Summary
Dylan Patel, founder of semiconductor and AI infrastructure research firm SemiAnalysis, joins Dwarkesh Patel for a detailed conversation on the economics driving the AI industry in 2026 and beyond. Patel documents the dramatic shift in lab financials: Anthropic turned profitable in Q2 2026, with revenue per megawatt reaching approximately $50 million — far above the $10–15 million infrastructure cost that previously generated negative gross margins. OpenAI is on a similar trajectory, with GPT-5.6 and Codex driving revenue growth that could push it to profitability in Q3.
The conversation covers compute centralization in granular detail. OpenAI and Anthropic currently absorb roughly 30% of all new global compute capacity, with that share projected to reach 40–50% by end of 2027 based on already-signed contracts. SpaceX is identified as a major new compute supplier likely to lease heavily to both labs. Global AI CapEx is projected to exceed $1 trillion in 2026 and more than $2 trillion by 2028. Patel makes a non-consensus argument: as revenue per megawatt rises, labs will increasingly reallocate compute from inference to training — because the marginal value of accelerating AGI development outweighs near-term inference revenue and shareholder returns.
The episode closes with broader implications for economic power concentration and what it means when two private labs control an increasing share of the world’s compute — and by extension, a growing fraction of automated labor.
📺 Source: Dwarkesh Patel · Published August 25, 2026
🏷️ Format: Podcast
