Writing
Semiconductors, AI infrastructure, and robotics - sourced from public material and reasoned through from the engineering constraints up.
- 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.
- Physical AI
The battery lasts two hours. No one publishes who changes it.
Charging a humanoid takes less time than draining it, so energy is not the ceiling. Three battery swaps a shift are, and neither maker says who does them.
- Physical AI
A robot hand's warranty is written in cycles, not years
Gripper makers rate hands in cycles, not years. So throughput and wear are one variable, and the case for faster robots is a case for faster wear.
- Bottleneck Watch
Why a hot robot joint needs a Chinese export licence
China's control list names terbium- and dysprosium-bearing magnets, not plain NdFeB. What decides which one a robot needs is heat in the joint.
- Bottleneck Watch
The AI bottleneck isn't chips. It's the grid connection.
PJM has cleared its capacity auction at the price ceiling three years running and still bought less power than the region needs. Two queues explain why.
- Physical AI
What robot operator job ads tell you that demos don't
Deployment is running ahead of autonomy and people fill the gap. The operator-to-robot ratio is a restatement of how often the robot needs help.
- Physical AI
What a 95% benchmark score actually means on a factory floor
A policy at 95% per step finishes a twenty-step task a third of the time. The best entry in a 2025 robotics benchmark completed 12% of its tasks.
- Physical AI
The three numbers humanoid demos never give you
Industrial buyers decide on cycle time, availability, and MTBF. Humanoid demos publish none of them - and the semiconductor industry has had a standard for two of the three since 1986.