GLM 5.2 and the AI Inference Margin Collapse
Original: GLM 5.2 and the coming AI margin collapse
Why This Matters
Open-weights models at frontier quality threaten the high-margin inference business model of leading AI labs.
Z.ai's GLM 5.2 is identified as the first open-weights model competitive with Anthropic's Claude Opus and OpenAI's GPT-5.5. Analyst Martin Alderson argues this signals a coming collapse in AI inference margins, where frontier labs currently earn roughly 90% gross margin on compute costs.
Martin Alderson, writing on his personal blog, argues that the real disruption in AI economics is not model training costs — a fixed, one-time expenditure — but inference, which scales with demand and carries genuine marginal costs. He estimates that frontier labs like Anthropic and OpenAI currently operate at approximately 90% gross margin on compute when charging $25 per million tokens, citing OpenAI's leaked financials showing ~60% gross margin on total revenue as a lower-bound reference. Alderson identifies Z.ai's GLM 5.2 as a meaningful milestone: the first open-weights model he considers a genuine competitor to Claude Opus and GPT-5.5. After several weeks of testing, he found quality near-indistinguishable from Claude Opus, his primary daily model. However, GLM 5.2 has notable weaknesses: it is slow due to extended chain-of-thought reasoning, lacks vision/multimodal support, and offers poor or absent web search capabilities. Alderson notes these gaps are significant for agentic workflows, where web search and image parsing are now routine. He flags third-party web search APIs as a high-potential area given this gap. The post is part one of a two-part series examining how open-weights models reaching frontier quality could structurally compress inference margins across the industry.