Zhipu AI Hong Kong IPO Analysis: R&D Investment and Commercialization Balance Challenges Behind HK$4.3 Billion Fundraising

Zhipu AI (Knowledge Graph Technology Co., Ltd.) officially listed on the Hong Kong Stock Exchange on December 29, 2025, becoming an important new member of the AI sector in the Hong Kong stock market [1]. This IPO issued 37.42 million H-shares at a price of HK$116.20 per share, raising a total of approximately HK$4.35 billion (equivalent to US$435 million) with a corresponding valuation of about HK$51.2 billion [1][2].

From the market environment perspective, 2025 was a bumper year for HKEX IPOs, with total annual IPO financing reaching approximately US$75 billion, ranking first in Asia. Zhipu AI’s successful listing also benefited from the investment boom brought by the rise of Chinese local AI enterprises like DeepSeek, as investor interest in Chinese AI companies significantly increased [1].

Zhipu AI’s core technology product is the GLM-4.5 series large models, which were released in July 2025 and adopt a Mixture of Experts (MoE) architecture design, representing the cutting-edge level of domestic large models.

• GLM-4.5
  : Total parameters 355 billion, active parameters 32 billion
• GLM-4.5-Air
  : Total parameters 106 billion, active parameters 12 billion

Notably, GLM-4.5 performs outstandingly in parameter efficiency. Its parameter count is only 1/2 of DeepSeek-R1 and 1/3 of Kimi-K2, but it performs better in comprehensive capability evaluations [3].

In 12 authoritative benchmark tests such as MMLU Pro, AIME24, MATH 500, SciCode, and GPQA, GLM-4.5 ranked third globally, first domestically, and first among open-source models in comprehensive scores. GLM-4.5-Air ranked sixth globally [3].

The core technical breakthrough of GLM-4.5 lies in

, making it the first model in the industry to achieve this feature. The model has the following advantages:

1. Efficient Reasoning Capability
   : The high-speed version generates over 100 tokens/second, and the GLM-4.5-Air version can reach 250 tokens/second [3]
2. Native Agent Capability
   : Can directly generate fully functional applications and support multi-modal content creation
3. API Compatibility
   : Provides endpoints compatible with Anthropic API for easy developer integration

As a technology-driven enterprise, Zhipu AI’s R&D investment is expected to continue growing at a high speed. According to industry practices, large model training costs are extremely high, with a single training possibly costing tens of millions of dollars. A significant proportion of the IPO fundraising is expected to be used for:

• Next-generation model training and iteration
• Computing infrastructure expansion
• Recruitment and training of top talent
• Deepening multi-modal capabilities

Zhipu AI’s commercialization model is characterized by

. According to public information, 85% of its revenue comes from on-premises deployment business for enterprise customers [4]. This revenue structure has the following characteristics:

• High customer unit price with project-based charging
• Relatively strong customer stickiness
• High certainty of payment collection

• Gross margin is limited (on-premises deployment involves large hardware and implementation costs)
• Scalable expansion is limited by the size of the sales team
• Long project cycles make it difficult to achieve exponential growth
• Highly dependent on sales and business development capabilities

Zhipu AI’s revenue structure is similar to that of listed AI company SenseTime, and both face typical challenges of

[4]:

1. Labor-intensive
   : On-premises deployment requires a large number of on-site implementation and custom development personnel
2. Marginal cost is difficult to reduce
   : Each project requires human input, making it difficult to achieve the scale effect of software products
3. Sales-driven characteristics are obvious
   : The usability of the technology itself has limited correlation with revenue, and competition is more reflected in business relationships

Zhipu AI faces multiple challenges in commercialization implementation:

• International giants: OpenAI, Anthropic, etc., continue to expand into the Chinese market
• Domestic competition: Strong opponents like Baidu (Wenxin), Alibaba (Tongyi), ByteDance (Doubao), and MoonShot (Kimi)
• Open-source ecosystem: Open-source models like Meta Llama and DeepSeek seize market share

• Large model technology evolves rapidly, requiring continuous large investments to maintain competitiveness
• Training costs continue to grow with improvements in model capabilities
• New directions like multi-modal and Agent require new investments

• Enterprise customers’ concerns about AI return on investment
• Difficulty in quantifying the value of general large models
• Need for deep industry know-how to create differentiated value

The HK$4.3 billion fundraising needs to be reasonably allocated between R&D investment and commercialization implementation. Recommended allocation directions include:

Usage Direction	Recommended Proportion	Core Objective
R&D Investment	40-50%	Model iteration, computing power construction, basic research
Commercial Expansion	25-35%	Sales team, industry solutions, ecosystem building
Operational Reserve	15-25%	Talent incentives, daily operations, risk buffer

1. Focus on differentiated capabilities
   : Continuously strengthen GLM-4.5’s leading advantages in Agent and native reasoning to avoid homogeneous competition with general large models
2. Improve parameter efficiency
   : Maintain the technical route of “small parameters, high performance” to reduce reasoning costs
3. Open-source strategy
   : Expand the developer ecosystem through open-source to form network effects

1. Industry deep cultivation strategy
   : Select 2-3 high-value vertical industries (such as finance, medical care, education) for in-depth layout to form industry solutions
2. API service expansion
   : Increase the proportion of cloud API service revenue to reduce dependence on on-premises deployment
3. Developer ecosystem
   : Cultivate a third-party application ecosystem to achieve commercialization through platform models
4. Productization of enterprise services
   : Precipitate project experience into reusable product modules to improve delivery efficiency

Zhipu AI needs to pay attention to the following key indicators to evaluate the balance effect between R&D and commercialization:

• Unit parameter performance improvement幅度
• Training cost/parameter ratio
• Academic influence (papers, citations)

• Increase in the proportion of cloud API revenue
• Customer retention rate and expansion rate
• Gross margin improvement trend
• Customer acquisition cost (CAC) vs. lifetime value (LTV) ratio

1. Technical leadership
   : GLM-4.5 ranks among the top in domestic large model evaluations with obvious technical barriers
2. Hong Kong stock scarcity
   : As one of the first large model enterprises listed in Hong Kong, it has a first-mover advantage
3. Policy support
   : AI is a national strategic emerging industry and enjoys policy dividends
4. Market scale
   : The Chinese AI market is still in a high-growth period with broad space

1. Continuous loss risk
   : Large model enterprises are generally in a loss state with unclear profit paths
2. Technical iteration risk
   : AI technology changes rapidly and may be disrupted by new technologies
3. Commercialization uncertainty
   : The sustainability of the business model with 85% revenue relying on on-premises deployment is questionable
4. Valuation pressure
   : The HK$51.2 billion valuation corresponds to high expectations and requires sustained high growth to support
5. Talent competition
   : Core talent has high mobility, and team stability is crucial

Zhipu AI’s Hong Kong IPO raised HK$4.3 billion, providing important financial support for its subsequent development. However, how to achieve effective commercialization implementation while maintaining technological leadership remains the core challenge facing the company.

From a strategic perspective, Zhipu AI needs to achieve breakthroughs in the following directions:

1. Business model transformation
   : Gradually reduce dependence on on-premises deployment and increase the proportion of cloud service revenue
2. Differentiated competition
   : Build moats in directions like Agent capabilities and industry solutions
3. Efficiency improvement
   : Optimize simultaneously in both model efficiency and commercialization efficiency
4. Ecosystem construction
   : Build a developer ecosystem and enterprise customer network to form a flywheel effect

Looking ahead, Zhipu AI is expected to occupy a favorable position in the Chinese AI large model competition with its technical accumulation and capital market support. However, whether it can achieve an effective balance between R&D investment and commercialization will determine whether it can become a true AI leader.

---

[1] Reuters - “Six China IPOs debut in Hong Kong after raising $900 million to cap banner year” (https://www.reuters.com/world/asia-pacific/six-china-ipos-debut-hong-kong-after-raising-900-mln-cap-banner-year-2025-12-30/)

[2] Reuters - “Chinese AI firm MiniMax, others launch Hong Kong IPOs in year-end rush” (https://www.reuters.com/world/asia-pacific/chinese-ai-firm-minimax-targets-up-539-million-hong-kong-ipo-2025-12-30/)

[3] Zhihu - “How to Evaluate Zhipu’s New Generation Open-Source Large Model GLM-4.5: What Technical Breakthroughs Does It Have?” (https://www.zhihu.com/question/1933282570436589214)

[4] Zhihu - “Zhipu and MiniMax Compete for the ‘World’s First Large Model Stock’; Hillhouse Bets on Both Lines” (https://www.zhihu.com/question/1985979705279673214)