Why AI Compute Costs May Rise 10x
Dwarkesh Patelgo watch the original →
the gist
Frontier AI labs face a widening gap between 10x revenue growth and 3x compute capacity growth, likely forcing a massive increase in compute prices as models become more efficient at monetizing hardware.
The Compute-Revenue Gap
Frontier AI labs are currently experiencing a structural mismatch: revenue is growing at 10x annually, while available compute capacity is only scaling at 3x. To bridge this gap, labs are already increasing inference margins (reportedly from 40% to 80% for some top models), shifting compute allocation from training to inference, and paying premiums for hardware. As AI models reach human-level capabilities, their ability to monetize compute will likely drive market prices significantly higher, potentially reaching 15x current spot prices if an H100-equivalent model performs the work of a professional software engineer.
The Alchian-Allen Effect and Scarcity
The increasing cost of compute will trigger the Alchian-Allen effect, where the relative cost of using inefficient models becomes prohibitive. Labs that train the most efficient models will capture higher margins because they maximize the utility of expensive, scarce tokens. This dynamic creates a barrier to entry, as smaller competitors cannot outbid frontier labs that generate higher revenue per unit of compute. The supply of compute remains inelastic due to physical bottlenecks, including the slow pace of building new ASML EUV machines and the saturation of leading-edge wafer capacity at TSMC, which is projected to reach 86% allocation for AI by the end of next year.