Summary of Key Points
Recently, there have been two contrasting news stories in the AI community: on one hand, MiniMax experienced a significant drop of 18% on its first day of trading after the lock-up period was lifted, with its stock price falling by 80% from its March high; on the other hand, Zhipu’s stock price increased by 10% despite the dual pressures of its cornerstone investors releasing their shares and the company issuing new shares at a discount. Zhipu was able to withstand these challenges because its GLM-5.2 model has become very popular among developers worldwide, competing with models from OpenAI and Anthropic. However, despite its stock price having increased 14 times in value, Zhipu is still urgently seeking to raise over HK$30 billion. The reason for this is that the development of large-scale models is extremely costly; the funds raised through its IPO are almost depleted, and the more popular the model becomes, the higher the推理 costs, necessitating the development of the next generation of models. Moreover, while GLM-5.2 has gained global recognition, Zhipu’s main revenue comes from “one-time” localized deployments, with API subscriptions accounting for only a small portion of its total income. To match OpenAI’s success, Zhipu needs to address this issue.
I. Zhipu vs MiniMax: Model Strength Determines Stock Price Resilience
The significant drop in MiniMax’s stock price, compared to Zhipu’s rise, is primarily due to the difference in model capabilities. The decline in MiniMax’s stock price from HK$1330 to below HK$300 indicates that the market is not optimistic about its future prospects. In contrast, Zhipu’s GLM-5.2 model has outperformed other open-source and closed-source models in third-party evaluations, with its programming capabilities even surpassing those of Claude Opus. Overseas developers, such as the founder of Vercel, have praised it as “astonishing,” and OpenRouter has noted that Zhipu has incorporated Claude-level capabilities into an open-source model, which was used more frequently than Anthropic’s models. This strong technical foundation has led the market to believe in Zhipu’s ability to compete with international giants, allowing its stock price to rise despite the recent events.
II. The HK$30 Billion Is Not Just for Fundraising; It’s a Ticket to Feed the Growth of Large-Scale Models
With a market value of around HK$70-80 billion, why does Zhipu still need to raise HK$30 billion? The reason is that large-scale models are extremely costly to develop and maintain:
1. High Cost of Development: Zhipu raised HK$4.896 billion in its IPO in January this year but spent 93% of it within half a year (only HK$300 million remaining). In 2025, it plans to invest HK$3.18 billion in research and development, which is more than four times its revenue (HK$724 million).
2. Increasing Costs as Models Become More Popular: The more people use GLM-5.2, the higher the推理 costs (e.g., server power requirements). Since large-scale models are updated frequently (OpenAI releases new versions every few months), Zhipu must continue to develop GLM-6 and GLM-7 to stay ahead.
3. The Next Round of Investment: While GLM-5.2 has helped Zhipu join the global elite, staying there requires continuous investment in top researchers, purchasing computing power, and training the next generation of models.
III. How Will the HK$30 Billion Be Spent?
Zhipu has clearly outlined how it plans to spend these HK$314 million:
- R&D and Talent: Recruiting and retaining model developers (with annual salaries in the millions for top talents).
- Computing Power: Purchasing or leasing cloud computing resources (training a large-scale model can cost hundreds of millions), as well as repaying loans from previous purchases.
- Next Generation Models: Developing GLM-6 and GLM-7 to enhance programming and reasoning capabilities.
- Other: Strategic investments and acquisitions to support operations.
Even more ambitious, Zhipu plans to spend all of this HK$30 billion within one and a half years, with an average monthly expenditure of HK$1.7 billion. To further finance its growth, it also intends to raise another RMB 15 billion on the STAR Market by 2030, highlighting the intense competition in the large-scale model industry.
IV. Global Recognition of Models, but Income Lags Behind: Localized Deployments for Now, API Subscriptions for the Future
Although GLM-5.2 has gained global attention, Zhipu’s revenue structure is not yet aligned with its global presence:
- Current Main Revenue Source: Localized deployments (73.7% of 2025 revenue). This involves installing the model on clients’ servers (e.g., banks and governments) for a one-time fee. While this model is secure and generates substantial amounts, it provides only a one-time income stream.
- Future Potential: API subscriptions (26.3% of 2025 revenue). This model generates recurring income based on the number of times the model is used; OpenAI relies heavily on this model for its subscription-based services (e.g., ChatGPT Plus).
Although Zhipu’s API revenue has grown rapidly (292.6% year-on-year in 2025, with an 83% increase and a 400% increase in usage this quarter), the current HK$190 million in revenue is still small compared to its market value of several billion. To become a global powerhouse like OpenAI, Zhipu must make API subscriptions its main source of income. After all, a single model deployed in the cloud can serve countless developers globally, offering low replication costs and significant growth potential.
V. Conclusion: Opportunities and Challenges for Zhipu
Zhipu’s opportunities lie in the fact that GLM-5.2 has proven its technical strength, providing it with a favorable financing opportunity (the current share price is 13.7 times higher than its IPO price, allowing it to raise funds with fewer shares issued). However, the challenges are also clear:
1. High Financial Pressure: Spending HK$30 billion in one and a half years and needing to raise more capital.
2. Revenue Transformation: It must shift from relying on one-time projects to recurring API subscriptions for sustainable growth.
3. International Competition: With OpenAI and Anthropic continuing to advance, Zhipu must continuously develop even more advanced models.
In summary, Zhipu has a strong foundation, but it still needs to invest heavily to realize its potential. Whether it can become a Chinese version of OpenAI depends on its ability to convert the global impact of its models into substantial revenue.