Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028
Summary
Dwarkesh Patel interviews Dylan Patel of Semi Analysis on the rapid centralization of AI compute, where labs like OpenAI and Anthropic are capturing most new capacity. They discuss how these labs are transitioning from venture-funded losses to massive profitability, driving up compute prices and reshaping the supply chain. The conversation covers the widening gap between US and Chinese compute capabilities, regulatory impacts on model releases, and the strategic shift toward prioritizing training over inference to accelerate AGI development.
Topics discussed
Intro: AI CapEx ballooning to $2T and lab profitability
Centralization of compute: Labs vs. the world
Compute supply constraints and arbitrage opportunities
Revenue per megawatt and the race for 100GW
Value capture: Why labs dominate over app layer
Compute pricing dynamics and supply chain bottlenecks
Timeline for Recursive Self-Improvement (RSI)
Training vs. Inference compute allocation shifts
China's AI compute growth and domestic chip production
Global AI CapEx reaching $10T by decade's end
Sponsor: Antithesis deterministic testing platform
Sovereign debt crisis risks from AI infrastructure
Interest rates, credit markets, and equity valuations
Regulatory slowdowns and government intervention
Sponsor: Jane Street ML internship opportunities
Future of labor, centralization, and AGI alignment
Listen ad-free on Castria