- DeepSeek's R1 caused market panic but misread AI economics: training is a fixed cost, while inference scales with demand and carries high margins.
- Frontier labs like Anthropic/OpenAI spend heavily on training then profit from inference, with gross margins estimated at ~90% on compute cost.
- GLM 5.2 from Z.ai is a genuine open weights competitor to Opus and GPT, offering similar quality but with weaknesses: slowness, no vision support, and poor web search.
- Switching costs are very low due to OpenAI/Anthropic compatible endpoints, making migration trivial compared to traditional vendor lock-in.
- Cost savings are significant: GLM 5.2 costs ~$4.40/MTok (less than 20% of Opus price), with expected further reductions as serving stacks optimize.
- Data privacy concerns with Z.ai's terms exist, but open weights allow on-premise deployment or alternative providers with better contractual protections.
- The article hints at a coming inference margin collapse that could reshape the AI industry, referencing Bezos's 'your margin is my opportunity'.