Open weights don’t need full parity with frontier labs. They only need what I call “pleb parity”: get close enough on high-value, high-volume task at a time, so 20 specialized plebs can divide the work and level the game against one elite.
When you have a gang of equally talented, frugal and relentless Chinese AI founders, who stays #
1# on the leaderboard matters less. What matters is that whenever one falls behind, another open-weight peer fills the gap very soon
My simple mental model for Anthropic/OpenAI’s future ARR is the frontier-call ratio: Across an end-to-end business outcome, what % of steps still MUST call the most expensive frontier model?
As plebs pushes that ratio down, pricing power and future ARR growth expectations get eroded. Think of the most elite revenue-generating team inside a company. How many seats truly need to be filled by Ivy/Stanford grads?
AI stacks will look the same: frontier intelligence for the key steps where it materially changes the outcome; open weights everywhere else.
The only domains I can think of with a real case for unlimited frontier-token budgets
1) Quant trading, mm and pod shops: RenTech, Jane Street, Point72 etc where raw intellectual horsepower can be translated directly into P&L
2) Ultra-high-stakes discovery: next-gen chips and materials, cancer drugs, etc. But even there, the upside/scale is limited by the even more scarce giga brain human minds to define the right problem and know where to tinker, not by token or compute
For the rest of the world, and most knowledge-work domains, pleb parity is probably the long-term equilibrium.
Long live the plebs. Kudos to every Chinese AI founder and engineer working against the odds to make it happen