The Nine-Month Doubling

The number that matters this year is not a benchmark. It is a doubling time. AI chip deployment is doubling roughly every nine months, according to Epoch AI's analysis reported this month [1] — a rate faster than Moore's law at its peak, achieved not by shrinking transistors but by pouring concrete.

Compute used to be a research input. It is now an asset class with a construction pipeline, a labor market — 4.7 million projected construction and operations jobs through 2030 [2] — and a power bill that utilities plan decades around. When an input doubles every nine months, everything priced against it reprices too: electricity futures, GPU depreciation schedules, and the valuation of any company whose moat is 'we have compute.'

What a doubling buys

The July release wave shows where the capacity goes. Six labs shipped frontier-class models within six weeks of each other [3]; million-token context windows, priced at $5 per million input tokens at the top end [4], are becoming table stakes. Capability gains that once took a year now arrive quarterly — not because ideas got faster, but because the queue for training runs got shorter.

The bear case is equally physical. A build-out this fast is a bet that demand doubles on the same schedule. If agentic adoption disappoints by even a year, the industry will have built the fiber-optic glut of the 2020s — capacity whose price collapses precisely because it exists. Whether that is a catastrophe or a subsidy for the next decade of startups depends on where you sit.