GLM 5.2 signals the inference-margin squeeze coming for frontier AI labs

The author argues that the market misread the DeepSeek scare: training a model is a large one-time fixed cost, but inference carries real per-token marginal costs and is where frontier labs make their money. His napkin math puts Anthropic and OpenAI gross margins on raw compute around 90%, meaning the whole business model hinges on amortizing training over highly profitable API calls. That cushion is the thing now under threat.

Z.ai’s GLM 5.2 is presented as the first open-weights model good enough to pass as a daily driver against Opus and GPT — nearly indistinguishable in the author’s coding work, if slower because it over-thinks, and hampered by two real gaps: no vision support and weak web search (worked around with a CLI search tool like ddgr). The migration story is what makes it dangerous to incumbents: providers like Z.ai and Fireworks expose OpenAI- and Anthropic-compatible endpoints, so swapping GLM into Claude Code or Codex is just a base URL and API key change, with none of the years-long lock-in of enterprise software. Enterprises wary of Z.ai’s China ties and loose data-retention terms can use other hosts or run the weights on-premises.

On price, GLM 5.2 runs around $4.40 per million tokens — under 20% of Opus and roughly 15% of GPT 5.5 — and even accounting for its heavier token use, the author expects most workflows to come out at least 50% cheaper at comparable quality. He also cites reports that serving it on AMD hardware is ~2.75x cheaper per token than Nvidia Blackwell, and expects costs to keep falling as serving stacks optimize. This is part one of a two-part series; part two will cover who wins and loses when inference margins compress, framed around Bezos’s line that “your margin is my opportunity.” (Disclosure: Fireworks gave the author free credit to test GLM.)