A second edition. Capability can keep advancing while the companies financing it fail, but the September post looked for the stress in the wrong layer. This version says where it sits, what would show it, and what we withdraw.

The old scaling recipe turned more high-quality human text into lower loss. That trade is still real, but the easy part is ending: future gains depend less on raw token count and more on data quality, synthetic-data mixtures, and multimodal transfer. A modeled look at the regime change.

The AI buildout is not a simple bubble story. It is a race between capability, demand breadth, and the delayed arrival of compute supply.

The most AI-exposed sectors show no productivity breakout. What early evidence there is points, tentatively, toward large operational labor pools — and it softened on the latest revisions. Causality remains unproven.