The dominant story this week is the squeeze. The cost of producing AI capability and the price of selling it are both collapsing simultaneously, compressing margins and reshaping business models across the entire stack. A phone company trained a frontier model for $3M. Every lab cut prices by 40-50% in lockstep. And an essay appeared arguing that when capability is cheap, what becomes valuable is something AI can't give you: your own judgment.
Xiaomi — a consumer electronics company, not an AI lab — produced MiMo-V2.6-Pro (1T parameters, 42B active), now the top open-weights model in the world. Training cost: approximately $3M. A year ago, frontier AI required nine-figure compute clusters and the resources of three or four labs. That narrative is broken. The barrier to entry is collapsing while the barrier to differentiation stays high. If $3M gets you top-tier open weights, the moat around frontier labs narrows dramatically.
The six-month question: does this democratisation hold, or do the labs pull ahead again with capabilities that can't be cheaply replicated?
Every frontier lab cut API prices by 40-50% in the same week. Claude Opus 5.5 launched as the new default. OpenAI shipped more efficient GPT-6 variants. When everyone cuts by the same magnitude simultaneously, it's structural, not competitive — the cost of serving capability is falling faster than capability is rising.
The structural implication: margin compression on token sales is changing the labs' business model. Investor Nick Carter's "cynical read" — if selling tokens is a commodity business, labs are incentivised to hold the true frontier for internal IP capture, with "pacing the frontier" as convenient cover. The API you build on may be a tier below what the lab actually runs internally.
For someone trying to take control: if the API is a tier below the real frontier, you're building on last year's capability. The people who kept fine-grained model selection have a lever the squeeze is removing from everyone else.
Ethan Mollick argues that when AI capability is abundant and cheap, the bottleneck shifts from the model to the person. Deep knowledge, wide knowledge, taste, and agency become the capabilities that retain value. The "overhang" is the gap between what AI can do and what humans actually use it for.
The counter-narrative: the feeds are full of model releases and price wars. Mollick is pointing at the other side. The squeeze rewards expertise, not access. Capability is becoming a public good. Agency is the scarce resource.
The people who are ahead right now aren't the ones with the most expensive subscriptions. They're the ones with deep domain knowledge who can deploy cheap capability toward hard problems.
Bonds tied to a Jane Street-leased data center — issued at 8.9% in August — now trade at 11.3%. $18B of Oracle data center debt in New Mexico is quoted as low as 89¢ on the dollar. Stressed levels. The bond market is pricing risk into AI infrastructure before equity markets notice.
The AI infrastructure buildout has been treated as infinitely fundable. If the next tranche comes at materially higher rates, it constrains the buildout — and the entire AI revenue thesis depends on continuous infrastructure growth. The cost of producing capability is collapsing, but the cost of housing and serving it may be rising.
The throughline from Episode 6 to Episode 10: value is moving up the stack. From the model to the harness to the person.
The squeeze connects to all of them. The cost collapse validates the anticipation. The "cynical read" is a control-layer story. The feedback loop gets sharper when labs have incentives to keep the best models internal. The governance framework from Episode 9 looks different when the labs' business model is shifting from selling tokens to capturing IP.
One consistent trend: capability is diffusing, control is concentrating, and the gap between the two is where the future gets decided.