Dwarkesh says frontier-model monetization could drive H100 rents far above today’s prices

inductionheads · x · 2026-07-29

Dwarkesh Patel argues that if frontier labs keep scaling revenue while compute only grows about 3× a year, compute prices may have to rise sharply.

He uses Anthropic’s reported revenue growth, margin expansion, and the increasing share of spend going to inference to argue that the current pricing regime may not hold.
The core claim is that a human-level software engineer running on H100-class hardware could justify more than $250k/year in rent, which would imply compute being far more expensive than today’s spot prices.
The post frames this as an AI infrastructure economics problem: better models may simply monetize the same compute more efficiently, pushing up the value of scarce accelerators.

More from Venture