The trillion-dollar AI arms race just discovered it has a $500 drone problem.
The Summary
- Ukrainian drones struck three Yandex data centers last week, taking down facilities that housed AI supercomputers and disrupted services for banks and retailers across Russia.
- Data centers are massive, power-hungry, heat-generating targets that can't be hidden, and the infrastructure powering military AI systems is particularly vulnerable.
- As nations pour billions into AI compute for defense applications, the physical security gap creates asymmetric attack opportunities that favor cheap drones over expensive infrastructure.
The Signal
Microsoft, Google, and Meta are racing to spend $50 billion each on AI infrastructure this year. OpenAI is plotting data centers that will consume gigawatts of power. Oracle's Larry Ellison talks about facilities so large they'll need their own power plants. Meanwhile, a $500 commercial drone just proved it can take the whole thing offline.
The strikes on Yandex facilities hit one site in Sasovo that housed tens of thousands of servers and two supercomputers training AI models. Services went dark for banks, retailers, everyday users. Russia had already hit Ukraine's Parkovyi Data Center so hard the operators said it couldn't be restored. This is the new math of infrastructure warfare, and it doesn't favor the side spending billions.
"If the Russians or the Chinese were using AI to accelerate their operations, I would definitely want to target those data centers."
Jeffrey Lewis at the Middlebury Institute points to AI inference as the critical vulnerability. Training runs can be distributed, interrupted, restarted. But inference, the real-time deployment of trained models, that's different. When a military uses AI to analyze satellite imagery, recommend targeting decisions, or coordinate drone swarms, those models need to run now. Take out the data center running inference, and you've just blinded or paralyzed an entire operational system.
The physical constraints make hiding impossible:
- Data centers need massive power infrastructure that shows up on satellite imagery
- Cooling systems generate heat signatures visible from space
- Network connectivity requirements create traceable digital footprints
Ukraine's military is already running AI systems that help drones navigate without GPS, analyze battlefield intelligence, and recommend commands to officers. All of that requires computing power, and that power lives in buildings full of servers. The U.S. and China are further along, deploying AI for everything from missile defense to logistics optimization. Every one of those applications has a data center dependency.
The economics are brutal. A modern AI data center costs $10 billion to build and millions per day to operate. A military-grade drone costs thousands. The attacker's advantage is so lopsided it makes Cold War mutually assured destruction look balanced. You can harden a facility, bury it, surround it with air defenses. But you can't hide the power lines, the cooling towers, the fiber optic cables, or the fact that it needs to sit somewhere accessible enough to staff and maintain.
The Implication
The companies building Web4 are building on infrastructure that assumes peacetime. Every agent platform, every AI model API, every automated system depends on compute that lives in a fixed location. As AI moves into military and critical infrastructure applications, that assumption breaks. Watch for three shifts: decentralized inference architectures that distribute computing across smaller, harder-to-target nodes; edge computing that pushes AI processing closer to where it's used; and a fundamental repricing of risk in data center real estate.
If you're building on centralized AI infrastructure, you're building on a target. The agents may be intelligent, but their homes are sitting ducks.