With a market capitalization of HK$57.9 billion! Knowledge Atlas has secured the title of the "world's first large model stock" to go public! Based on the issuance valuation, investment institutions that participated in its three rounds comprising 14 financings achieved an average paper gain of 3.4 times, including Meituan, Ant Group, and Tencent Investment. Following the listing, coupled with the multiplier effect of the rising share price, the profits for these investment institutions are set to increase further.
However, alongside rapid revenue growth, structural issues within Knowledge Atlas's business model have begun to surface. While it relies on on-premise deployment as its primary revenue pillar, and its cloud deployment business has experienced rapid growth in recent years, intense competition in the cloud sector has driven down unit prices. Compounded by a significant surge in computing power service fees, which exceeded 1.5 billion yuan annually, this has significantly dragged down gross margins and weakened profitability.
As business volume expanded, Knowledge Atlas's losses also continued to widen, reaching nearly 3 billion yuan in 2024. By the end of June 2055, its liabilities had climbed to 11.252 billion yuan, with negative net assets of -6.151 billion yuan. While the HK$4.3 billion raised from the IPO will alleviate its debt pressure, the company will still require continuous capital infusion. The uncertainty surrounding when it will reach an inflection point for self-sustaining profitability raises questions about its subsequent valuation.
On January 8, 2026, the Hong Kong stock market welcomed Knowledge Atlas (02513.HK), the "world's first large model stock." Its share price initially fell below the issue price after opening but quickly rebounded, finally closing at HK$131.5 per share, up 13.17%, resulting in a market cap of HK$57.9 billion. CICC acted as the exclusive sponsor and overall coordinator for its IPO, with Huatai International, BOCOM International, Guotai Haitong International, and China Merchants Securities International jointly serving as overall coordinators and joint bookrunners.
For this IPO, Knowledge Atlas issued 37.4195 million H shares at an issue price of HK$116.2 per share. This comprised a Hong Kong public offering of 7.4839 million shares and an international offering of 29.9356 million shares, with an over-allotment option of 5.6129 million shares, raising a total of HK$4.348 billion.
Among the companies listed on the Hong Kong stock exchange in January 2026, its fundraising scale ranked fourth, lower than Biren Technology's HK$4.55 billion, OmniVision Technologies' HK$4.8 billion, and GigaDevice's HK$4.68 billion.
Seventy percent of the net proceeds from Knowledge Atlas's IPO will be used for R&D investment in general-purpose AI large models, further consolidating its market competitiveness in general-purpose foundation models. Approximately 10% will be allocated to continuously optimizing the company's MaaS (Model as a Service) platform, including providing the latest foundation models, pre-training tools, and infrastructure development.
Knowledge Atlas's IPO valuation exceeded HK$51.1 billion, which is 2.1 times its post-money valuation of 24.4 billion yuan after the last financing round in May 2025. This means investors who entered in the final round achieved a 2.1x return in just over six months. For instance, in May 2025, Shanghai Wisdom Navigation, whose LP shares are 79.84% held by the Shanghai Pudong State-owned Assets Supervision and Administration Commission, exclusively participated in Knowledge Atlas's B6-4 round, investing 500 million yuan. Just over six months later, the market value of its shares reached 1.05 billion yuan. As Knowledge Atlas's share price rises post-listing, the market value and investment returns for these investors will correspondingly increase.
So, is Knowledge Atlas's valuation high? Let's measure it using the price-to-sales (P/S) ratio. Its IPO valuation corresponds to a P/S ratio of approximately 147 times based on 2024 revenue, slightly higher than the 122.6x P/S ratio of domestic GPU company Moore Thread at its IPO. If estimated based on the 2024 growth rate, Knowledge Atlas's full-year revenue for 2025 could reach 1.33 billion yuan, which would correspond to a P/S ratio of 38.4 times.
The genuine confidence in the investment target is reflected by real financial commitments. Knowledge Atlas's IPO is backed by a powerful consortium of cornerstone investors, including JSC International Investment Fund SPC, JinYi Capital Multi-Strategy Fund SPC, Perseverance Asset Management, Shanghai Gaoyi & CICC Financial Trading Limited, WT Asset Management, Taikang Life Insurance, GF Fund Management, 3W Fund, Wusong, Optimas Capital Limited, and Lingyun Guang Technology International Co., Ltd., totaling 11 investment institutions.
These cornerstone investors collectively subscribed to shares worth HK$2.984 billion, accounting for 68.63% of Knowledge Atlas's intended total fundraising. This indicates that the cornerstone investors provide a high degree of certainty for the company's post-listing trading base.
Knowledge Atlas's IPO represents a triumph for its shareholders. Beyond the soaring wealth of the founding team and employees, external investor shareholders from its three rounds and 14 financings have all reaped substantial gains.
Calculated at the issue price, the 57 external investors held shares with a market value of HK$31.357 billion, compared to a total investment amount of 8.36 billion yuan, resulting in an overall paper gain of 3.4 times. Factoring in the 13.17% closing gain on the IPO day, the overall paper gain increases to 3.85 times.
Breaking it down, various investment institutions achieved respectable returns. For example, Junlian Xiangdao, which followed with investments six times, invested a total of 454.7225 million yuan, and its shares were worth HK$2.168 billion at IPO, representing a paper gain of approximately 3.3 times. Ant Group's subsidiaries Shanghai Yunya and Shanghai Feiya invested a total of 600 million yuan, with their shares valued at HK$1.871 billion at IPO, yielding a paper gain of about 2 times. Meituan's subsidiary Tianjin Sankuai invested 300 million yuan, and its stake was worth HK$1.999 billion at IPO, resulting in a paper gain of approximately 5.7 times. Tencent Investment invested 200 million yuan, with its shares valued at HK$808 million at IPO, a paper gain of 3.6 times. Beijing Shunying, controlled by Lei Jun and others, invested 150 million yuan, and its shares were worth HK$603 million at IPO, a paper gain of 2.6 times. Quande Meijia, controlled by Capital Today's Xu Xin, invested 255.3 million yuan, and its stake was valued at HK$1.313 billion at IPO, a paper gain of 3.6 times. Such investment returns are likely one reason why Knowledge Atlas attracted these cornerstone investors.
Founded in 2019, Knowledge Atlas immediately commenced development of general-purpose language models. In 2021, it released its first pre-trained large model with billions of parameters. Since then, it has continuously iterated and expanded its product line, now boasting a comprehensive model matrix including foundation models, reflection models, multimodal models, and AI agents, capable of providing integrated AI services encompassing computing, networking, training, communication, and inference acceleration.
Knowledge Atlas has capitalized on the pivotal moment where AI is profoundly reshaping global economic growth, business operations, and human life. The industry anticipates that by 2030, AI will empower at least 20% of daily business decisions globally, support mainstream intelligent devices for at least 80% of global consumers, and create an AI economy exceeding $20 trillion.
China's AI market size also grew from 93.7 billion yuan in 2022 to 160.7 billion yuan in 2024, with a compound annual growth rate (CAGR) of 31%. It is projected to expand to 993 billion yuan by 2030, with a CAGR of 35.5% during that period, outlining the future growth trajectory for Knowledge Atlas.
In 2024, Knowledge Atlas achieved revenue of 312 million yuan, ranking first among independent general-purpose large model developers in China and second among all general-purpose large model developers, capturing a 6.6% market share. Following its Hong Kong listing, its financial strength and international influence are expected to increase further.
Knowledge Atlas serves customers through its integrated MaaS platform, featuring a comprehensive large model product matrix. First, foundation models. Its flagship foundation model GLM-4.5 was open-sourced upon release, with a parameter scale of 355 billion. The lighter version, GLM-4.5-Air, has 106 billion parameters. In September 2025, Knowledge Atlas released the upgraded foundation model GLM-4.6 with enhanced coding capabilities, ranking first globally on the authoritative evaluation system CodeArena. Second, reflection and rumination models. Building on the foundation models, it developed the reflection model (GLM-Z1) and the rumination model (GLM-Z1-Rumination), which incorporate deep thinking capabilities. Third, multimodal models. Its CogView (image generation), GLM-4.5V (visual understanding and reasoning), CogVideoX (video generation), GLM-Realtime (real-time video call), and GLM-4-Voice (voice model) can process and integrate information from different modalities like text, images, audio, and video. Fourth, AI Agents, which can autonomously execute multi-step tasks without continuous human input. Knowledge Atlas's foundational AI agent model, AutoGLM, can translate human-like reasoning into concrete actions. In August 2025, it released AutoGLM2.0, capable of simulating human operations across a wider range of mobile applications and websites, performing specified tasks autonomously in the cloud without occupying the user's phone or computer. Building on this, Knowledge Atlas launched the advanced version AutoGLM Rumination, an autonomous AI agent designed for exploring open-ended propositions and taking action based on findings, integrating the reasoning power of the GLM-Z1-Rumination model with the interactive capabilities of AutoGLM, enabling "thinking while doing."
According to the large language model hallucination rate ranking published on GitHub, evaluated by the Hughes Hallucination Evaluation Model (HHEM-2.3), Knowledge Atlas's GLM-4.5-Air (released July 2025) and GLM-4.6 had hallucination rates of 9.3% and 9.5% respectively, placing them at relatively low levels. In September 2025, based on the LLM hallucination ranking in the Retrieval-Augmented Generation (RAG) domain, GLM-4.5's hallucination rate was the second lowest globally and the lowest in China. This benchmark tests by posing deliberately misleading questions to large models and evaluating them based on the frequency of generating non-existent answers (i.e., hallucinations).
Currently, Knowledge Atlas's models and agents are applied across various sectors including internet and traditional industries. First, applied in the technology and internet industry, for instance, helping platforms like Kingsoft Office, Zhaopin.com, and NieTa improve service quality, reduce operating costs, and enhance user experience. Second, applied in traditional manufacturing, for example, by deploying the GLM series models, Knowledge Atlas has helped an automotive manufacturer's smart cockpit system evolve its interaction capability from simple Q&A to intuitive and natural communication, possessing personalization, emotional depth, and adaptive interaction capabilities. Third, applied in the retail industry, for instance, Knowledge Atlas assisted Mengniu Dairy in creating the AI nutritionist "Meng Meng," providing consumers with expert-level, personalized nutrition and health services. The continuous rollout of products and services has driven the rapid revenue growth of Knowledge Atlas.
From 2022 to 2024, Knowledge Atlas's revenue increased from 57 million yuan to 312 million yuan. In 2025, its revenue continued to grow, reaching 191 million yuan in the first half of the year, a year-on-year increase of 3.2 times.
Knowledge Atlas's revenue primarily comes from two main deployment businesses for large models.
First, the core revenue stems from localized deployment business, mainly targeting B-end clients with strict data security requirements like government, finance, and energy sectors. It adopts a dual-drive model of "open source + commercialization," providing clients with private, exclusive AI model solutions. Revenue from this business grew from 54.815 million yuan in 2022 to 264 million yuan in 2024. In 2025, this business continued its rapid growth, generating revenue of 162 million yuan in the first half, a year-on-year increase of 5 times. The localized deployment business has consistently maintained a gross margin exceeding 50%, reaching a high of 66% in 2024, and slightly decreasing to 59.1% in the first half of 2025, contributing stable cash flow for the company. Second, the cloud deployment business, which contributes a smaller proportion. Revenue from this business rose from 25.94 million yuan in 2022 to 485 million yuan in 2024, increasing its share of total revenue from 4.5% to 15.5%, becoming its second growth curve. In the first half of 2025, this business generated revenue of 290 million yuan, a year-on-year increase of 60.7%. Due to advantages like lower entry barriers and costs for enterprises, flexibility, speed, and rapid iteration, its revenue has grown rapidly.
However, with the increase in cloud deployment clients, its token consumption (the basic unit for AI model data processing) has skyrocketed. Daily average consumption surged from 500 million to 4.6 trillion by the end of June 2025, and reached 4.2 trillion in November 2025.
It is worth noting that in 2024, China's large language model market size reached 5.3 billion yuan, with institutional clients contributing 4.7 billion yuan and individual clients 600 million yuan, meaning institutional procurement accounted for 89%. The market expects China's large language model market size to grow to 101.1 billion yuan by 2030, with a CAGR of 63.5% from 2024 to 2030. Institutional clients are expected to remain the core driving force in this process.
Therefore, the acquisition and retention of institutional clients are key indicators of a large model enterprise's competitiveness. From 2022 to 2024, Knowledge Atlas's number of institutional clients grew from 48 to 3,156, mainly driven by a significant increase in cloud deployment clients, which nearly doubled year-on-year in 2024. In 2025, its institutional client base accelerated, exceeding 12,000 by September 2025, a nearly threefold increase from 3,061 at the end of June.
However, as market competition intensifies, Knowledge Atlas's cloud deployment business also faces profitability challenges. Firstly, rising costs have caused its gross margin to drop from 76.1% in 2022 to 3.4% in 2024, and further to -0.4% in the first half of 2025. The prospectus explains this as due to "continuously and strategically reducing service prices and顺应ing to market trends," essentially trading lower prices for market share.
So, with industry competition expected to remain fierce in 2026, can Knowledge Atlas maintain its competitiveness in the cloud market? A more objective assessment might be possible by examining its technical strength and R&D investment.
Knowledge Atlas is a unicorn that took flight on the shoulders of giants. Its technological foundation originates from the Knowledge Engineering Group (KEG) laboratory established in 1996 within Tsinghua University's Computer Science Department, one of China's earliest laboratories engaged in natural language processing and knowledge graph research. Its core member, Tang Jie, developed the AMiner platform in 2006, which became the prototype for Knowledge Atlas's SaaS (Software as a Service) platform.
After more than 20 years of accumulation, in 2019, Tang Jie, along with Tsinghua alumni Liu Debing, Zhang Peng, Li Juanzi, Xu Bin, and others, co-founded Knowledge Atlas with a registered capital of 10 million yuan, aiming to industrialize the laboratory's research成果. Consequently, Knowledge Atlas's core founding team consists entirely of data scientists and engineers,堪称 China's AI large model "Dream Team."
Tsinghua University's KEG remains Knowledge Atlas's strongest technical backbone. As stated in the prospectus, "Our core technical team maintains a close and enduring partnership with KEG, forming a unique and powerful channel for technical exchange. This connection is rooted in a shared academic lineage and helps inject continuity and coherence into our technology roadmap. This enables us to systematically collaborate with KEG to achieve cutting-edge research results."
Knowledge Atlas's Chief Scientist, Zhang Bo, is an academician of the Chinese Academy of Sciences, a foreign academician of the Russian Academy of Natural Sciences, and the Dean, Professor, and Doctoral Supervisor of Tsinghua University's AI Research Institute. He is one of the founders of AI research in China, having published over 400 papers in the AI field, with profound积累 in fundamental AI theories and applications like pattern recognition, knowledge engineering, and robotics, playing a key role in the company's establishment and subsequent product development.
Chairman Liu Debing, a student of Chinese Academy of Engineering academician and AI research expert Gao Wen, holds over 50 invention patents globally and has played a leadership role in setting the company's technological innovation and strategic research direction.
Co-founder and CEO Zhang Peng focuses on knowledge graphs and large-scale pre-trained models and is a core contributor to the development of Knowledge Atlas's GLM model series and AMiner.
Co-founder Li Juanzi is a top expert in semantic content management and text and social network mining, bridging fundamental AI research and practical system deployment.
Co-founder Xu Bin specializes in knowledge graphs, data mining, and AI, developing large-scale educational knowledge graphs and scalable technologies for scientific intelligence mining and knowledge services.
With such a background, the founding team of Knowledge Atlas views R&D as the primary driver of the company's development, stating that "R&D permeates everything we do." This is reflected in three main characteristics of its R&D.
First, its R&D expenditure far exceeds its revenue. In 2024, its R&D expenditure was 2.195 billion yuan, a threefold increase year-on-year, which is 7 times its 2024 revenue of 312 million yuan. In 2025, it continued to increase R&D investment, spending 1.595 billion yuan in the first half, which is 8.3 times its revenue for that period.
Second, its R&D team constitutes a high proportion of total staff. As of the end of June 2025, Knowledge Atlas's R&D team had 657 people, accounting for 74.4% of its total employees.
Third, it possesses 86 registered patents and 234 patent applications in China, along with 160 copyrights, 314 trademarks, and 58 domain names.
However, with the rapid growth in the size of the R&D team and investment, the structure of its R&D expenditure has also undergone significant changes.
As business scale expanded, Knowledge Atlas's expense支出 also increased significantly, growing from 132 million yuan in 2022 to 2.716 billion yuan in 2024, a 20.5-fold amplification over two years. R&D expenditure constituted the bulk, accounting for 81%, and showed the fastest growth rate.
Breaking down its R&D expenditure, Knowledge Atlas's R&D cost structure changed considerably over three years, shifting from being dominated by salaries to being dominated by computing power costs.
In 2022, the total salary expenditure for the R&D team was 420 million yuan, accounting for 49.8% of its R&D支出. By 2023, its computing power service fees reached 1.37 billion yuan, a year-on-year increase of 2.26 times, accounting for 58.9% of R&D支出 and becoming the largest expense item. In 2024, its computing power service fees further skyrocketed to 1.553 billion yuan, accounting for 70.7% of R&D支出. In the first half of 2025, its computing power costs continued to surge, reaching 1.145 billion yuan, a 90% year-on-year increase, accounting for 71.8% of R&D expenditure, the highest proportion historically.
The prospectus explains that iterating foundation models and investing in more advanced model training infrastructure led to an increase of 542 million yuan in computing power service fees paid to third-party providers.
Furthermore, due to the迅猛 growth in computing power procurement, Knowledge Atlas prepaid some fees. In 2024, prepaid computing power service fees were 39.522 million yuan, which increased significantly to 611 million yuan in the first half of 2025, highlighting the substantial burden of computing power costs and the tight demand for computing power faced by large model enterprises.
High computing power costs will also be a key constraint hindering the expansion of Knowledge Atlas's cloud business.
As a technology innovation enterprise, beyond its substantial R&D investments, Knowledge Atlas has over half of its employees holding shares, firmly aligning employee interests with company growth. The successful IPO is likely to create a group of new millionaires.
As of the end of June 2025, Knowledge Atlas had 883 employees, of whom 452 held shares, representing a high proportion of 51.2%.
These employees hold company shares through two employee stock ownership platforms: Hui Hui and Zhi Deng, both established in June 2021.
The GP (General Partner) of Hui Hui is Liu Debing, the current Chairman of Knowledge Atlas, holding 30.33% of the partnership interests. There are 426 LPs (Limited Partners) in total, among whom Zhang Peng and Executive Director Zhang Xiaohan hold 20.98% and 0.46% of the partnership interests, respectively.
The GP of Zhi Deng is Liu Debing, who holds 39.01% of the partnership interests. The 25 LPs are all company employees and advisors, among whom Executive Director Zhang Peng holds 4.63% of the partnership interests. The 16 advisors are all full-time interns, engaged by the company as algorithm experts (independent third parties).
As of the prospectus submission date, the two employee持股 platforms collectively held 6.6679 million shares, representing 16.55% of the company's total shares, and had signed a concerted action agreement with the founding team.
Liu Debing, as the common and sole GP, holds absolute control over Hui Hui and Zhi Deng. No single LP among them can influence this control.
Knowledge Atlas achieved its listing just over six years after its founding, leading to a significant increase in the market value of the holdings of Hui Hui and Zhi Deng. Based on the issuance market capitalization, Hui Hui's holding market value is HK$4.586 billion. Excluding the 51.77% partnership interests held by Liu Debing and two others, the remaining 48.23% interests are held by 424 employees, resulting in an average per capita holding value of HK$5.217 million. Zhi Deng's holding market value is HK$3.16 billion. Twenty-four employees hold 56.36% of the interests, resulting in an average per capita holding value of HK$74.207 million.
The wealth of the founding shareholders has also multiplied. Among them, Beijing Lianpai, in which Liu Debing holds a 92.7% equity stake, holds a 7.73% interest in Knowledge Atlas. Based on the issuance market cap, this stake is worth HK$3.663 billion. Adding his direct holding value of HK$107 million, his total holding value reaches approximately HK$3.77 billion. Tang Jie's holding value also reached HK$3.119 billion.
External fundraising capability is an indicator of a high-tech startup's technical strength and competitiveness. Knowledge Atlas attracted industry attention from its inception. From January 2022 to May 2025, over approximately three and a half years, Knowledge Atlas conducted three rounds comprising 14 financings, raising a total of over 8.3 billion yuan.
Its investor lineup is impressive, including Ant Group (via Shanghai Yunya, Shanghai Feiya), Tencent Investment, Fortune Capital, HuaKong Capital, Zhongguancun Science City, Zhuhai Huafa, the Haihe Fuxin Youda Fund controlled by Tianjin Wuqing State-owned Assets, the Artificial Intelligence Fund controlled by Beijing SASAC, Hangzhou City Investment Industrial Fund, the Daxing Industrial Fund controlled by Beijing Daxing District SASAC, Beijing Shunying controlled by Lei Jun and others, Quande Meijia controlled by Xu Xin, among others. These institutions collectively contributed 8.37 billion yuan.
Behind the attraction of nearly ten billion yuan from this豪华 investor consortium, each major product release or commercialization milestone for Knowledge Atlas was accompanied by capital infusion and valuation increases.
Its post-Series A valuation reached 665 million yuan, doubling from the angel round valuation of 387 million yuan, primarily because its early knowledge graph-related products began generating stable revenue, and its R&D team had initiated pre-training work on large language models.
In September 2021, Knowledge Atlas released its first 50-billion-parameter pre-trained large model GLM-10B, enhancing its market recognition and driving its Series B1 valuation to 2.1 billion yuan.
In August 2022, Knowledge Atlas released the open-source large model GLM-130B, and in September 2022, it released the high-performance code model CodeGeeX, expanding its product portfolio and client base, pushing its Series B2 valuation to 3.2 billion yuan.
In March 2023, it released the 1-billion-parameter foundation model ChatGLM and open-sourced ChatGLM-6B, driving its Series B3 valuation to 4.572 billion yuan. The increased market impact following these releases also raised its Series B4 financing valuation to 7.228 billion yuan.
In August 2024, Knowledge Atlas released one of China's first batch of large model products to pass regulatory filing—the generative AI assistant ZhiPu QingYan, supporting universal Q&A, multimodal understanding and generation, customized agents, and other broad application scenarios, lifting its Series B5 valuation to over 13.36 billion yuan.
Following the releases of GLM-4 and GLM-4V, Knowledge Atlas's Series B6 valuation steadily climbed, finally reaching 24.4 billion yuan after the May 2025 financing round.
In 2025, while completing the largest Series B6 financing round (4.377 billion yuan), Knowledge Atlas accelerated its listing process.
In March 2025, Knowledge Atlas was restructured into a joint-stock limited company with a registered capital of 362.244 billion yuan. On June 28, 2025, Knowledge Atlas further split its shares, changing the par value from 1 yuan to 0.1 yuan, resulting in a registered capital of 402.8 million yuan, divided into 4.028 billion shares.
Large model development and operation are extremely capital-intensive businesses. Even with nearly ten billion yuan raised, Knowledge Atlas still faces the situation of continuous losses and an expanding debt scale.
Continuous reliance on external financing for capital infusion is a common survival mode for most high-tech enterprises in their startup phase. From 2022 to 2024, Knowledge Atlas reported losses of 144 million yuan, 788 million yuan, and 2.958 billion yuan, respectively. In the first half of 2025, the loss was 2.358 billion yuan. The loss amount continues to widen, indicating the company still lacks self-sustaining profitability.
Its debt scale has also expanded急剧. Total liabilities surged from 542 million yuan in 2022 to 8.331 billion yuan in 2024, and further climbed to 11.252 billion yuan in the first half of 2025.
Under high debt, Knowledge Atlas's financial statements were already "insolvent" (liabilities exceeding assets), with negative net assets of -6.151 billion yuan as of the end of June 2025. Even after the HK$4.3 billion IPO融资, its debt burden remains substantial.
Post-listing, Knowledge Atlas still faces the test of multiple balancing acts. First, balancing the intensity and quality of R&D investment to ensure the commercial转化 of technology and thereby maintain market competitiveness. Second, balancing the growth rates of the localized deployment business and the cloud deployment business, using the high-margin localized business to secure the cash flow foundation, while using the high-cost, low-margin cloud business to increase market share. Third, balancing the speed of business scale expansion with the scale of assets and liabilities, gradually reducing the debt ratio.
When will Knowledge Atlas achieve economies of scale and reach the profitability inflection point? This is the puzzle it leaves for the market.
Comments