AI capex deep-dive: user penetration at 50%, per-capita token use still 100x from ceiling

dotey · x · 2026-08-29

A long series projecting the AI semiconductor endgame. Core thesis: internet infrastructure had only one exponential dimension (user penetration), while AI hardware demand is the product of user penetration × per-capita usage, both convertible into tokens — hence a longer buildout cycle.

User penetration hit 50% months ago, near the sigmoid inflection; but per-capita token usage is still early: median US enterprise AI spending is $12/person/month, with a plausible ceiling near 10% of white-collar wages ($1000/month) — nearly two orders of magnitude of headroom.
On memory: HBM satisfies 2.5 of 3 conditions for escaping commodity cyclicality (customization, structural exponential demand, rapid tech iteration). Samsung's failed NVIDIA qualification crashed its share from 60% to 20%, yet the capacity was redirected to Google TPU and AMD — showing HBM is still largely standardized.
Also covers the next ARR growth driver after coding and the open- vs closed-source trajectory.

Original post →

AI capex deep-dive: user penetration at 50%, per-capita token use still 100x from ceiling

More from Infra