Summary
Dwarkesh Patel examines a structural tension building inside the AI industry: frontier lab revenues are growing roughly 10x per year while compute supply is expanding at only 3x annually. Using Anthropic as a case study — with revenue potentially reaching $100–150 billion this year after ending 2025 at $9 billion — Patel works through the arithmetic of how that gap gets bridged.
Three escape valves emerge: rising lab margins (Anthropic’s inference margins reportedly jumped from 40% to over 80% in roughly a year), higher compute prices (spot GPU rates are already more than 40% above their February 2026 trough), and a growing share of compute dedicated to inference rather than training. Patel argues the margin path has limits — sustaining 90%+ margins on a commodity like intelligence is historically unusual — which leaves compute price inflation as the more likely outcome. His anchor example: a Google deal paying $900 million per month for 110,000 GPUs at roughly 2x market spot rates.
The analysis extends to a provocative valuation thought experiment: if a human-level software engineer could run on an H100-equivalent chip, that chip should rent for over $250,000 per year at current software engineer salaries — more than 15x today’s spot price. Patel engages seriously with the counterarguments, including the lump-of-labor fallacy and historical commodity-scarcity predictions, making this a useful framework piece for anyone tracking AI infrastructure economics.
📺 Source: Dwarkesh Patel · Published August 03, 2026
🏷️ Format: Opinion Editorial
