How GLM 5.2 Is Driving the AI Margin Collapse

$25 per million tokens for model inference. That's what we're used to, but not for long. GLM 5.2 by Z.ai is shaking things up and threatening steep pricing and profit norms. The AI margin collapse isn't just a story—it's coming, driven by advances like GLM 5.2.

Key Takeaways

• Inference costs drive AI economics, not training costs.
• GLM 5.2 challenges high-margin models like GPT.
• 'Open weights' models pressure existing AI business models.
• $25/MTok inference rates hide high gross margins.

The Real Cost in AI: Inference Over Training

Most think training is the big cost in AI, but it's actually inference. Training demands a fixed investment—$6 million or more for top-notch models—but inference scales with demand, hitting real marginal costs Martinalderson. When Anthropic or OpenAI set $25 per million tokens for inference, their gross margins likely sit between 60% and 90%, before extra operational expenses.

Why Inference Costs Matter More

AI companies bank on spreading these large initial investments across lucrative inferences. But this financial cushion weakens when new rivals offer cheaper alternatives without losing quality.