in their pocket, and the alternatives exist. Here comes the unpleasant side, because the alternatives in the Russian market are, in large part, even more controlled by the State. In other words, a blow to Yandex may push users towards platforms more permeable to surveillance. Third-order effects rarely point in the direction the planners imagined. I don't know whether that is what will happen, but it wouldn't surprise me. And then there is the enterprise market. Yandex Cloud sells computing to Russian companies that, since 2022, were left without Western suppliers. These customers have nowhere to go. If Yandex fails, plan B is another Russian cloud, a few kilometres away and with the same hardware problems, with a fraction of the engineering muscle. That's why the risk premium of Russian domestic infrastructure goes up, even without Yandex disappearing. Raising the risk premium is the practical goal of many infrastructure attacks: not destroying everything, just making everything more expensive and more uncertain. And it works better than people think. I contradict myself here, since I'm halfway through. I said earlier that Yandex will withstand the blow. I stand by it, but with an enormous asterisk: it will withstand it if the attacks stop here. If it's three in eight days and the pace holds, what is being measured is not the hardness of each blow, it's the cadence. Any operations team knows that a week with three serious incidents is different from three serious incidents in three years. In the week of the three, people stop sleeping, procedures turn into improvisation, and improvisation produces errors that do more damage than the original attack. I've read post-mortems in which the fire was the cause of the first outage and human fatigue was the cause of the second, the worse one. Here, I have the impression that the cadence is part of the message. I also wonder what lessons Ukraine itself is drawing from the receiving end, because the irony is huge. Kyiv suffered the attacks first and suffered them at a moment when it had much less margin. Ukraine has been, curiously, an involuntary laboratory of digital resilience: data spread across foreign providers, administrations working with copies kept outside the country, State services that survive attacks on the network. When people talk about "data sovereignty," the Ukrainian case shows that sometimes sovereignty is precisely being able to put your data somewhere else. The opposite of Russian rhetoric. This is one of those sentences that sounds like a slogan to me and that I still find true. What remains is the angle that scares me most, honestly, and it's the precedent. If today data centres are hit because they "process information for enemy militaries," tomorrow any data centre anywhere serves as justification, because practically all of them process something that someone can link to something military. The border between civilian and military in cloud computing is a line drawn in pencil, and the pencil is soft. The same server farm hosts a bakery's email, a city council's records, the training of a model, and maybe a defence contract. Hitting the building distinguishes none of it. Anyone who insists these are "legitimate targets" should explain where the list ends, and I don't remember hearing anyone answer that well. I'll close with something small and practical, because the big stuff ended up tiring me out. If you run something on the internet, even a modest project like the ones I run, go and check today where your backups are. Not "whether they exist." Where they are, on which platter, in which building, in which country. And test a real restore, not a pretty simulation. I'm going to do the same after writing this, and I'll probably discover something I'd rather not know. --- **Sources:** - Al Jazeera, Ukraine takes aim at Russia's AI data infrastructure: https://www.aljazeera.com/news/2026/10/9/ukraine-takes-aim-at-russias-ai-data-infrastructure - Al Jazeera, Yandex says data centre in Vladimir shut down after drone attack: https://aljazeera.com/news/2026/10/11/russias-yandex-says-data-centre-in-vladimir-shut-down-after-drone-attack - Reuters via ThePrint, Ukraine expands drone strikes on data centres owned by Russia's Yandex: https://theprint.in/world/ukraine-expands-drone-strikes-on-data-centres-owned-by-russias-yandex/3066688/ - Newsweek, Ukraine strikes major data center of "Russia's Google": https://www.newsweek.com/ukraine-strikes-major-data-center-of-russia-google-drone-attack-yandex-russia-war-12539331 - Newsweek, Ukraine strikes second data center of Russia's Google in two days: https://www.newsweek.com/ukraine-strikes-second-data-center-of-russias-google-in-two-days-12546205 - The Daily Beast, Putin's Nightmare Deepens as More Drones Hit "Russia's Google": https://www.thedailybeast.com/putins-nightmare-deepens-as-more-drones-hit-russias-google-yandex - The Record, Major Yandex data center in Russia hit by Ukrainian drone strike: https://therecord.media/yandex-russia-drone-ukraine