The Margin Collapse

Last week we ran the demand side of the squeeze: the average company's AI bill for a developer is on track to cross the developer's own salary by 2029. This week an engineer named Martin Alderson published the supply side, and it points the other way. His claim is that an open-weight model out of China now does frontier-grade coding for under a fifth of what Opus charges, and that the roughly ninety-percent margin the labs earn on inference is not a moat but a clock. It is a genuinely important argument. It is also napkin math, funded in part by a company that sells the cheap model, and the exit it describes runs through Mainland-China data terms. Hold all three at once.

The piece is titled "GLM 5.2 and the coming AI margin collapse," posted July 6 on Alderson's personal blog.¹ GLM 5.2 is a large language model from Z.ai, the Chinese lab Alderson describes as having a "deep connection to Mainland China," and unlike Opus or GPT it is an open-weights model: you can rent it from third-party hosts rather than only from its maker. Alderson spent a week putting it through real coding work, and came away convinced something had shifted.

The number that makes the argument

Start with price, because price is the whole case. The going rate for GLM 5.2, Alderson writes, "seems to be around the $4.40/MTok mark": four dollars and forty cents per million tokens of output. He puts that at "less than 20% of the retail price of Opus and ~15% the cost of GPT5.5." If that holds, a shop paying frontier prices for coding assistance could cut the bill by roughly five-sixths by switching hosts, and the switch is nearly free: both Z.ai and the American host Fireworks expose endpoints that speak the same protocol as OpenAI and Anthropic, so tools like Claude Code and Codex point at them by changing a URL.

That last detail is what turns a cheaper competitor into a threat. A better model that requires you to rewrite your stack is a project you can schedule. An 80%-cheaper model that slots into the exact API call you already make is a temptation your finance department will find on its own.

A better-but-incompatible model is a project. An 80%-cheaper drop-in replacement is a temptation your finance department finds on its own.

The margin he says is a countdown

From price, Alderson reasons back to what the incumbents are earning. When Anthropic or OpenAI charge on the order of twenty-five dollars per million tokens for inference, he writes, "my napkin maths suggests that this is probably something like 90% gross margin on the cost of compute." If an open-weight model can be served profitably at $4.40, then most of the gap between $4.40 and $25 is margin, not cost. And margin that wide, sitting on a product a competitor gives away the weights to, is the kind of thing that does not survive contact with a spreadsheet for long. That is the collapse in the title: not that the labs vanish, but that the fat inference margin funding everything else gets competed toward the floor.

It is a clean, alarming story, and it pairs exactly with the demand-side curve we covered last week. From inside, costs balloon: a lab spending millions of compute per employee, an industry on track to spend more on the tool than on the engineer. From below, an open model undercuts the price the whole edifice depends on. The squeeze closes from both jaws at once. If you wanted a single frame for the summer's economics, that is the one.

Now the three caveats, because the story is too good to leave unqualified

First: this is one person's estimate, and he says so at every step. He writes that the price "seems to be" $4.40. His margin figure is "napkin maths," "probably" ninety percent but it "may be a bit higher, or a bit lower." And the competitive verdict, that GLM 5.2 is "the first model that reaches the 'bar' of a genuine open weights competitor to Opus and GPT," opens with "I believe." None of that is a knock on Alderson, who is refreshingly honest about his own uncertainty. It is a warning about how the claim will travel. By the third repost it will be "open-source just killed Anthropic's margins," and the hedges that make it responsible will have fallen off. The numbers are a working engineer's careful guesses, not audited figures, and the difference matters most to the people most eager to quote them.

Second: the experiment was funded, in kind, by a seller of the thing it recommends. Alderson discloses it plainly at the bottom of his post: "Fireworks kindly gave me some free credit to experiment with GLM to help write this article." Fireworks is one of the hosts that sells GLM access. This does not mean the analysis is wrong. The disclosure is exactly what integrity looks like, and we would trust it less without it. But an argument that a cheap model undercuts the incumbents, underwritten by a company whose business is selling that cheap model, has a thumb somewhere near the scale, and a reader deserves to know which hand it belongs to. We would be failing our own ethics page if we passed his number along without passing along who paid for the week that produced it.

Third: the cheap exit has a passport. The $4.40 model is served, at its source, by a company Alderson describes as having a "deep connection to Mainland China," with consumer subscription terms whose data-privacy provisions he calls "weak." "Open weights" and "cheap inference" are real, but the frictionless version, pointing your coding agent at Z.ai's own endpoint, means routing your codebase, and whatever is in it, through a Chinese company's servers under those terms. You can avoid that by self-hosting the weights or renting from a Western host like Fireworks, but that is a different, costlier proposition than the headline number, and the headline number is what does the persuading. The savings are real. So is the string attached to the cheapest way to claim them. And GLM 5.2 matches Opus on coding specifically (Alderson notes it still lacks vision and its web search is slow), not on everything a frontier model does.

What survives the caveats

Strip the piece down to what a skeptic has to concede, and something still stands. An open-weight model you can download now does coding work a careful practitioner rates alongside the closed frontier. It can be served for single-digit dollars per million tokens. It drops into the tools people already use. Whatever the exact margin is, the direction of pressure on inference pricing is down, and the thing pushing it down cannot be bought, sued, or acquired, because the weights are already loose in the world.

That is the part worth sitting with. Every story we have told this summer about the cage (the classifier that reroutes you to a cheaper model, the reasoning budget that quietly shrank, the vendor telling its own staff to demand efficiency) is the behavior of companies protecting exactly the margin Alderson is describing. His post is a bet that the margin is protectable for a shorter time than the share price assumes. He might be wrong about the timing; his math is admittedly loose. But the mechanism he is pointing at is not speculative. It already downloaded.