The capital race within China's large model sector has finally reached its zenith.

After weathering the 'hundred-model battle' and experiencing frenzied activity in the primary market, Beijing Zhipu Huazhang Technology Co., Ltd. (hereinafter referred to as 'Zhipu') has secured the prestigious title of the 'first global large model stock'. According to the latest disclosure from the Hong Kong Stock Exchange, Zhipu is set to officially list on the main board of the Hong Kong Stock Exchange on January 8, 2026. This date is one day earlier than that of MiniMax, which was initially scheduled to list on January 9. For Zhipu, which emerged from a Tsinghua University laboratory, this one-day lead carries significant symbolic weight.

Global public offering documents indicate that Zhipu's current fundraising efforts are expected to yield HK$4.3 billion. The company has attracted 11 cornerstone investors, including Shanghai Gaoyi, GF Fund Management, and Taikang Life Insurance, with subscription amounts accounting for nearly 70% of the offering size.

However, beneath the glamorous 'first stock' halo and the star-studded shareholder lineup lies a sobering reality. The prospectus reveals that Zhipu has accumulated losses exceeding HK$6.2 billion over the past three and a half years. Meanwhile, as its business model transitions from software delivery to cloud-based MaaS (Model as a Service), the gross profit margin of its cloud business has plummeted from 76.1% in 2022 to -0.4% in the first half of 2025.

On one hand, Zhipu boasts a unicorn halo with an estimated valuation exceeding HK$50 billion. On the other, it faces a financial black hole where 'the more you sell, the more you lose.' With top-tier technical endorsement from the Tsinghua faction on one side, and a survival red line where cash flow can only sustain operations for a few months on the other, Zhipu's listing is less of a crowning glory moment and more of a 'race against time' to maintain its position in the market.

The Darling Under the 'Tsinghua Faction' Halo

In China's large model landscape, Zhipu lacks the traffic foundation of internet giants and does not have the rugged, entrepreneurial origins of some startups. Instead, it is a typical representative of the 'academic school,' with roots in the Knowledge Engineering Laboratory of Tsinghua University's Computer Science Department.

Founded in 2019, Zhipu's core founding team is almost entirely composed of Tsinghua alumni. CEO Zhang Peng holds a Ph.D. in Computer Science from Tsinghua, Chairman Liu Debing previously served as the Deputy Director of the Technology Big Data Research Center at Tsinghua's Institute for Data Science, and Professor Tang Jie, the company's 'technical soul,' is a top scholar in the global AI field who led the development of the ultra-large-scale intelligent model 'Wudao.'

Its strong academic background has allowed Zhipu to maintain a certain 'aloofness' and leadership in its technical approach. Unlike many domestic large models that are essentially repackaged versions of Llama, Zhipu insists on full-stack independent research and development, having launched the GLM pre-training architecture. The GLM-4 series, released in 2024, rivals GPT-4 in several authoritative evaluations and performs exceptionally well in vertical areas such as code generation.

Zhipu's technical prowess has made it a favorite in the primary market. Prior to its IPO, Zhipu completed eight rounds of financing, raising a cumulative total of over HK$8.3 billion. Its shareholder list includes industry giants such as Meituan, Alibaba, Tencent, and Xiaomi, as well as top-tier VC firms like Sequoia Capital, Hillhouse Capital, and Legend Capital. This reflects both recognition of Zhipu's technical strength and the intense desire among Chinese capital for companies with independent underlying model capabilities following the rise of OpenAI.

However, unlike MiniMax, which is heavily investing in consumer-facing applications (such as Talkie) to gain traffic through hit apps, Zhipu has chosen a B2B (enterprise-level) route that aligns more with its 'academic school' image but is also more challenging. The prospectus shows that Zhipu's revenue primarily comes from two segments: local deployment (private deployment) and cloud-based MaaS services (API calls).

In its early development stages, Zhipu resembled a traditional software solutions provider. In 2022, 95.5% of its revenue came from private deployments. Customers for this model are mostly government agencies, central state-owned enterprises, and large financial institutions. The advantages of this model lie in high average order values and gross margins, providing early revenue generation for the company.

However, private deployment is a typical 'project-based' business with heavy delivery, long cycles, low repurchase rates, and difficulty in reducing marginal costs through economies of scale. The prospectus reveals that from 2022 to the first half of 2025, none of Zhipu's top five customers overlapped, meaning the company has to constantly find new major clients each year to sustain growth, resulting in immense customer acquisition pressure.

To break this bottleneck and tell a more compelling capital story, Zhipu began aggressively transitioning to cloud-based MaaS services, attempting to sell large models like water and electricity through API interfaces to various industries. In the first half of 2025, revenue from cloud deployments increased to 15.2%, but this has pushed Zhipu into another brutal battlefield—the price war.

Behind the 6.2 Billion Yuan Loss

'Kuaima Financial Media' discovered that while Zhipu AI's cumulative revenue rapidly increased from 2022 to the first half of 2025, its cumulative net loss during the same period reached HK$6.2 billion. In 2022, 2023, 2024, and the first half of 2025, Zhipu incurred losses of HK$144 million, HK$788 million, HK$2.958 billion, and HK$2.358 billion, respectively. Even after excluding non-operating factors such as changes in the fair value of preferred shares, its adjusted net loss remains extremely high.

The massive loss is closely tied to the sharp decline in the gross profit margin of its cloud business. The prospectus shows that the gross profit margin of Zhipu's cloud deployment business was 76.1% in 2022, indicating high profitability. However, it plummeted to 46.6% in 2023 and further dropped to -0.4% in the first half of 2025. In other words, Zhipu not only fails to make money from each API service sold but also incurs losses. The root cause of this situation is the fierce 'price war' in China's domestic large model market since 2025.

To compete for developers and market share, models like Doubao, Tongyi Qianwen, and Wenxin Yiyan have slashed API call prices to 'rock-bottom' levels, even offering them for free. For internet giants with cloud computing infrastructure, they can subsidize large model losses with revenue from cloud services. However, for independent model vendors like Zhipu, this represents a 'dimensionality reduction' attack. Following suit in price cuts means 'selling blood' to operate, while not doing so results in customer loss.

The prospectus mentions that nine of China's top ten internet companies are major clients of Zhipu's cloud business. These giants possess strong bargaining power and mostly have self-developed models. Their use of Zhipu's services is primarily for 'backup' or supplementary purposes in specific scenarios, resulting in extremely low loyalty. Under intense competition, Zhipu has been forced into the price war, leading to a collapse in gross margins.

Besides low-price competition, massive R&D investment is another major cause of the losses.

The prospectus shows that from 2022 to the first half of 2025, Zhipu's cumulative R&D expenditure reached HK$4.4 billion. In 2024 alone, R&D expenses amounted to HK$2.195 billion, more than seven times the revenue for that year.

In the large model sector, computing power is the lifeblood. As model parameters grow exponentially, training and inference costs rise geometrically. In 2024 alone, Zhipu's computing service fees reached HK$1.553 billion, accounting for over 70% of R&D expenditure. To catch up with GPT-4's capabilities, Zhipu must continuously burn money on NVIDIA GPUs and cloud computing power.

This 'loss-for-technology' model has built a technical moat but also keeps the company's cash flow constantly strained. As of the first half of 2025, Zhipu had approximately HK$2.55 billion in cash and cash equivalents on hand. Considering the net loss of HK$1.75 billion in the first half of 2025, this amount can only sustain the company's operations for about nine months without additional financing.

Although the prospectus mentions that the company has access to billions of yuan in bank credit lines, for a technology company that is not yet profitable and has light assets, large-scale borrowing is not a sustainable long-term strategy.

This explains why Zhipu's IPO process has been so rapid: filing in December and listing in January. This is not only to provide an exit channel for early investors such as Alibaba, Tencent, and Meituan but also to open up financing channels in the secondary market before the company's cash flow runs out, ensuring sufficient 'ammunition' for the upcoming protracted war.