Keep pulling the thread on Yvonne Wassenaar.
Charles Adlesky estimates that approximately 70% to 80% of global AI compute is currently dedicated to proprietary models.
Together AI, a four-year-old company, has achieved a run rate of "several nine figures," implying hundreds of millions of dollars.
The total inference revenue generated by coding agent companies is estimated to be several hundred million dollars.
Charles Adlesky asserts that the top 50 AI-native model labs and agent companies currently consume more AI infrastructure than the Global 2000 combined.
AI-native companies typically begin to reconsider their foundational model strategy, shifting towards open-weight models, when they reach approximately $200 million in annual run rate.
Nader Broebrievyer predicts that token consumption will increase by 10x for both open source and closed source models this year.
NVIDIA has open-sourced its Nemotron 3 model, including its underlying data, weights, architecture, and training recipes.
Charles Adlesky predicts that the combination of open-weight models and open source agent harnesses will have a "Lampstack moment," leading to a proliferation of AI applications.
The entire AI industry, both open and proprietary, is palpably constrained by a lack of compute capacity.
Charles Adlesky believes that capital and energy are the fundamental limiters to the expansion of AI into the world, constraining both proprietary and open source development.
Charles Adlesky asserts that every major model provider, whether for closed-weight or open-weight models, uses an open source foundation for their inference engine.
Nader Broebrievyer's startup, Brev, was acquired by NVIDIA approximately a year and a half ago.