Dylan Patel of SemiAnalysis argues on the Dwarkesh Podcast that compute centralization is not a forecast but a signed contract: OpenAI and Anthropic are on track to control most of the world's usable FLOPs within a few years, because they can monetize compute better than anyone else and therefore outbid everyone else.
The numbers he gives:
That last ratio is the whole argument. GPT-4 on Hopper generated negative gross margin; GPT-5.6, Opus 5 and Fable 5 clear the incremental cost several times over. Patel says Anthropic turned a profit in Q2 and OpenAI may have in Q3, and that the labs are now recycling inference profit straight into training — a shift from serving to R&D that he ties to RSI drawing nearer.
The builders churn even as the tenants don't. SpaceX enters next year as a major compute developer and will likely lease much of it to the two labs, being the bidders with the highest marginal willingness to pay. Meanwhile OpenAI is moving to its own chips and Anthropic is buying TPUs from Google, deployed with .
The second half is the darker one. The two put total AI capex above $10T by the end of the decade and ask whether hyperscaler debt issuance at that scale pushes up interest rates enough to bankrupt non-AI-exposed countries and crash non-AI equities. Patel also puts China under 10% of new compute, while arguing its labs need less of it.
The question they explicitly fail to resolve: whether anything counters centralization at all, given economies of scale in training, compute scarcity, and — eventually — continual learning and RSI. Treat the trajectory as one well-sourced analyst's projection, not a settled fact.
Sources: Dwarkesh Podcast