Zhipu’s GLM 5.2 wowed devs, but Z.ai stock plunged 40% despite “open” weights

Z.ai, also known as Zhipu, launched its open-weight AI model GLM 5.2 last month and “wowed the industry.” The stock reaction tells a different story: shares have plunged more than 40% over the past month. The business problem is not that developers stopped caring. It is that open-weight AI is not the same thing as open-source software, and the economics that made open-source successful do not translate cleanly to AI models.

To make the mismatch concrete, Z.ai is publicly traded, so the numbers are visible. Last year, the company lost almost $500 million on revenue of about $107 million. That is a brutal gap between hype and the financial engine required to keep inference running, especially when the “open” distribution model puts price and infrastructure pressure on the wrong balance sheet.

This is where the analogy to open-source breaks. In software, once you write code, distributing it to the next customer can be close to free. That is why open-source can become an excellent business. Red Hat is the classic proof point: IBM bought it for $34 billion, and the model benefitted from a world where customers get the code, then pay for support, services, and enterprise editions.

Open-weight AI is different because AI output is expensive every time it runs. Each answer depends on chips, electricity, and data-center capacity. The “next unit of software” might be nearly free, but the “next unit of intelligence” is not. Model makers can provide the trained numerical parameters, which are like tiny numerical dials inside a model’s brain. But once other companies download those parameters and run the model, the cost structure shifts. Inference is ongoing compute, not a one-time distribution event.

Moonshot’s recent Kimi K3 example illustrates how quickly the bill can catch up to ambition. Its new open-weight model impressed the industry with frontier-level performance. But days after launch, it halted new customer sign-ups because it did not have enough computing power to run the model. If this were traditional software, adding millions of users would mostly be a scaling and licensing problem. In AI, every new customer can add to infrastructure costs, which caps growth and makes unit economics unforgiving.

The “who captures the profit” issue gets even sharper when you look at how these open-weight models are used in practice. Open-weight labs share trained parameters so outsiders can download and run them. Outsiders can also run models on cloud giants such as Amazon, Microsoft, Google, Oracle, and Alibaba. There are specialist providers too, like Fireworks AI and Baseten, though the source notes they largely rent capacity from major cloud companies. Or companies can use the model maker’s own inference service, but in the Western world, corporate customers often avoid that option for data security reasons.

Only that last path reliably generates real revenue for the model creator. The other three routes can leave the lab that trained the system with little or no ongoing revenue, even though it spent heavily in the first place to build the model. That dynamic helps explain why Chinese model labs may be struggling even when their models perform well.

And investors are noticing. MiniMax, another independent Chinese AI lab that is publicly traded, lost $250 million last year on revenue of just $79 million. Its shares have fallen more than 50% over the past month. At the same time, Alibaba’s stock is up about 13% over the past month, aligning with the idea that hosting and inference infrastructure can be where the money flows, not the lab that shipped the open weights.

Analysts call out the pricing and margin squeeze. In a recent note to investors, William Blair analyst Arjun Bhatia wrote that open-weight models have a “challenging path to making a profit,” and he later added: “Unlike open-source software, open-weight models do not generate significant sums of revenue by selling support, services, and enterprise editions around the free offering (the Red Hat playbook).” Bhatia also warned that inference workloads will flow to whoever operates the inference infrastructure most efficiently, and that is usually not the model provider. Raimo Lenshow of Barclays reportedly reached a similar conclusion after researching China’s AI sector, saying intense domestic competition has led to more aggressive pricing competition and that open-source or open-weight models accelerate pricing pressure throughout the system.

So why release open-weight models at all, if the business case looks so ugly? The source frames it as a challenger strategy: distribute technology widely to attract developers, pressure incumbents, and complicate premium pricing. That is exactly how open technology can work as a weapon. Even if Chinese labs make little money themselves, cheaper alternatives can force American competitors like OpenAI and Anthropic to cut prices and make it harder to recover the billions spent training new models.

The political layer matters too. The source says China’s leadership is now openly encouraging that openness approach. In a speech in Shanghai, President Xi Jinping said countries should seize the opportunity to encourage “open-source, openness, collaboration, and sharing.” The article notes that is not likely to be treated as casual guidance. Companies such as Moonshot, Zhipu, and MiniMax may face strong pressure to align with strategic priorities even if it makes their own path to profit harder in the near term.

For executives and boards evaluating AI strategies, the takeaway is uncomfortable but actionable: calling something “open” does not automatically produce open-source economics. Open-weight can drive adoption, but it can also shift the revenue center of gravity toward infrastructure operators and the providers best positioned to sell inference compute. If you are a model maker, your growth plan has to survive the reality that customers can run your model elsewhere, and every additional query comes with a compute cost you might not directly monetize. If you are a platform or investor, you should expect more volatility in “model-led” businesses when openness accelerates pricing pressure and compresses long-term margins.