Introduction: A First for China’s AI Race
Chinese AI startup Zhipu AI is on the brink of making history as the first of China’s new wave of foundation model companies to go public. The Beijing-based firm – founded just six years ago – has initiated the IPO process with regulators, aiming to outpace its domestic peers in the race to capital markets . This move comes amid a fierce AI arms race in China, where startups and tech giants alike are vying to build the most advanced large-scale AI models. Zhipu’s imminent listing is more than just a financing event; it signals a coming-of-age for China’s AI sector, highlighting Beijing’s drive to foster homegrown AI champions that can rival Western players. By seeking an IPO now, Zhipu AI is poised to transform from a research-driven upstart into a publicly traded leader – a landmark moment underscoring the broader significance of China’s AI ambitions on the global stage .
Academic Origins at Tsinghua and the Wudao Legacy
Zhipu AI’s story has uniquely academic roots. The company was spun out of Tsinghua University’s Department of Computer Science in 2019 , beginning as a humble research lab project. Its co-founders, Tang Jie and Li Juanzi, are both esteemed Tsinghua professors who initially focused the startup on knowledge graphs and academic AI research . Early on, Tang Jie was also a leading figure at the Beijing Academy of Artificial Intelligence (BAAI), where he helped develop the groundbreaking “Wu Dao” model – a colossal 1.75-trillion-parameter AI system unveiled in 2021 that set records in scale . In fact, Tang and his team created Zhipu.AI in part to apply and extend the Wudao research beyond the lab . Zhipu even collaborated on high-profile demos like “Hua Zhibing”, China’s first AI-powered virtual student, which was jointly developed by BAAI, Zhipu and others as a showcase of the Wudao 2.0 model’s capabilities . These academic ties gave Zhipu a head start in China’s AI boom: the young company inherited cutting-edge research from Tsinghua and BAAI, positioning it to leap into the emerging large model race ahead of many corporate competitors. By 2020, Zhipu pivoted squarely to large-scale AI pre-training, recognizing the technology’s transformative potential well before the ChatGPT frenzy kicked off .
Founders and Vision: Academia Meets Ambition
At the helm of Zhipu AI is CEO Zhang Peng, a Tsinghua alumnus who represents the new generation of AI entrepreneur in China. Zhang joined the professors early on and now leads the company’s commercial strategy, all while maintaining the research-first ethos from its university days. He is openly ambitious – going so far as to predict that Zhipu’s models will match OpenAI’s GPT-4 by the end of the year and put the company “very close” to achieving artificial general intelligence . Internally, Zhang espouses a quasi-applied-science idealism; he has remarked to his team that “no matter how much money we raise or make, it will be a hindrance on our road to AGI,” emphasizing that short-term profits shouldn’t derail their long-term mission .
Meanwhile, co-founder Professor Tang Jie serves as the chief scientific mind and spiritual guide for Zhipu’s innovation. Tang is a renowned scholar (an IEEE, ACM, and AAAI Fellow) and currently Director of the Foundation Model Research Center at Tsinghua’s AI Institute . He retains a significant influence in Zhipu’s direction – the company is effectively controlled by Tang and Chairman Liu Debing, who together lead a consortium of founders and early stakeholders holding about 37% of voting rights . Tang’s vision of integrating massive pretrained models with knowledge and reasoning (as seen with Wudao) continues to drive Zhipu’s R&D trajectory . While Zhang Peng focuses on scaling the business and product offerings, Tang Jie ensures the scientific rigor and pioneering spirit remain at the core. This founder duo – backed by fellow Tsinghua professor Li Juanzi in the early days – exemplifies how China’s AI startups blend academia and industry. Their influence and vision have shaped Zhipu into a company that straddles cutting-edge research and real-world deployment, with an eye on ultimately achieving AGI.
AI Portfolio: GLM, ChatGLM, and an Open-Source Arsenal
Zhipu AI made its name through an impressive suite of homegrown AI models and products, many of which it has openly shared with the developer community. Starting in 2020, the team began developing the GLM (General Language Model) architecture, which led to one of China’s first truly large-scale pretrained language models. In 2022, Zhipu and Tsinghua’s lab jointly debuted GLM-130B, a 130-billion-parameter bilingual Chinese-English model that demonstrated China’s ability to build GPT-3 class models . This foundation paved the way for a rapid rollout of derivative models and applications. Today, Zhipu’s AI suite includes:
-
GLM & ChatGLM – The flagship family of large language models, including the massive GLM-130B and the open-source conversational AI ChatGLM. ChatGLM, first released in 2023, is a chatbot tuned for smooth bilingual dialogue. A lightweight 6-billion-parameter version (ChatGLM-6B) runs on a single GPU, making advanced AI accessible on consumer hardware . Zhipu claims its latest-generation model, GLM-4, now surpasses even OpenAI’s GPT-4 on several benchmarks – a bold assertion that speaks to its confidence in this core tech.
-
CodeGeeX – A multilingual code generation model (on the order of 13 billion parameters) that serves as an AI coding assistant. Launched in late 2022, CodeGeeX has since evolved through multiple iterations; by mid-2024 the fourth-generation CodeGeeX4-ALL-9B model was released open-source . It offers code completion, debugging, and even code interpretation across more than 20 programming languages. The tool has gained over 1 million individual users via plugins for VS Code and other IDEs, and an enterprise version has been adopted across various industries . This wide developer uptake highlights Zhipu’s success in nurturing a community around its AI offerings.
-
CogView & CogVLM – A pair of multimodal AI models pushing beyond text. CogView is Zhipu’s text-to-image generation model (analogous to DALL·E) capable of creating images from prompts in Chinese or English. Its later iterations (CogView2, CogView3) introduced advanced techniques like diffusion to improve image fidelity. CogVLM, on the other hand, is a visual-language model with 17 billion parameters (10B visual + 7B language) that can understand and describe images . Together, these models allow Zhipu to offer AI that sees and imagines, complementing its text-centric LLMs .
-
AutoGLM (AI Agents) – Most recently, Zhipu has been exploring autonomous agents that use its models to perform complex tasks. In late 2024 it introduced a mobile assistant app (under the AutoGLM moniker) that can obey voice commands to execute actions like shopping online or managing errands – pitched as a Chinese answer to on-device smart assistants . And in March 2025, Zhipu unveiled AutoGLM Rumination, a more advanced AI agent designed for deep research and autonomous decision-making . This agent can concurrently browse the web, read and analyze information, then generate detailed reports with citations – essentially acting like an AI research analyst . Notably, Zhipu has made AutoGLM Rumination available free of charge via its platforms, showcasing its strategy of seeding the market with powerful open tools .
Central to Zhipu’s product philosophy is a strong commitment to open-source AI. The company has repeatedly open-sourced major models under permissive licenses (MIT, etc.), from the original ChatGLM-6B to the newest GLM-4 series. In April 2025, Zhipu open-sourced its latest 32B and 9B parameter GLM models – releasing base, inference, and reasoning versions for free on its new developer platform (z.ai) . It even branded 2025 as its “Year of Open AI,” pledging to release more high-performance models to foster transparency and accessibility in the AI community . This open approach not only accelerates Zhipu’s R&D via community feedback, but also differentiates it from more guarded competitors. By sharing its crown jewels (from large language models to code generators) openly, Zhipu has cultivated goodwill among developers and positioned itself as a leader of China’s open-source AI movement. Internally, the company notes that one of its new open models, GLM-Z1-32B-0414, achieves inference speeds up to 200 tokens/second – making it the fastest in China – while costing just 1/30th the price of a rival model (DeepSeek’s R1) to run . Such metrics underscore how Zhipu leverages openness and efficiency as competitive advantages.
Funding Journey: From Angel Round to Tech Unicorn
Backing Zhipu AI’s rise is a who’s-who lineup of Chinese tech giants, global investors, and state-sponsored funds. In its early days, however, funding did not come easy. As an academic spin-off in 2019, Zhipu initially struggled to find investors until it proved the promise of large models. A small angel round in 2021 (around $15 million from local VCs) finally provided a breakthrough after the startup’s early pivot to LLMs bore fruit . From there, the fundraising momentum only grew. By 2023, Zhipu AI had reportedly raised a cumulative ¥2.5 billion (≈$350 million) with the help of internet giants Alibaba and Tencent . Indeed, that year saw China’s tech heavyweights crowd onto Zhipu’s cap table: Alibaba, Tencent, Ant Group, Xiaomi, and Meituan all became stakeholders . At the same time, top-tier venture capital firms joined in – Sequoia Capital China (HongShan), Qiming Venture Partners, Legend Capital, Hillhouse’s GL Ventures, Lightspeed China, and others have all bet on Zhipu . This diverse investor mix reflects the company’s strategic value: both China’s private sector and the government see Zhipu as a key player in the AI race worth funding.
Government-linked funding, in particular, kicked into high gear in late 2023 and early 2024. Zhipu secured three consecutive state-backed investments in the span of a few weeks . Local governments are investing aggressively: for example, the Chengdu city government injected ¥300 million (≈$42 million) in one recent round , and state-guided funds from Beijing, Hangzhou, and other tech hubs have contributed large sums . By September 2024, Zhipu’s valuation had exceeded ¥20 billion ($2.8 billion) according to lead investors , making it a solid unicorn. That same year it attracted international capital as well – notably a $400 million investment from Saudi Arabia’s Prosperity7 (Aramco’s venture arm) in mid-2024 . And as of March 2025, just before the IPO push, Zhipu raised another ¥1 billion ($137 million) in fresh funding led by state-backed investors in Hangzhou .
All told, Zhipu AI has amassed over ¥10–16 billion in funding (estimates vary, roughly $1.4–$2.2 billion) across ~15 rounds since inception . Its shareholder roster balances government stakeholders (e.g. Beijing’s Zhongguancun Science City fund, the national AI Industry Investment Fund, and municipal funds from Chengdu, Zhuhai, etc.) with China’s tech giants and elite VCs . This gives Zhipu both the political backing and private-sector resources to pursue long-term AI research. Investors, for their part, are drawn to Zhipu’s rapid progress and the strategic importance of having an indigenous AI model platform. The company’s relatively distributed ownership and joint founder control are seen as positives for corporate governance, increasing its appeal to institutional investors . In short, Zhipu’s rise from a scrappy lab venture to a multibillion-dollar “AI tiger” has been fueled by a cascade of capital – a sign of the tremendous confidence placed in its technology and vision.
Strategy: Balancing Enterprise and Consumer AI
With substantial R&D funding in hand, Zhipu AI has been experimenting with how to monetize and deploy its AI at scale. The company’s commercialization strategy appears to straddle both enterprise (ToB) and consumer (ToC) markets – a dual approach not unlike that of OpenAI itself. “Our aim is to become China’s version of OpenAI, with operations spanning both business and consumer sectors,” CEO Zhang Peng told Caixin recently . On the consumer side, Zhipu has rolled out apps and services to build a broad user base. Its ChatGLM chatbot was made freely available to the public, reportedly gaining 25 million users for its chatbot app by early 2025 . Likewise, the newly launched AutoGLM agent is offered free of charge via Zhipu’s website, mobile app, and PC client – a tactic to quickly acquire users and iterate with real-world feedback. By prioritizing open access, Zhipu has managed to achieve significant adoption; for example, the open-source ChatGLM-6B model saw over ten million downloads shortly after release . This user growth, while not immediately lucrative, seeds an ecosystem that Zhipu can later monetize through premium services or partnerships.
On the enterprise front, Zhipu is actively cultivating a B2B business in model-as-a-service and bespoke AI solutions. The company launched its Zhipu AI Open Platform (Z.ai), where businesses and developers can access its large models via APIs, cloud services, or on-premise deployments . With its models spanning language, code, and multimodal tasks, Zhipu can cater to a variety of use cases. Already, ChatGLM has been widely used for customer service chatbots and education tools in China , demonstrating value in handling bilingual queries for companies and institutions. The firm also provides an enterprise edition of CodeGeeX for corporate coding assistance, which has been adopted in sectors like software and finance . More generally, Zhipu’s enterprise clients are exploring applications of its AI in healthcare, education, finance, and government services, mirroring the trend seen by peers in the industry . To support these deployments, Zhipu offers model fine-tuning, training support, and consulting to tailor its foundation models to each client’s needs. Notably, its models can be run at a fraction of the cost of U.S. alternatives (as with the 1/30 cost claim versus GPT-4 or DeepSeek models) , a compelling selling point for budget-conscious enterprises or those concerned about data sovereignty.
Zhipu AI is also building out a developer ecosystem around its technology. By open-sourcing code and models, it encourages developers to experiment and build applications on top of its GLM models. The company has forged partnerships with hardware makers like Huawei and Qualcomm to optimize its models for local chips – a strategic move given U.S. export restrictions on high-end GPUs. In mid-2024, when OpenAI’s API access became limited in certain regions, Zhipu launched a “Special Migration Program” to woo developers over to its platform , offering support to port their applications from GPT-4 to ChatGLM. Such moves indicate an aggressive push to capture the domestic developer mindshare. Zhipu is effectively nurturing a full-stack AI ecosystem: open research to attract academia, free consumer apps to attract users, and enterprise services to generate revenue. While its current annual revenue (reportedly just over ¥10 million, or ~$1.4M, in recurring revenue as of 2024 ) is modest, the company is playing the long game. The bet is that by cementing itself as the go-to Chinese AI platform – trusted by government, integrated by companies, and loved by developers – significant monetization will follow. This approach is backed by its investors and by Chinese authorities, who see strategic importance in mass adoption of indigenous AI. As one early VC backer noted, the large-model race in China is still fluid: some players will fall away, but those that build robust ecosystems now could surge ahead and define the market’s future .
Competition: Standing Out Among the “Six AI Tigers”