What happened in this episode?
Dwarkesh Patel delivers a short, numbers-heavy monologue arguing that AI labs’ revenues could outpace their ability to scale compute, implying higher margins, higher compute prices, or a larger share of compute devoted to inference rather than training. He asserts that Anthropic’s revenue has grown about 10x year-on-year for three years and speculates about a continuation of that trend, while claiming lab compute grows closer to 3x annually. Patel cites reported improvements in inference margins and an increase in spot compute pricing since an earlier-year trough, and says inference’s share of total compute has likely risen since 2024.
He argues that if frontier models become more economically valuable per unit of compute, compute prices could rise sharply; he offers a thought experiment that a human-level software-engineering agent on an H100-class GPU could justify far higher annual rental prices than current spot rates. Patel also claims that large buyers may already be paying substantial premiums to secure capacity at scale, giving a purported example of Google’s monthly GPU spend.
Patel closes by contending that compute supply is less elastic than commodity supply, pointing to limits from semiconductor tooling (including EUV machines) and to a finite ability to reallocate leading-edge wafer capacity from smartphones/PCs to AI. He ends by expressing concern that strong economies of scale in ‘intelligence’ could concentrate power.
