Keep pulling the thread on Gavin Baker.
The overwhelming majority of economic value created at the AI model layer has been captured by frontier models.
The current AI model "Pareto frontier" for intelligence versus cost is dominated by Anthropic, OpenAI, and xAI's Grok 4.3.
A violation of Richard Sutton's "bitter lesson" — the idea that more compute and data will always outperform human algorithmic ingenuity — is the biggest risk to the AI investment thesis.
The AI model industry is shifting from fixed-price, "all-you-can-eat" plans to usage-based pricing models.
Gavin Baker predicts that the shift to usage-based pricing will enable OpenAI and Anthropic to exceed $200 billion in Annual Recurring Revenue (ARR) this year.
The disaggregation of prefill and decode workloads in AI inference will extend the useful life of GPUs to 10 or 15 years.
The profits in the AI stack are currently accruing to the energy, data center, chip, and model layers, but not to the application layer.
Gavin Baker believes that whenever Jensen Huang chooses, NVIDIA can develop its own model that is "pretty close to the frontier" of AI capabilities.
A new "prisoner's dilemma" exists for frontier AI labs: if they all withhold API access, they can cut off Chinese open-source models, but if one company defects, it will gain a significant revenue and intelligence advantage.
Microsoft CEO Satya Nadella is making the strategic decision to use the company's internal compute to improve its own products like Copilot, rather than selling that compute to others, even though it hurts short-term revenue.
AI has so far been a net destroyer of value at the application layer, with trillions of dollars of value destroyed.
Gavin Baker asserts that Ukraine is winning its war against Russia primarily because it possesses the best battlefield AI outside of the United States and Israel.