Dwarkesh x Dylan Patel: Anthropic and OpenAI on track to control most usable FLOPs

scaling01 · x · 2026-08-25

Dwarkesh Patel chats with Dylan Patel (SemiAnalysis) about AI lab economics over the next few years.

Key points:

As RSI draws near, the shift moves from inference to training; Anthropic and OpenAI are on track to control most of the world's usable FLOPs within a few years, because they monetize compute better and can outbid everyone.
They debate whether the >$10T of total AI capex by decade-end could trigger a sovereign debt crisis — hyperscaler debt raising interest rates, driving non-AI-exposed countries into bankruptcy and crashing non-AI equities.
Unresolved: whether anything can counter the forces barrelling toward centralization — training economies of scale, compute scarcity, and eventually continual learning and RSI.

More from AGI Musings