Deep Dive — Drones just proved AI compute is a wartime target

Four days of Ukrainian drone strikes took out half of Yandex's data centers, including the buildings that hold all three of its training supercomputers. The rest of the industry is now asking a question it had been avoiding: what happens when the machines that make AI are also the easiest thing to hit.

The four days

Russia started it. Through late September, Russian drones and missiles hit Ukrainian data centers and telecom infrastructure, temporarily cutting internet and cell service for up to 100,000 households in Kyiv. Russia's defense ministry mocked the campaign on Telegram as a "digital detox." Ukraine answered in kind, and the answer landed on the country's largest technology company.

On October 8, drones hit Yandex's facility at Sasovo in the Ryazan region southeast of Moscow. That campus holds two of the three supercomputers Yandex uses to train YandexGPT: Chervonenkis, rated at 29.42 petaflops and ranked 101st on the Top 500 list of the world's fastest machines, and Lyapunov at 20.03 petaflops, ranked 161st — both built on Nvidia A100 GPUs. Yandex said operations there were "completely halted."

On October 9, a second strike knocked out the Kaluga region site southwest of Moscow, one of the country's largest with a maximum capacity of 63 megawatts. Local officials reported nine people injured. On October 11, a third strike forced Yandex to suspend activity entirely at its Vladimir facility east of Moscow, which houses the third training system, Galushkin — 20.64 petaflops, ranked 134th — and left about 1,600 customers in the region without electricity.

Three of Yandex's six data centers, in four days. Per the BBC's Russian service, only three escaped attack: one more in the Vladimir region and two near Moscow. Yandex has confirmed the service damage in its own Telegram statements and it moved fast on the legal front too, amending its user agreement on October 11 to classify attacks "using drones... and other weapons" as force majeure events.

What actually broke

The buildings are the visible damage. The part that decides how expensive this gets is the machinery inside, and nobody has said. Ars Technica reported that the extent of damage to the supercomputers was not immediately clear; Yandex's own statements describe facilities, not hardware. That distinction is the whole story: a wrecked campus is a real-estate and insurance problem measured in months, a wrecked training cluster is a compute problem measured in quarters, because you cannot buy 70 petaflops of A100-class capacity from a catalog.

The service layer fell over anyway. Yandex Cloud — more than 64,000 corporate customers — switched to emergency mode after the third strike, which analysts read as the company running out of spare capacity to shift load onto. "That's a very bad sign," Eldar Murtazin, lead analyst at Mobile Research Group, told Fontanka. Companies that rent from the cloud went dark with it: Kommersant's review of outage data named grocery chain Magnit, fast-food chain Vkusno i Tochka and electronics retailer M.Video; the Moscow Times added retailer Wildberries. Yandex's own consumer apps followed — taxi, delivery, maps, music — and Yandex Pay went down hard enough that the fintech division's CEO, Dima Pirogov, posted a public reassurance: "All your money is safe. The money itself isn't stored in the data center."

The blast radius crossed borders, which is the part worth lingering on. Users in Kazakhstan, Kyrgyzstan, Uzbekistan and Armenia reported trouble reaching Yandex apps; in Yerevan, taxi demand surged onto a local competitor. Meanwhile Yandex's IT staff told customers to arrange backup capacity with competitors Selectel, K2Cloud and VK Cloud — an unusual thing for a company to say out loud, and evidence that the rebuild is measured in months.

What Yandex was supposed to be able to absorb: BBC News Russian reported the company ran regular drills in which it shut one of its six data centers and shifted load to the rest. The drill scenario was one site down. Ukraine took three, two of them fully offline. Analysts told Kommersant that replacing the damaged server equipment could take up to six months.

The concentration trap

The strategic critique did not come from Kyiv. It came from pro-Kremlin military bloggers, translated by the independent researcher ChrisO_wiki. The milblogger writing as Ratio Militaris argued that Yandex's hyper-centralized ecosystem — a near-monopolist at the market level, with the company's entire capacity concentrated in a handful of very large nodes — had flipped from national advantage to wartime weak link. A second blogger, Artjockey, pointed at the trap underneath it: Russia's 2019 Sovereign Internet Law effectively bars key services from being hosted outside the country, and the capacity to mirror them in Belarus or Kazakhstan does not exist. Two-thirds of Russia's data centers sit in the Moscow region and St. Petersburg.

Russian officialdom had already drawn the conclusion and not acted. On October 6 — two days before the first strike — Deputy Digital Development Minister Yevgeny Filatov urged companies to move computing capacity "somewhere beyond the Urals, where, fortunately, the strikes haven't reached yet." Digital Development Minister Maksut Shadayev told a State Duma committee on October 12 that his ministry and the Defense Ministry were "working on the issue of protecting these facilities," calling the attacks unprecedented and crediting backup systems and distributed capacity for keeping core services alive.

Ukraine's advantage in this exchange is structural, not technical. Its digital estate is smaller and more dispersed, some of it moved underground or abroad, and Western hyperscalers handed its government a resilience layer for free: Amazon Web Services alone hosts more than 15 petabytes of data from over 60 Ukrainian government organizations, with Microsoft providing similar support. Russia's equivalent escape hatch is a law that forbids using it. Personal data and telecom traffic — six months of it, plus three years of metadata under the Yarovaya laws — must stay inside the country, which means Russian companies must absorb the next strike among Russian data centers.

Containers, netting, and radar

The market's response is already visible. Russian IT companies and mobile operators have been buying container-sized mobile data centers in growing numbers since late summer, according to BBC News Russian — buyers include operators and mid-sized businesses that need more compute than a workstation. Yadro, part of X Holding, launched its Borey containerized data center this year at 18 million to 25 million rubles without servers, and hundreds of millions with equipment installed; the BBC linked the holding's deputy CEO, Boris Korolev, to a family connection with FSB First Deputy Director Sergey Korolev. Rostelecom subsidiary RTK-TsOD is another supplier. Procurement records reviewed by Meduza also show VK, MegaFon and Rosatom buying protective frames and netting for facilities, drone detectors, and employee shelters.

Containers are a real answer for a company that needs somewhere to put a rack. They are not an answer for a training cluster that needs a substation, a chilled-water loop and tens of megawatts of steady power. The honest read is that Russia's compute map will fragment into smaller, harder-to-hit, less efficient sites — buying resilience with performance.

The defense side has a procurement pipeline in the West that Russia lacks. In the United States, Echodyne announced on October 1 a contract supporting Joint Interagency Task Force 401's Domestic Shield program with a $250 million ceiling and a 36-month ordering period, putting its MESA radars on the government's counter-drone marketplace for base defense and force protection. The company's EchoShield radar is also the sensor inside Trust Automation's Small Unmanned Air Defense System under a US Air Force contract worth up to $490 million, and Echodyne says it can track a small quadcopter at roughly 3 kilometers and a larger multirotor at about 5.3 kilometers. It recently opened an 86,350-square-foot factory in Washington State rated for more than 30,000 radars a year.

Here is the catch the industry admits in public, through vendors writing in the trade press: current FAA and FCC rules limit drone mitigation to special conditions, so in nearly all cases data center operators cannot jam a drone's signal, shoot it down, or block its flight path — even during a clear threat. What they can do is detect, and then run "soft mitigation": lock partitions, power systems down, evacuate, call law enforcement. More than one million drones are registered in the United States, and drone incursion incidents at EU airports rose fourfold in under two years. Physical security consultants are now pushing a Drone Vulnerability Risk Assessment into the design phase of new builds, because retrofitting airspace sensing onto an operating campus is more expensive and leaves it exposed while the work happens.

Why the rest of the world should care

Yandex is a specific case — sanctioned, state-adjacent, monopolistic, and legally required to stay concentrated. Its problems do not transfer cleanly to Microsoft or Amazon, which build multiple regions and availability zones and say so in their architecture documents. But the design assumption underneath all of it is technical failure: a burned-out switch, a tripped breaker, the loss of one zone. The system was never architected against an adversary that deliberately picks the day.

The scale only raises the stakes. Epoch AI's tracking has the record for the largest single AI data center by IT power capacity doubling roughly every ten months since mid-2024 — xAI's Colossus 2 at an estimated 950 megawatts as of June, with Anthropic's Amazon-built New Carlisle campus projected past a gigawatt this month and Microsoft's Fairwater in Wisconsin projected at more than two gigawatts by 2028. Epoch's cost model prices a typical one-gigawatt AI data center at about $38 billion in up-front capital spending and roughly $8.5 billion a year in annualized total cost, with servers alone at 60%. Fixed assets of that size are not portable, not relocatable, and not insurable on the assumption that nobody aims at them.

The precedent is only seven months old. Iranian drones struck three AWS facilities in the Gulf on March 1, 2026 — two data centers in the United Arab Emirates hit directly, one in Bahrain affected — and AWS declared the Bahrain region disrupted again on March 23 and April 2. We covered where that ended: AWS can't restore Bahrain and a UAE zone, six months after the drones. Gulf states responded by redrawing their build maps, as we wrote in Iran war forces UAE to redraw its 5GW AI data center map. Yandex is the third data point in the same series, and the first where training hardware, not just serving capacity, sat inside the blast radius. The anatomy of what was hit is the same building type we walked through in AI 101 — What is an AI data center?.

What the skeptics say

The contrarian case starts with the missing damage report. Nobody has confirmed that a single GPU was destroyed. Yandex describes halted operations and damaged buildings; the supercomputers may be intact behind dead power and networking, in which case the training pipeline is throttled rather than gone and the six-month repair estimate is about facilities, not compute. Treat "a national frontier lab lost its compute" as unproven until the company or satellite imagery says otherwise.

Second, Yandex is close to a worst case by construction: one dominant provider, six large nodes, an unusually concentrated national market, and a legal regime that forbids the obvious fix. Most hyperscale operators deliberately spread workloads, and the ECB-style assumption that a single strike removes half a company's capacity is not the design point of a multi-region cloud.

Third, Russia's own numbers were never that impressive on the global curve. Yandex's three machines total roughly 70 petaflops. That is a serious national asset and a rounding error next to the gigawatt campuses now going up — which is precisely why the strikes worked economically: cheap drones versus nine-figure fixed sites is a ratio no defender wins. The counterargument is that Russia is the outlier, that nobody will shoot at Abilene, and that hardening every campus for a threat that has materialized in exactly two theaters is expensive theater. The Gulf and Ryazan are the answer to that until they are not.

What to watch

The first test is YandexGPT itself. If the training pipeline is genuinely idled, expect silence on model releases for months; if the supercomputers survived, expect a demonstration of capacity sooner than the repair estimate suggests. A single published training run would settle more than any statement.

Second, watch where Russian compute goes. Filatov's "beyond the Urals" advice, containerized capacity, netting and detectors are all symptoms of the same decision — smaller, distributed, less efficient, closer to air defense. That is what resilience costs when you buy it after the first strike.

Third, watch the West's paperwork. Counter-drone radar is now a federal procurement line item in the United States, but the operating rules still stop short of letting a data center defend its own airspace. The regulatory question — who can jam what, above whose fence — is the bottleneck, not the sensor. And watch the insurers, who have to price a risk class that just moved from theoretical to demonstrated twice in seven months.

If a cheap drone can idle a national AI program, should physical security be a condition of siting permission for every new gigawatt campus? Tell us in the comments.