Drones Shut 3 Yandex Data Centers in 4 Days. The Cloud Behind Your Exchange Has a Street Address Too

What Happened

Yandex, the Moscow-based company behind Russia's dominant search engine, its largest ride-hailing service, and 1 of its largest cloud platforms, said on Sunday, October 11, that a drone attack had damaged its data center in Vladimir, about 180 kilometers east of Moscow, and that 'the data center's operation has been completely suspended.'

Its cloud status page put the platform in emergency mode with 2 availability zones, ru-central1-a and ru-central1-b, unavailable, suspended the issuing of new resource quotas, and recommended that customers 'use alternative disaster recovery plans.' Coverage counted more than 80 Yandex Cloud services down, from compute and managed databases to the YandexGPT API, alongside consumer services from the Alice assistant to taxi ordering, with disruptions reported in Kazakhstan, Belarus, Armenia, and Serbia.

Yandex reported no injuries and gave no restoration date.

It was the third Yandex facility hit in 4 days. On Thursday, October 8, drones set fire to the company's data center in Sasovo, Ryazan region, which houses 2 of its 3 supercomputers; operations there stopped completely, and services that rent Yandex infrastructure, including Russian Railways and the mobile operator MegaFon, went down with it.

The Moscow Times called it the first confirmed strike on data infrastructure inside Russia since the 2022 invasion. On Friday, October 9, the company told Interfax that drones had hit its Kaluga campus, its largest at a designed 63 megawatts, and that 'several data center modules are completely disabled.'

Yandex's notices say only 'UAV attack.' President Volodymyr Zelensky, asked about Sasovo on October 8, said 'we always respond in kind' and that Russia 'has been and still is striking our data centers.' No Ukrainian military statement on Vladimir had surfaced as of this writing.

Yandex told its customers to run their own recovery plans while its buildings burned. Keys you hold do not need a building. Most of what sits between you and the chain still does, and the week showed what a list of buildings is worth to a drone.

Multi-Zone Was Built for a Fire, Not a Campaign

Cloud providers sell resilience as geography. An availability zone is a data center, or a cluster of them, far enough from the others that a fire or a flood takes out 1 zone and not the region. Yandex's allocation documentation tells customers to spread instance groups across its zones in the Vladimir, Moscow, and Ryazan regions.

The design assumes failures are independent. A targeting cell with a list does not fail independently: it hit Ryazan on Thursday, Kaluga on Friday, and Vladimir on Sunday, and by Sunday morning 2 zones were down at once, the platform was in emergency mode, and the provider's advice was that customers should execute their own recovery plans.

The redundancy had not vanished. It had been enumerated.

The Gulf is the control case. On March 1, Iranian drones struck 2 Amazon Web Services (AWS) facilities in the United Arab Emirates and a nearby blast damaged a third in Bahrain; AWS said 2 of its 3 UAE zones were 'significantly impaired,' and CNBC reported outages at UAE banks, a ride-hailing app, and 2 payments firms.

On September 15, AWS told customers that resources and data which had existed only in 1 UAE zone could not be recovered, that the Bahrain damage 'exceeded what our regional and multi-AZ services are designed to withstand,' and that most customers had re-established operations in other regions 'by restoring backups or copying data that remained accessible.'

Those were the customers who had backups somewhere else. Iran's Fars news agency said the Islamic Revolutionary Guard Corps (IRGC) chose the targets for their support to the US military, and on March 31 the IRGC named 18 companies in the Gulf, Microsoft, Apple, Meta, and Oracle among them, as legitimate targets.

A Semafor column observed that of 233 data centers across the Gulf, only 3 were affected. That is true, and it is also the point: 3 were enough to leave a region unavailable for 6 months, and data a hyperscaler cannot recover is no longer hypothetical.

Compute Is a Target Class by Doctrine Now

None of this is accidental collateral. Russia's defense ministry said on September 26 that it was striking Ukrainian data centers because they 'process and transmit intelligence data for the Ukrainian Armed Forces,' claimed hits on facilities of Kyivstar and Omega Telecom the next day, and described the result for Ukrainian soldiers as an involuntary 'digital detox.'

Ukraine's digital authorities said the strikes cut internet service to about 100,000 households around Kyiv, and Meduza counts at least 8 companies whose data centers have been damaged or destroyed since late September. Kyiv's answer was the same target class in reverse, and the choice of Yandex is instructive: 1 company's buildings carry the search engine, the taxis, the cloud that a railway and a mobile carrier rent, and the supercomputers behind its language models.

Drones that cost a fraction of a cruise missile produced outages in 4 countries, and Semafor's Sunday summary was that the 2 sides 'have traded strikes on digital infrastructure in recent weeks, triggering major outages on both sides.'

This is the fifth-generation warfare pattern the site tracks: inexpensive autonomous weapons aimed at the infrastructure a society runs on rather than the army that defends it, with the effects delivered to civilians through services they did not know shared a building. The counter-argument is that data centers serving military networks can be lawful targets, and that is correct as far as it goes; the Russian and Iranian statements both lean on it.

What is new is the ratio. Sam Winter-Levy, a Carnegie Endowment fellow, told Anadolu Agency after the Gulf strikes that data centers 'become much more attractive as targets for states or non-state actors that want to cause disruption.' The concentration that makes the cloud efficient is what makes it worth a drone.

What This Means for You

Translate it to your own stack. A balance on an exchange is a row in a database that lives in a cloud region, next to the matching engine, the hot-wallet signer, and the compliance system. Coinbase's retrospective on the October 20, 2025 AWS outage counted 3 hours and 17 minutes of degraded service from a fault in 1 US region, and nobody had fired anything.

The Gulf and Russian cases are the kinetic version of the same dependency, and in the Gulf some of the data did not come back. That is custodial risk in its most literal form: the service, the record of what it owes you, and the staff who would process your withdrawal all have a street address.

Bitcoin held in self-custody does not. The ledger is replicated on thousands of independently run nodes in dozens of countries, there is no availability zone to lose, and the private keys on your hardware wallet were never inside anyone's building.

Name the limit, because this story is mostly about it. Keys survive the building; your access to the chain runs through buildings all the same. Most wallet apps query balances and broadcast transactions through a server the vendor runs, by default; the price you see comes from an exchange's API; the on-ramp you would buy on and the venue you would sell on are cloud tenants.

A holder who has never connected a wallet to anything but the vendor's backend has self-custody of the keys and a custodial dependency for everything else. The distribution of the Bitcoin network protects the ledger. It does not automatically protect your path to it. Yandex's advice to its customers mid-incident, run your alternative recovery plan, was only useful to the customers who had one.

So build the plan while nothing is burning. Run a node, or at minimum know how to point your wallet at one that is not the vendor's: Trezor Suite, for example, accepts a custom Electrum or Blockbook backend in its network settings, and many desktop wallets can reach a node over Tor.

Keep a seed backup that depends on no service, in a second place; for multi-sig, keep the wallet descriptor and each cosigner's public key offline with it, because a recovery without them is a recovery that needs someone's server. Keep exchange balances sized for use rather than storage, and confirm you can withdraw from each venue before an outage makes that the question.

Then test the fallback once: broadcast through your own node, restore the backup on a spare device. Any step that needs a company's servers to be up is the building.

What to Watch

Yandex has given no restoration dates for Sasovo, Kaluga, or Vladimir, and whether the 2 dark zones come back in days or, as in Bahrain, not at all is the first signal of how much of the damage is to buildings rather than to servers. Any Russian strike on a Ukrainian cloud or payments provider in the coming week will show whether the exchange has settled into a tit-for-tat on compute.

AWS has promised a further update on the 2 UAE zones it is still trying to recover, and that update will say whether 'cannot be recovered' spreads. Closer to home, the next cloud-region outage, kinetic or not, will produce the same exchange status page; the useful question is which of your tools kept working through it.

Tracked entities this report names, with total coverage across the archive. Follow one into the Actor Network to see every report it appears in.

Other reports that name the same tracked actors — a different thread than “related,” which follows wording. These are linked by who appears in them.