Core theses
What we believe, on what evidence, and what would change our mind.
7 essays, each in the same structure: the claim, the mechanism, the evidence, the strongest counterargument, what would change our mind, what we do as a result, and the open questions. Judgments change with evidence, and the changes are logged. Updated 2026-11-02.
01
Applications and open-source chokepoints hold AI's value; the model layer cannot
With the best US and Chinese models 2.7% apart and a model swap costing one endpoint change, this cycle's AI value lands in application ARR and in the open-source chokepoints every model passes through, and the model layer cannot hold it.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →
02
The bubble sits in the middle layer: 7-to-10-year debt on 2-to-3-year GPUs
The AI bubble sits in the middle layer that must keep refinancing: debt-funded compute lessors buy GPUs whose economic life may be 2 to 3 years with 7-to-10-year debt, and the losses land on creditors.
8 pieces of evidence · 3 counterarguments · 4 falsification tests →
03
The cost curve has a new author: Chinese open models at 1/40th the price
The cost curve's author has changed from US closed labs to Chinese open models: GLM-5.3-Flash is priced at about 1/40th of Claude Opus 4.8, and value is moving to applications and open-source chokepoints.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →
04
Verifiability times feedback cycle sets the order; each window lasts 12 to 18 months
The order in which AI remakes industries equals verifiability times feedback cycle: code checks itself in seconds and changes first; medicine, finance and manufacturing get feedback in months and change last; and each sector gets a 12-to-18-month window after capability crosses its threshold.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →
05
Models are replaceable parts; app profits rest on four moats and outcome pricing
Models are replaceable parts, so an application company keeps its profit through four moats and outcome pricing and proves it on three screens: distribution, retention, and gross margin that climbs as token prices fall.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →
06
14x apart in capital, 2.7% in capability: the US and China in one cycle
The US and China sit at two positions in one AI cycle: the US late in its mania phase with about 14x the capital, China early to mid, only about 2.7% behind on models yet valued at about a third of US peers, a discount that founders who bill in dollars are closing themselves.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →
07
Embodied AI: 2026 is the hardware year, and 80 points equals zero
2026 is the hardware year for humanoid robots and the intelligence year has not arrived: a robot that completes eight tenths of a task has completed nothing, so 80 points equals zero, and value lands in dexterous hands and data while the body turns into a commodity.
7 pieces of evidence · 4 counterarguments · 4 falsification tests →