This is a contrarian take: instead of open source models running on private infrastructure being the dominant mode of AI, they instead argue that the returns to scale from cacheing intermediate reasoning steps could provide enough of a performance advantage to centralized models.