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 among labs like OpenAI and Anthropic. They discuss how these firms are shifting from venture-funded losses to massive profitability, enabling them to capture half the world’s incremental compute by 2027. The conversation covers supply chain bottlenecks, rising compute prices, regulatory impacts on model releases, and China’s lagging but accelerating domestic chip production.
Topics discussed
Intro: AI CapEx ballooning to $2T and labs turning profitable
Compute centralization: Labs vs. the world and SpaceX entry
Global compute growth projections and ROI on Fab CapEx
Arbitrage opportunities and revenue per megawatt predictions
Regulatory impacts, value capture, and supply chain dynamics
Sponsor: Using GrokBot to hire a podcast editor
Revenue forecasts, RSI, and internal vs. external model use
US vs. China compute deployment and semiconductor subsidies
Training vs. inference split and $10T CapEx forecast
Sponsor: Antithesis deterministic software testing
Sovereign debt crisis risks and crowding out effects
Interest rate hikes and economic doubling scenarios
Political constraints on AI speed and slow takeoff hopes
Sponsor: Jane Street LLM research for financial markets
Future centralization of AI labor and economic concentration
Listen ad-free on Castria