Similar to the model performance iteration that leads the AI industry day by day, the capitalization process of big model startups is much faster than in other industries. The rise of the big model investment boom has only been two years, and the "world's first big model stock" is about to come.

From this perspective, the landing speed of AI in China is indeed much faster than in North America. OpenAI has just announced plans to go public, and Chinese big model companies have already lined up at the doorstep of ringing the bell.

On December 19th, Zhipu and MiniMax, representatives of a large model startup company, both passed the listing hearing on the Hong Kong Stock Exchange. Two days later, Zhipu was the first to disclose its prospectus before MiniMax, which was interpreted by many media outlets as taking the lead in the battle for the "first stock of global big models". According to media reports, Zhipu is also one of the fastest companies from mainland China to pass the hearing since the implementation of the "reporting system" for listing in Hong Kong.

Compared to other AI companies that are still waiting for valuation windows, the reason why Zhipu is accelerating its IPO is largely due to the increasingly heavy financial pressure on the big model track. Against the backdrop of global technology giants increasing their investment in talent and computing power, start-up companies are facing almost synchronous increases in research and development costs, while the financing heat in the primary market has significantly cooled down. Accelerating the IPO process is an inevitable choice.

According to the prospectus data, although the compound annual growth rate of Zhipu's revenue in the past three years has exceeded 130%, the speed of loss expansion has also been almost synchronous, and the growth rate is more significant than the revenue.

From 2022 to the first half of 2025, the cumulative revenue of Zhipu in the three and a half years is only 685 million yuan, while the cumulative loss during the same period exceeds 6.2 billion yuan. The proportion of losses to revenue has been increasing from 250.2% in 2022 to 1235.3% in the first half of this year

Most of the losses come from high R&D investment. In the past three years, Zhipu's cumulative R&D expenditure has reached 2.809 billion yuan, and will further climb to 1.595 billion yuan in the first half of 2025. The total R&D investment in the past three and a half years has exceeded 4.4 billion yuan.

Sales and marketing expenses have also rapidly increased, from 15 million yuan in 2022 to 209 million yuan in the first half of 2025. The cumulative marketing expenses over three and a half years are about 712 million yuan, which is more than the cumulative revenue of Zhipu during the same period.

Zhipu was founded in 2019, and its core founding team comes from Tsinghua University, including Chairman Liu Debing, Chief Scientist Tang Jie, CEO Zhang Peng, and others, known as the "Tsinghua Gang" in the industry.

The prospectus shows that the founding team holds a total of 33.03% equity in Zhipu through a concerted action agreement. Since its establishment 6 years ago, Zhipu has raised 8 rounds of financing, with a total fundraising of over 8.3 billion yuan, and the latest post investment valuation before IPO exceeded 24 billion yuan.

As of June 30th this year, Zhipu's year-end cash and cash equivalents were approximately RMB 2.552 billion, but its net current liabilities for the same period were RMB 7.089 billion, a significant increase from the end of last year. The newly added RMB 2.888 billion came from financial instruments issued to investors. In the first half of this year, Zhipu received strategic investments from multiple local state-owned assets.

The first independent large model, with an annual revenue of only 300 million yuan

As the first big model startup company to publicly list its shares, Zhipu provides the market with the most intuitive way to observe the commercialization of big models.

In the prospectus, Zhipu cited the Frost&Sullivan report, stating that based on 2024 revenue, Zhipu ranks first among independent general model developers in China and second among all general model developers, with a market share of approximately 6.6%.

According to the definition of Zhipu, independent large models can be briefly summarized as large model startups, Zhipu MiniMax、 The so-called "AI Six Tigers" such as the Dark Side of the Moon belong to this category, while non independent big models generally refer to tech giants such as Alibaba, Tencent, Baidu, and ByteDance.

In its prospectus, Zhipu specifically listed the top big model manufacturers in China for 2024. Except for Zhipu, which ranks second as an independent big model developer, the rest are technology giant companies such as iFlytek, Alibaba, SenseTime, and Baidu.

But this ranking may differ from people's actual perception, partly because the commercialization of China's large model market is limited to B-end government and enterprise customers, and C-end payment willingness is not strong.

According to Frost&Sullivan data, in terms of revenue, out of the 5.3 billion yuan market size of China's large language models in 2024, institutional customers contributed 4.7 billion yuan, while individual customer revenue was only 600 million yuan.

In fact, in the past few years, Zhipu's revenue has mostly come from B-end customers, and the localization deployment of top government and enterprise customers is its main source of revenue. In the past three years, localization deployment revenue has accounted for about 85% or more of Zhipu's total revenue for a long time.

Even though Zhipu has vigorously promoted the cloud API calling service of MaaS platform in the past two years, the proportion of localized deployment revenue in the first half of this year still reached 84.8%, while cloud deployment was only 15.2%. As a comparison, about 70% of OpenAI's revenue comes from paid subscriptions from C-end consumers.

According to the prospectus data, Zhipu Service Institution has over 8000 clients, but the top five clients have long accounted for nearly 50% of the company's revenue. In the past three years, the top five customers of Zhipu accounted for 55.4%, 61.5%, and 45.5% of the revenue, respectively, until the first half of this year when it dropped to 40%.

This may explain why Zhipu has already ranked second in annual revenue among Chinese general model developers and first among independent general model developers, but its revenue last year was only 312 million yuan. According to the historical cash consumption rate of Zhipu in the first half of this year, 312 million yuan is not enough for Zhipu's monthly expenses.

It is somewhat comforting to note that although Zhipu's previous revenue relied excessively on the localization deployment of B-end large customers, the company's gross profit margin has always remained above 50%.

Among them, the gross profit margin of localized deployment is higher, reaching 66% last year, while that of cloud deployment is only 3.4%. In the first half of this year, the gross profit margin of localized deployment reached 59.1%, while that of cloud deployment was negative 0.4%.

Zhipu extensively described its MaaS platform in its prospectus, while emphasizing the growth potential of its cloud business. The Science and Technology Innovation Board Daily quoted Zhipu as saying that the cloud based MaaS and subscription businesses of Zhipu have shown "exponential growth", and "paid traffic revenue exceeds the sum of all domestic models." Zhipu CEO Zhang Peng said that the annual recurring revenue of model services for global developers has exceeded 100 million yuan.

Zhipu is also actively exploring overseas markets, currently more concentrated in Southeast Asia. Starting from 2024, Zhipu's localization deployment service for large models will begin to land overseas and generate revenue. In the first half of 2025, the proportion of overseas revenue from localization deployment will reach 11.6%.

70% of R&D expenses are consumed by computing power

The current big model industry is still in the early stage of high input and low output development. Even if it is as strong as OpenAI, it has to face the pressure of slowing down subscription revenue and try to expand its revenue sources through diversified methods. In the past few years, the big model industry has often made more profits not from big model manufacturers, but from GPU manufacturers such as Nvidia and a large number of cloud service providers.

Tech giants have more ample funds and time to wait for the industry to mature. They can continue to push up computing power and talent investment without considering current losses, and correspondingly often obtain models with stronger performance. Model performance remains the key to determining AI competition.

In the past few months, Google has been pressing OpenAI hard to catch its breath with the super performance of Nano Banana and Gemini 3 Pro, forcing Sam Altman to sound a red alert internally and demand strengthened model performance research and development.

Domestic companies such as Alibaba, Tencent, and ByteDance are also strengthening their competition for AI talent in addition to investing in computing power. At the beginning of the year, Alibaba CEO Wu Yongming announced an unprecedented three-year investment plan of 380 billion yuan in AI, and at the end of the year, he also stated that he did not rule out further expansion of investment. Tencent has just recruited Yao Shunyu, a former OpenAI researcher known as the "genius boy," while ByteDance has poached Wu Yonghui, Vice President of Research at DeepMind.

In contrast, even though Zhipu has accumulated a research and development investment of as much as 4.4 billion yuan in the past three and a half years, it is not the most expensive one in this big model competition.

In order to demonstrate its scientific research capabilities in model competition, Zhipu emphasized in its prospectus that it has established a research and development team of 657 people, accounting for 74.4% of the company's total employees.

But the cost of talent is not the main component of Zhipu's R&D expenses. In fact, most of Zhipu's R&D expenses were spent on computing power consumption, with salary costs accounting for only 14.8% of R&D expenses last year. In 2024 and the first half of 2025, over 70% of R&D expenses will be used for computing service fees.

In July of this year, Zhipu released its flagship model GLM-4.5, and in September, it released the GLM-4.6 model with upgraded coding capabilities. In the prospectus, Zhipu cited various data including benchmark tests, global rankings, and token consumption to demonstrate the leading performance of the model. However, considering the current frequency of model updates on a monthly or even daily basis, it is difficult for the model lead in the first half of the year to continue into the second half.