Why smarter AI models could drive up compute prices 10x

Summary

Dwarkesh Patel argues that AI lab revenue is compounding far faster than compute supply, making sharply higher compute prices likely. He examines whether higher lab margins, higher compute prices, or more inference spending closes the gap and concludes compute pricing is the main escape valve. He also identifies specific supply bottlenecks at ASML and TSMC and warns that weaker AI applications may get priced out.

• Anthropic revenue has 10x'd for three consecutive years while lab compute only 3x's yearly.
• The revenue-compute gap can be closed by higher lab margins, higher compute prices, or a bigger inference share.
• Spot compute is up over 40% since February and frontier secured GPU rentals are around 2x spot.
• As AI models become human-level, equivalent compute could monetize over 15x current H100 rental rates.
• Compute supply growth depends on Moore's Law, new fabs, and wafer reallocation, all of which appear constrained.
• ASML EUV machine production is a bottleneck for fab buildout through 2030 and possibly beyond.
• TSMC's leading-edge AI wafer allocation is approaching saturation at N3 nodes.