GLM 5.2 and the Coming AI Margin Collapse

GLM 5.2 and the coming AI margin collapse

I argue that the real cost of AI lies in inference, not training, and GLM 5.2 is the first open weights model to genuinely rival Opus and GPT. With costs under 20% of frontier labs and trivial migration via OpenAI-compatible endpoints, this shift threatens to collapse AI margins. While it lacks vision and web search, its price advantage makes it a dangerous drop-in replacement for many agentic workflows.

"Training is a fixed, up-front cost, but inference scales with your demand and has genuine marginal costs."