What's this about?
Dylan Patel of SemiAnalysis argues that Anthropic and OpenAI are on track to control most of the world's usable computing power within a few years, for a simple economic reason: they monetise compute better than anyone else, so they can outbid everyone for it. The episode also covers a shift from inference back toward training as recursive self-improvement approaches, and questions the more than $10 trillion of announced infrastructure.
What exactly does "outbid everyone" mean here?
Compute is a global auction with finite supply. Whoever converts a GPU-hour into the most revenue can pay the most for the next one. If two labs are meaningfully better at that conversion than everyone else, concentration is not a strategy they chose — it is the arithmetic of the auction.
What's the community saying?
This is a long-form interview rather than a discussion thread, so there is no comment signal. The claim worth carrying forward is the mechanism: concentration follows from monetisation efficiency, not from any deliberate land grab. Note this is one analyst's forecast, not a reported fact.