The first stock surge, the quality, capability, and ambition of China's largest independent model manufacturer.

24.377 billion yuan, on the 17th, Zhiyu confirmed its latest valuation in the prospectus publicly released online after the hearing at the Hong Kong Stock Exchange, marking the first time that people have a precise understanding of the valuation amount of Chinese large model companies.

This is a timely moment; it has been three years since ChatGPT made a shockwave in the industry. After the tumult of the hundred model competition, the Chinese large model players remaining at the table, having proven that their innovation and technical capabilities are no less than anyone else's, are now making a push towards the capital market.

During the critical transition period of “youth stepping into adulthood,” the market anticipates how large model companies will demonstrate to everyone the process of transitioning from “novelty and flashy” model technologies to the end-to-end implementation of “practical and adaptable” large models.

As the first company to successfully “go public” during the Hundred Regiments Offensive, Zhipu's response is not very satisfactory. The losses disclosed in the prospectus far exceed the revenue growth, and R&D expenses continue to increase significantly, seemingly with no trend of stopping the “burning money”.

If it were a mature company, such a balance sheet would be truly unsatisfactory, but large models are special.

Zhipu is one of the earliest independent large model companies established in China, but it has only been around for less than 6 years. The reason why the AI industry excites countless people is its potential for revolutionary change in the future, which is difficult to directly deduce from past history, as transformations often occur suddenly in an exponential form at a certain moment.

This is Zhipu's bet, or it can be called a vision. Before the AGI future, which many are undoubtedly convinced will arrive, Zhipu hopes to be as well-prepared as possible, which includes a series of complex trade-offs and convincing the market.

It makes people believe that this is a bet full of hope.

Core Data Summary

The core commercial revenue of Zhipu is mainly composed of localized deployment and cloud deployment. The former refers to providing private deployment of large models for B-end customers, while the latter provides model API interfaces and token calling services to users through the MaaS platform.

The prospectus shows that, based on 2024 revenue, Zhiyu has already become the largest independent large model vendor in China and the second-largest large model vendor in the enterprise LLM field, surpassing Alibaba and SenseTime, with a market share of 6.6%.

From 2022 to 2024, Zhipu's revenue was 60 million yuan, 120 million yuan, and 310 million yuan respectively, with a CAGR of 130%. The revenue for H1 2024 and H1 2025 is expected to be 40 million yuan and 190 million yuan respectively, with a year-on-year growth of over 300%;

Among them, nearly 85% of Zhipu's revenue comes from localized deployment services, mainly distributed across vertical fields such as Internet technology, public services, telecommunications, traditional enterprises, consumer electronics, as well as retail, media, and consulting industries.

In the first half of 2025, Zhipu's overseas revenue share quickly increased to nearly 12%, mainly from the Malaysian and Singaporean markets. Under the “Belt and Road” strategy, Zhipu has helped overseas countries establish sovereign large models, achieving a breakthrough in the export of our country's large model technology.

From 2022 to 2024 and in the first half of 2025, Zhipu's gross margins were 54.6%, 64.6%, 56.3%, and 50.0%, respectively. In comparison, domestic software outsourcing companies in China generally have around 30%, indicating that the gross margin for Zhipu's localized large model deployment business is not low.

It is worth noting that the decline in gross margin mainly comes from the price fluctuations of the MaaS platform business. Affected by market price wars, Zhibu has increased its price reduction efforts to gain more customers and more application scenarios.

The result of this is that in the first half of 2025, the number of institutional clients for Zhipu's localization and cloud deployment is 95 and 3061 respectively, while for the entire year of 2024, the numbers are 123 and 5457 respectively. The price reduction of the MaaS platform business not only increased the number of cloud institutional clients but also had a significant effect in attracting clients to the localized deployment business.

As of Q3 2025, the company's model has empowered over 80 million end-user devices and more than 45 million developers, with over 12,000 institutional clients, a significant increase from 8,000 clients in H1 2025.

In addition, by November 2025, the average daily token consumption of Zhipu had reached 4.2 trillion, while this figure was 500 million in 2022. This is a result of the release of Zhipu's new generation GLM 4.5/4.6 open-source base model and directly reflects the effects of user growth and call volume.

Income is more important than anything else, but it also requires patience.

Zhipu introduced a “crosswise strategy” narrative in its prospectus for the growth prospects of commercialization revenue.

From a vertical perspective, the price reduction and expansion of MaaS platform services have gained widespread user adoption and application scenarios. The introduction of more users will drive traffic to localized deployment services, and this “MaaS traffic generation + localized monetization” model will initially establish a cycle between customers and businesses.

Horizontally speaking, similar to OpenAI and Anthropic, Zhipu's MaaS platform features high flexibility and scalability. Unlike traditional impressions of heavily customized delivery projects, Zhipu points out that the majority of revenue-generating enterprise-level MaaS projects deliver more generalized model capabilities, allowing for a quicker scaling of token call volume.

The vertical customer cycle model and the horizontal capability expansion of the generic solution strengthen Zhipu's label as the largest independent model company in China by revenue. This is the commercialization synergy that Zhipu expects. The business revenue scale of the MaaS platform is likely to achieve excess growth. According to Zhipu's forecast, the future revenue from localized deployment and cloud services will be evenly split. To achieve this, Zhipu is willing to sacrifice its gross profit margin: unlike the high gross margin of localization projects, the gross profit margin of cloud services is relatively low due to significant costs incurred in computing power procurement. This is also an inevitable result of the strategic price reduction of MaaS and a compromise that must be made in the short term to compete for market share.

The model is the product, and the product is growth.

A popular saying in the large model industry is “model is product,” which reflects a very simple first principle: users pay for the strongest models, and the process of building the model itself is to create a highly competitive product.

This also constitutes the core growth logic of MaaS. In the second half of this year, the latest GLM 4.5/4.6 base models released by Zhipu achieved the native integration of reasoning, coding, and agent capabilities in a single model for the first time. Notably, in coding, GLM ranks first alongside Anthropic and OpenAI's models on Code Arena, and after its release, the call volume on its MaaS platform has grown exponentially.

Similar to the “Vertical and Horizontal Strategy” in the first part, the Powerful Model product is expected to trigger a data flywheel effect, which will be the ideal moat in the competition of large models. The commercialization goal of Zhipu will be realized on a complete platform that includes the latest models, APIs, and development tools. MaaS will become the “operating system” level infrastructure of the AI era, serving thousands of industries and organizations of various sizes, from individuals, small development teams, to large enterprises. On this complete platform, there is no delivery, only product managers.

According to the data from Zhipu in the prospectus, MaaS has over 2.9 million businesses and application developers, making it one of the most active large model API platforms in the country. Currently, 9 out of the top 10 internet companies in China are using Zhipu's GLM large model. Among them, the GLM Coding package (a standardized product of large models that developers subscribe to on a monthly basis) has over 150,000 paying developers within 2 months of its launch, and the annual recurring revenue is rapidly exceeding 100 million.

The cost of the strongest model

Users pay for the strongest model, and Zhipu's cost for this judgment is clearly revealed in the prospectus.

From 2022 to 2024, as well as in the first half of this year, Zhiyu's adjusted net losses were 97 million yuan, 621 million yuan, 2.466 billion yuan, and 1.752 billion yuan, respectively, with research and development expenses amounting to 80 million yuan, 530 million yuan, 2.2 billion yuan, and 1.59 billion yuan, accounting for a significant proportion.

Zhipu pointed out that the proportion of computing power service fees to research and development costs has rapidly increased from the initial 17% to over 70%. This means that a large portion of the R&D costs is used to purchase computing power, which is in sync with the rapid growth in the usage of the MaaS platform.

On one hand, the extremely high investment in R&D is a strategic choice for Zhipu. The capability of the pre-trained base model itself consumes a significant amount of computing power. As we advance through the AGI development stage, it also requires the derivation of more application models to enhance the comprehensive performance of the intelligent agent, thereby solidifying its comparative advantage in underlying model capabilities and attracting more developers and customers.

On the other hand, in the foreseeable medium to long term, the competition among large technology companies around large models will not cool down. This is a race without an endpoint in the short to medium term, and the competition for ranking is extremely fierce. Only by consistently maintaining a position among the leading players can one gain market recognition and attract capital.

Therefore, “running while replenishing energy” is the only choice.

Unlike the internet bubble around the year 2000, participants in the competition for large models all understand that while long-term cash burning is not a good commercialization narrative, it is a widespread consensus in the market. In a situation where both model capabilities and market scale are rapidly and intensely changing, staying in the game is the top priority. When people look forward to the systemic changes brought about by AI, creating a trillion-dollar market space, Zhihui bets that a larger market space will create more revenue growth opportunities and more significant scale effects. The strong foundational model capabilities and the MaaS business model imply a higher monetary multiplier. This is an arithmetic problem that everyone participating in the AI game can understand.

Burning money in an attempt to secure a place for Zhipu in an AI-dominated future.

The implied counterprice of professionalism, focus, and determination

AGI is a long struggle, as Zhang Peng, CEO of Zhipu, put it, like “running a marathon at the speed of a sprint.” This contradictory description explains both sides of Zhipu: it not only interprets Zhipu's technological determination but also explains the reason for choosing capitalization.

Founded in 2019, Zhipu originated from the Knowledge Engineering Laboratory of the Computer Science Department at Tsinghua University. Starting from the underlying architecture of the foundational large model, Zhipu has continuously been the first to launch several “first models” in China, accumulating a set of domestically produced original model combinations.

Take the base model GLM-4.5 released in July this year as an example. According to data from Frost & Sullivan, GLM-4.5 has achieved the following leading positions globally.

According to the evaluation results of 12 industry standard benchmark tests conducted in July 2025, GLM-4.5 ranks third globally, first in China, and tops among global open-source models.

In September 2025, according to the LLM hallucination rankings in the field of Retrieval-Augmented Generation (RAG), the hallucination rate of GLM-4.5 is the second lowest globally and the lowest in China.

Since the release of GLM-4.5, Zhiyu has consistently ranked among the top ten globally and the top three in China for token consumption on OpenRouter.

At the same time, Zhipu's paid API revenue on OpenRouter exceeded the total of all domestic models.

In the second week of December, Zhupu continuously open-sourced the GLM multimodal and agent series models for five days, occupying 5 seats in the top 10 of the Hugging Face Trending list.

This includes not only AutoGLM, which can replace users in operating their mobile phones, but also the GLM-4.6V multimodal large model, the GLM-ASR series of speech recognition models, the GLMTTS industrial-grade speech synthesis system, and the core technological achievements in video generation. At first glance, this operation seems almost contrary to the logic of emphasizing commercialization and profitability. However, at this sensitive time point just before going public, Zhipu's open-source strategy demonstrates a strong vision of a technology-centric artificial intelligence company. Promoting the prosperity of the entire AI technology stack and open-source community is not limited by time nodes, even during the critical phase when AI companies shift from technological competition to capital operations.

This temperament is precisely why Zhipu was initially referred to as China's “most like OpenAI company.”

With OpenAI recently suspending several non-core projects, including the Sora video generation model, and focusing all resources over an eight-week period on enhancing the performance and user experience of its core product, ChatGPT, the two companies have once again reverted to a single path in their model product strategy: the capabilities of the model itself are decisive, and all other matters should give way to it.

The only significant difference may be the huge valuation gap between Chinese and American AI companies. OpenAI's extremely high valuation and continuous funding ensure that it has a steady stream of resources for model development, even if it does not go public. In contrast, Chinese large model companies have valuations that are lower by more than an order of magnitude, even though Zhizhi has already become one of the best-performing companies among them. In its 6 years of establishment, Zhizhi has raised over 8 rounds of funding, with a total financing scale exceeding 8.3 billion yuan.

Now, Zhipu needs a broader market to raise more ample ammunition, and the capital market's response will confirm the quality of independent large model companies represented by Zhipu, thereby fully pricing the entire large model industry in China.

Future War

There is currently no independent large model company in the Chinese capital market that can be referenced, and various doubts surrounding the intense competition for large models still linger in the market. In the growth stage of technology companies, measuring their fundamentals and growth potential using financial profitability indicators is generally ineffective. The commercial value or growth logic of such companies is often assessed through multiple complex dimensions, including revenue performance, product capabilities, market space, and business models, as well as, most importantly, expectations for the future.

Tech optimists firmly believe in the future envisioned by artificial intelligence. Two days after Zhipu's public offering documents, another large model company, MiniMax, also disclosed its prospectus. Regardless of who ultimately succeeds in going public first, they are destined to leave a significant mark in the history of AI development.

AI is an intelligent infinite war about productivity and social change, and the battle has just begun.