Wall Street set up for record highs Monday as, in Reuters' phrasing, the "AI push eclipses US-Iran war worries." Same morning, the front page of Hacker News: a 10-year-old Xeon is all you need to run Gemma 4, and a 4-billion-parameter image model now fits on a local device. Two stories about the same technology. They point in opposite directions.
The capital markets are pricing scarcity at the exact moment the technology is proving abundance. AI capex is being floated on corporate debt now — Reuters reports AI-linked bond sales are reshaping global credit markets — on the premise that compute is the moat and the bill runs to the trillions. The engineers keep shipping evidence that the bill is collapsing.
The capital markets are pricing scarcity at the exact moment the technology is proving abundance.
Look at what "a 10-year-old Xeon is all you need" actually claims. Not that frontier training got cheap — it didn't. But the thing most users want, capable inference, is sliding down the hardware stack fast enough to land on machines people already wrote off. A 1-bit image generator running on-device. Gemma 4 on a 2016 CPU. The marginal cost of a good-enough model is falling toward the price of the electricity to run it. That is the opposite of a moat.
The bull case has a serious answer: demand scales faster than efficiency. Every drop in the cost of inference unlocks more of it — agents that run all day, models called in loops, workloads nobody would have paid for at the old price. Cheaper compute doesn't shrink the market; it floods it. Jevons, not deflation. Maybe. But the Jevons paradox makes the application layer valuable — not the balance sheets that borrowed billions betting the compute itself stays scarce.
One of these bets is wrong. Either the moat holds and the Xeon is a toy, or the Xeon is the tell and a great deal of debt was raised against a commodity. The market is buying the first story. The front page of Hacker News keeps quietly publishing the second.