Gavin Baker with Patrick O'Shaughnessy provides one of the best analyses of what is currently happening with the AI economy.
Baker is an investor, an early investor in Cerebras, and focuses on national security investing. His perspective on numbers and market dynamics is a rare combination of depth and specificity.
Here are a few theses that seemed the most non-obvious:
- Anthropic added $11B ARR in just one month. This is more than what Palantir, Snowflake, and Databricks built together over 10 years. The current ARR is around $50B, with a 1000% year-over-year growth. Moreover, Anthropic is compute-constrained - if there were enough hardware, the ARR would already be $100-200B.
- Anthropic burned 80% less capital than OpenAI to reach comparable revenue. And the token value is lower. OpenAI compensates for this with aggressive compute acquisition.
- Disaggregating the prefill and decode phases of inference can extend the life of GPUs from 3-4 years to 10-15. According to Baker, this alone could save private credit - the cost of financing GPU clusters would drop from 7%+ to 5-6%.
- The main indicator of whether a bubble is inflating is TSMC's discipline regarding capacity. Nvidia could sell $2-3T worth of GPUs in 2026-27 if TSMC agreed to everything Jensen asks for. TSMC's decisions on capacity expansion are cleaner than any other signal.
- Frontier labs are capturing almost all the economy at the model level. At the application level, on the contrary: trillions of dollars of value have been destroyed by AI, and Cursor and Cognition are no exceptions.
And separately - a metaphor that Baker repeats: AI is a machine gun, while most are still fencing. Even a master swordsman loses to a peasant with a machine gun. We need to master new weapons, not refine the old ones.
