I initially thought the capital intensity of frontier AI development meant that it's impossible to catch up with AI labs. But with a bunch of near-frontier model releases recently (some from China), the market is no longer an Anthropic / OpenAI duopoly.
I think this creates two scenarios:
**One lab achieves recursive self-improvement.** It develops a compounding capability advantage that captures most of the market.
**No lab gets to recursive self-improvement.** The frontier remains a fluid oligopoly, token prices continue falling, and labs see significant margin compression.
In the second scenario, labs have to compete via conventional moats: distribution, workflows, proprietary data, cost, access to specific geos, etc.
Where this matters for us (public-market investors), is that it's actually fine to not have access to private companies because all of these labs need compute (the bottleneck suppliers give us another way to capture this AI beta). But, the easy version of the AI bottleneck trade has ended.
In this leg of the AI trade, we need to do way more analysis on ultimate market growth & the ability for the bottleneck suppliers to maintain pricing power. Given the current fluctuations in the market, there are also much more tactical / short-term signals we need to track (e.g. CXMT / YMTC fab ramps and model efficiency improvements if we are looking at the memory layer).
I wrote a longer version with the lab-by-lab breakdown and investment framework here: [https://eastwind.substack.com/p/the-warring-states-period-frontier](https://eastwind.substack.com/p/the-warring-states-period-frontier)
