Why Smarter AI Models Could Drive Up Compute Prices 10x

Smarter AI models could make computing capacity more expensive if they let customers earn more from each hour of GPU use. That is the argument Dwarkesh Patel makes in the essay this video records. The essay starts from a reported trend: Anthropic's revenue has grown about tenfold a year, while the essay estimates that lab compute grows roughly threefold a year. If the revenue trend continued, something would have to give: lab margins rise, compute gets more expensive, or a larger share of compute goes to serving customers.

The essay argues that margins alone would have to become implausibly high and that labs prefer not to shift compute away from training, leaving rising compute prices as the likeliest outlet. The reasoning is that models able to do more valuable work can monetize the same hardware better, so their operators can outbid others for scarce capacity. The author frames this as conditional on continued capability gains and notes ways it could fail, such as falling marginal value of labor as supply grows. Original essay.