An open model at $4.40/MTok threatens frontier labs' 90% inference margins
Martin Alderson argues the frontier's real weakness is the cost of serving tokens, not training them, and that GLM 5.2 from Z.ai now lands near-Opus quality at a fraction of the price.
GLM 5.2 lists near $4.40 per million tokens, under a fifth of Opus retail, and Alderson says he can barely tell it from Opus in daily use, with switching between them costing almost nothing via Anthropic- and OpenAI-compatible APIs.
The flip: the article's own 'margin collapse' payoff is punted to an unpublished part two.
The frontier's soft spot is inference, not training
Martin Alderson's argument starts with a number: frontier labs run something like 90% gross margins on inference at roughly $25 per million tokens. The exposure, he writes, is the variable cost of serving tokens, not the one-time cost of training. Into that gap walks GLM 5.2 from Z.ai, which he calls the first real open-weights competitor, priced at about 15% of GPT 5.5. It burns more tokens per task, yet he estimates it comes out more than 50% cheaper for nearly every workflow. He points to AMD inference running 2.75x cheaper per token than Nvidia's Blackwell, and to switching costs kept near zero by OpenAI- and Anthropic-compatible endpoints. The weaknesses are real: it's slower, has no vision, and its web search is poor. The actual collapse argument is held back for a part two.
"I don't see any historical analogues"
The cleanest pushback asked for evidence the pattern has ever played out:
Compute costs collapsed since the advent of Cloud and yet hyperscalers still have fat margins… I understand the arguments for a margin collapse, but I don't see any historical analogues.
The part-two problem
Others noted the piece never delivers on its title:
This article only promises to get into 'the coming AI margin collapse' in a yet to be published 'part two'. This part only makes the point that GLM 5.2 is pretty good (no shit).
— a reader unimpressed by the setup
"The gap is huge" in practice
Benchmarks flattered GLM; daily use didn't. A developer who keeps subscriptions to nearly every provider:
the difference between glm5.2 and opus is not even close, the gap is huge. raw benchmarks glm is impressive, but in practice these models are lacking so much.
— a developer who pays for nearly every provider
Where the cost actually hides
On the economics, one voice relocated the whole question:
In agentic coding, cached input tokens is 90% of the API 'cost'. It doesn't require GPU compute, and DeepSeek has shown that it can be done 50~100x cheaper with MLA/CSA/HCA, and a whole bunch of disks.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
The Supreme Court limited location tracking. Does it reach plate-reader cameras?
On June 29, 2026, the Supreme Court ruled 6-3 in Chatrie v. United States that obtaining cell-phone location data is a Fourth Amendment search. A Yahoo piece casts the decision as a lit fuse under Flock Safety's automated plate readers.