Dwarkesh Patel argues that AI companies could see revenue grow faster than their available compute. That gap would have to resolve through higher margins, higher compute prices or a larger share of infrastructure devoted to inference rather than training.
More capable models may increase the value created by each unit of compute instead of reducing total demand. This would favor efficient models and well-capitalized incumbents, while lower-value applications could be priced out when capacity is scarce.
The physical bottlenecks include fabrication plants, advanced lithography equipment and wafer capacity, all of which expand slowly. The argument remains a scarcity scenario rather than a certainty, but it shows why better model efficiency does not automatically imply cheaper AI services.
Watch the original on YouTube