Martin Alderson · 2026-07-06 · notable
Martin Alderson — GLM 5.2 and the coming AI margin collapse
Martin Alderson argues GLM 5.2 is the first open-weights model whose quality per dollar breaks the ~90% gross margin frontier labs earn on inference, at ~$4.40 per million tokens versus Opus and GPT-5.5 rack prices.
GLM 5.2 is the first open-weights model whose price-per-token cracks the ~90% inference margin propping up frontier labs.
What is it?
Martin Alderson's essay is Part 1 of a two-part argument that GLM 5.2 changes the economics of hosted LLM inference. He treats GLM 5.2 as the first credible open-weights alternative to Claude Opus and GPT-5.5 for real workflows, priced around $4.40 per million tokens — less than 20% of Opus's and roughly 15% of GPT-5.5's rack rate.
How does it work?
Alderson estimates hosted-inference gross margins at frontier labs sit near 90% over raw compute, because closed frontier models had no like-for-like open competitor. GLM 5.2 removes the moat: it clears his bar for 'more than 50% cheaper for nearly all workflows', it exposes the same API surface, and inference on AMD chips runs at roughly 2.75x lower cost per token than Nvidia Blackwell, further widening the pricing gap.
Why does it matter?
If Alderson is right, the ~90% inference margin that funds Anthropic and OpenAI is about to compress toward commodity — hosted GLM-5.2-style tenants will underprice frontier APIs while matching most of their quality, and labs will either lean on training subsidies (DeepSeek V3 reportedly under $6M) or pivot up-stack. Part 2 promises the follow-on for training economics.
Who is it for?
Founders and infra teams picking a model provider, investors modeling frontier-lab economics, open-weights operators.
Try it
Read Part 1 at martinalderson.com — Part 2 (training economics) is trailed at the end of the post.