Today’s Sharp Thought: Beating Moore’s Law

Source: Nate’s Newsletter Date: 2024-11-09 URL: https://natesnewsletter.substack.com/p/todays-sharp-thought-beating-moores

Summary

The piece argues that AI capability improvement has decoupled from semiconductor lithography progression by shifting to multi-datacenter training at scale. Where Moore’s Law delivered roughly 2x improvement every two years, the author claims AI capabilities are doubling or tripling annually through coordinated large-scale training runs. The mechanism is architectural — concentrating specialized infrastructure across geographically distributed facilities for AI training — rather than chip-level transistor density gains.

Implications

• Feeds the Anthropic pricing restructure + capital thread: the $65B capital infusion (Google $10B + Amazon $5B, both including compute commitments) is the financial expression of this thesis — staying on the steep part of the capability curve requires continuous multi-datacenter infrastructure investment that compounds faster than traditional IT spending.
• Connects to token economics competition: if capability improves faster than the pace Moore’s Law would predict, pricing pressure compounds faster too — vendors who bet on stable capability curves are repriced by the market before their infrastructure pays off (the OpenAI subscription collapse signal).
• Background for TurboQuant and KV cache compression: the rapid capability gains at training time create inference-side pressure; techniques like KV cache compression are the local/edge response to a cloud-side improvement curve that consumer hardware cannot directly match.