The Physical Layer
Fabs, process nodes, packaging, HBM, equipment - why only a handful of companies can make any of it.
Three things I write about
The Bottleneck Watch
Power, cooling, materials, capacity. What is actually constraining AI buildouts, and what breaks next.
Physical AI Reality Check
What a robotics demo does and does not prove, and what would have to become true for it to ship.
All three end in the same place: what it means for margins, capex, and the companies you may own.
Recent writing
All posts →- Physical Layer
The 35% chip-fab tax credit has a start-by date, not a finish-by date. It is December 31.
A 35% US chip-fab tax credit applies only if construction begins by December 31, 2026. What counts as a start, and what missing it costs.
- Physical AI
Anthropic's hardware standard makes the wiring cheap. Someone still has to write the limits.
Partners report that Model Hardware Standard cuts device integration from weeks to hours. The part it cannot automate is the knowledge of how a machine may safely move, and the standard hands that to the people who use the hardware.
- Bottleneck Watch
The efficiency number improved. The water number in the same report did not.
Berkeley Lab projects average US data center PUE falling to 1.15-1.35 by 2028 while average site WUE rises to 0.45-0.48 L/kWh, and names the shift to hyperscale and colocation as a cause of both. Most of the water is not on the site meter at all.
- Physical AI
Nobody scraped the robot data. Fifty people performed it.
Language models train on text that already existed. The dataset robotics calls large-scale and in-the-wild is 350 hours, produced by 50 people over a year.