GLM 5.2 and the coming AI margin collapse
6 hours ago
- #inference costs
- #AI economics
- #open-source models
- DeepSeek's R1 model caused a market reaction due to low training costs, but AI economics are more complex.
- Training costs are fixed, while inference costs scale with demand and have high gross margins for frontier labs.
- GLM 5.2 is a strong open-weights competitor to models like Opus and GPT, with near-equivalent quality.
- Drawbacks of GLM 5.2 include slower speed for interactive use, lack of vision support, and poor web search capabilities.
- Migration to open-weights models like GLM 5.2 is easy due to OpenAI/Anthropic-compatible endpoints, with low switching costs.
- Cost savings are significant: GLM 5.2 is priced at ~$4.40/MTok, less than 20% of Opus, and may become even cheaper with optimizations.
- Enterprise concerns about data privacy with providers like Z.ai can be mitigated by using other providers or on-premises hosting.
- Inference margin collapse could reshape the AI industry, benefiting those who leverage lower costs.