GLM 5.2 and the coming AI margin collapse

Martin Alderson published a detailed analysis on July 6 arguing that GLM 5.2, an open-weights model from Chinese lab Z.ai, marks the real “DeepSeek moment” for the AI industry. Not because of training cost, but because of what it does to inference pricing.

The DeepSeek R1 panic in early 2025 was a misread. Markets assumed cheap training meant cheap AI, and Nvidia’s stock cratered. The reality is that training is a fixed cost. Even if it costs hundreds of millions, you spend it once. Inference is the variable cost that scales with every query. And the frontier labs have been charging as though inference were scarce.

Alderson’s napkin math suggests that when Anthropic and OpenAI charge $25 per million tokens for inference on their flagship models, the gross margin on compute alone is roughly 90%. OpenAI’s leaked financials show a ~60% gross margin on total revenue, but that includes support, payment processing, and other services. The core business model is straightforward: spend heavily on training and salaries, then amortize that cost over highly profitable inference. If you can sell enough tokens at 90% margin, the math works.

GLM 5.2 breaks that math.

Alderson reports that GLM 5.2 is the first open-weights model he considers a genuine competitor to Anthropic’s Opus and OpenAI’s GPT-5.5. He struggled to tell the difference in day-to-day use. The model is slower due to extended thinking, and it lacks vision support and good web search. But for non-interactive agentic tasks like reviewing pull requests, it works as a drop-in replacement. Both Z.ai and Fireworks offer OpenAI-compatible and Anthropic-compatible endpoints. Alderson says migrating from Claude Code or Codex to GLM 5.2 is as simple as changing the base URL and API key.

The price is the story. GLM 5.2 runs at roughly $4.40 per million tokens. That is less than 20% of Opus’s retail price and about 15% of GPT-5.5’s. Even accounting for the model’s tendency to generate more tokens through extended thinking, Alderson estimates it is more than 50% cheaper for nearly all workflows. Z.ai also offers a “coding plan” subscription that mirrors Anthropic and OpenAI’s plans, but with higher claimed usage limits.

The switching costs are near zero. Alderson notes that this is not a Microsoft or Salesforce level of lock-in requiring years of migration planning. The friction of keeping up with policy changes and term updates from frontier labs is actually higher than switching inference providers. Data privacy concerns around Z.ai’s Mainland China ties are real, but open weights mean enterprises can host the model on premises or use any of the dozens of providers with proper contractual provisions.

Part two of Alderson’s series, not yet published, will address what a collapse in inference margins does to the industry. The question is not whether margins compress, but how fast and who gets squeezed.

The frontier labs have been running a business model that depends on 90% gross margins for inference. That margin is not a reflection of cost. It is a reflection of scarcity. Open-weights models that match frontier quality at one-fifth the price erase that scarcity. The labs can try to differentiate on speed, vision, web search, and agentic tooling. But those advantages are narrowing. GLM 5.2 lacks vision and good search today. Those gaps will close.

The AMD angle matters. Wafer published an analysis of running GLM 5.2 on AMD hardware, reporting that inference is 2.75x cheaper per token on AMD versus Nvidia Blackwell. If open-weights models can run on cheaper hardware, the cost floor drops further. The frontier labs are tied to Nvidia’s pricing. The open-weights ecosystem is not.

Alderson closes with a line from Jeff Bezos: “Your margin is my opportunity.” The opportunity here is for any company that can serve open-weights models at cost plus a small markup. The risk is for labs that have built their entire financial model on inference margins that no longer hold.

The coming margin collapse does not mean AI becomes unprofitable. It means the profits shift from inference providers to those who can build applications and services on cheap, commodity models. The frontier labs will need to find new revenue streams or accept that their core product is becoming a low-margin utility.