Zhipu AI Lab IPO: From GLM-130B to $170B Valuation Surge
In early 2026, two prominent AI labs, MiniMax and Z.ai, made their initial public offerings (IPOs) on the Hong Kong stock exchange. Z.ai, also known as Zhipu (智譜), is particularly recognized for its GLM series of open-weight models. Often compared to Anthropic in the Chinese AI landscape, Zhipu has a rich history rooted in academic research and a strong drive towards commercialization.
Beginnings in Academia
The journey of Zhipu began with Zhang Peng, an alumnus of Tsinghua University. After graduating in 2002, Zhang joined Tsinghua's Knowledge Engineering Group (KEG) as a master's student. Established in 1996, KEG specialized in knowledge graphs and information extraction, initially focusing on semantic web, data processing, and domestic Java projects.
At KEG, Zhang collaborated with distinguished computer science professors like Tang Jie. In 2006, Tang Jie's team launched ArnetMiner (AMiner), a scientific and technology research intelligence mining program. AMiner functioned similarly to Google Scholar, building a searchable knowledge graph and ranking of researchers by extracting data like emails and profiles from unstructured web content using machine learning.
As AMiner matured, Zhang Peng, influenced by KEG's "Paper to Product" (P2P) philosophy, began exploring commercialization. The faculty encouraged this approach, emphasizing the inseparable link between AI and industry. By 2015, the AMiner team started generating revenue through partnerships with private companies, including Microsoft, Elsevier, Google, ExxonMobil, and IBM. The success of these partnerships fueled the idea of fully commercializing AMiner into a standalone company.
The Path to Commercialization
Several KEG lab members had already ventured into entrepreneurship, notably Tang Wenbin and Yang Mu, who co-founded Megvii, a facial recognition AI company known for its Face++ API and valued at $4-5 billion.
In 2015, Zhang Peng observed the excitement surrounding computer vision but also noted the ongoing debate about its potential for achieving Artificial General Intelligence (AGI). With the definition of AGI still unclear, Zhang decided to focus on "cognitive intelligence," aiming to enable machines to think like humans.
By 2016, Zhang Peng and his co-founders—Tang Jie, Liu Debing, Li Juanzi, Xu Bin, and Wang Shaolan—decided to spin off AMiner into a company. However, university policies at the time made it challenging to commercialize academic products. This changed in 2018 when China's Department of Education issued guidelines permitting the commercialization of research institution work. After extensive negotiations, Zhipu, meaning "Knowledge Atlas," was formally founded in June 2019. Tsinghua University transferred the AMiner product and remains a significant shareholder.
Initially, Zhang served as CTO, overseeing engineering and product delivery, while Tang, Li, and Xu focused on research. Liu managed university negotiations and intellectual property. Wang Shaolan was the founding CEO, a role Zhang Peng assumed in late 2022, with Wang transitioning to president.
The GPT-3 Breakthrough and GLM
The release of OpenAI's GPT-3 in 2020, with its 175 billion parameters, marked a significant turning point. Zhang recognized it as a breakthrough, demonstrating that traditional, complex Natural Language Processing (NLP) pipelines were not always necessary. GPT-3's ability to generate human-like text surpassed existing models like Google's BERT.
Zhang discussed GPT-3 with Professor Zhang Bo (no relation), who, despite acknowledging its novelty, pointed out its tendency to "spout total nonsense." Nevertheless, Zhang Peng, with support from Debing Liu, pushed his colleagues to train their own multi-billion parameter bilingual model, famously stating, "Unless you take the leap and make the pursuit, you’ll never know just how fast or how far you can run." Lacking detailed information on GPT-3's inner workings, they had to reverse-engineer much of it.
In late 2020, co-founder Tang Jie led a government-backed project called WuDao, involving a hundred scientists from various Chinese universities, to train ultra-large models. In March 2021, WuDao 1.0 unveiled four major models: - Wenyuan: A 2.6 billion parameter model with GPT-3-like performance in Chinese NLP tasks. - Wensu: A biology-oriented LLM. - Wenlan: Connected languages to images using 50 million web-scraped captions. - Wenhui: The most notable, an 11.5 billion parameter model approaching "Cognition," capable of complex reasoning, image generation, and poetry after fine-tuning.
The Wenhui model was trained using a pretraining framework developed by Tang Jie called "General Language Model" (GLM), first introduced in a March 2021 paper. GLM aimed to be an alternative to GPT and BERT, blending their approaches. While GPT generates text unidirectionally (left-to-right), and BERT fills in blanks bidirectionally, GLM fills in blanks token by token using surrounding context. Zhang noted this approach's stability during training, observing unexpected capabilities in understanding human input beyond 100 billion parameters. He cited an example where the model correctly divided by four to answer a simple math problem, indicating an understanding of intent rather than just rote multiplication.
Though WuDao was not explicitly a Zhipu project, Tang Jie's leadership and the GLM framework became central to Zhipu's foundational models.
Building a Business and GLM-130B
While developing GLM, Zhipu also created other AI products through a "Model-as-a-Service" (MaaS) framework launched in 2021. This led to diverse projects beyond AMiner, including: - Xiaoke: A physical robot for libraries and museums. - Hua Zhibing: China's first "virtual AI student" to enroll at Tsinghua University, powered by the 1.75 trillion parameter WuDao 2.0 model, with AI-generated image and voice. - Contribution to a digital sign language broadcasting system for the Winter Olympics.
The MaaS model aimed to deploy large models to customers via cloud API or on-premise. However, this "AI system integrator" approach mirrored the business models of older Chinese AI companies like Megvii and Sensetime, which had struggled to make a significant impact, dampening investor interest.
In December 2021, Zhang and his KEG lab team at Tsinghua decided to develop their next major model. Unlike the sparse mixture-of-experts (MoE) WuDao 2.0, which only partially activated its parameters, they aimed for a "dense" model like GPT-3, known for simpler architecture, consistent output, and easier deployment. Their goal was a model runnable on a single A100 GPU server.
The training began in January 2022, overcoming challenges like hardware failures, exploding model gradients, and unexpected memory usage. Zhipu contributed compute, and the final training run commenced in May 2022 on 96 A100 Nvidia servers, completing in July 2022. The compute cost was 4 million RMB ($600K), with an all-in cost of 10 million RMB ($1.5 million).
Despite the significant investment, Zhipu open-sourced the model, GLM-130B. Zhang explained that keeping it closed would not have generated immediate revenue, and they wanted to share it, reflecting the open-source spirit prevalent before ChatGPT. This move garnered a good reputation for Chinese AI efforts. In November 2022, Stanford University benchmarked GLM-130B as the only Chinese model among 30 prominent models, acknowledging its progress as an open model, though it underperformed closed models.
Before ChatGPT's release, investors were not particularly impressed with Zhipu's GPT-3-like model, with some even suggesting halving its valuation due to the weak economy. This changed dramatically on November 30, 2022.
The ChatGPT Moment and "Battle of the Hundred Models"
When ChatGPT launched, Zhang Peng was in Shenzhen, raising funds and securing contracts. While aware that a chat product was coming due to OpenAI's publications and connections with Tsinghua alumni, the scale of its impact was unexpected. AI researchers, who had worked in obscurity for a decade, suddenly found themselves at the center of a global "AI fever."
Tang Jie's online lecture, "Reflections on the Journey from 100-Billion-Parameter Models to ChatGPT," in late December 2022, attracted over a thousand Tsinghua alumni, eager to discuss data, applications, and robotics. For Zhang, this validated Zhipu's direction. As one of the few Chinese companies with its own LLM models, Zhipu's popularity surged, overwhelming its COO, Zhang Fan, with messages.
Internally, Zhipu worked through the Lunar New Year holiday to develop its own chatbot and open-source a smaller LLM, ChatGLM-6B, to allow Chinese users to explore a real-life model.
2023 was a period of both excitement and anxiety for Zhang Peng. The excitement stemmed from the immense opportunities, while anxiety arose from the "battle of the hundred models"—a chaotic influx of new AI models in Mainland China. At the start of 2023, only a few GPT-3 class LLMs existed (from Baidu, Huawei, ByteDance, Alibaba's DAMO Academy), all a generation behind GPT-3.5.
The following months saw numerous new startups and models emerge, including: - Baichuan AI: Founded in April by the former CEO of Sogou, focusing on healthcare. - Moonshot AI: Founded in March by Yang Zhilin (a former student of Tang Jie), known for the Kimi chatbot. - 01.AI: Founded by Kai-fu Lee (former head of Microsoft Research China and Google China), which raised significant early funding. - DeepSeek: Formally founded in July by hedge fund billionaire Liang Wenfeng, quickly gaining recognition for its coding model released in November.
A September 2023 survey identified over 130 models in Mainland China, excluding industry-specific variants, signaling an imminent shakeout and price war. Despite the intense competition, Zhipu launched its foundational chat models—ChatGLM, ChatGLM-2, and ChatGLM-3—and open-sourced ChatGLM-6B, which became popular. The company also secured substantial funding, completing multiple Series B rounds.
Shifting to Enterprise and International Prominence
In August 2023, Zhipu launched its consumer chat product, Zhipu Qingyan ("elegant words"). Despite being an early entrant, it lagged behind competitors like Baidu's Ernie bot and ByteDance's Doubao, which leveraged powerful existing distribution channels. Doubao eventually became China's leading chatbot with an estimated 324 million monthly active users (MAU). Other players like Moonshot's Kimi, DeepSeek, Tencent's Yuanbao, and Alibaba's Qwen also joined the fray, spending heavily on customer acquisition.
Qingyan never surpassed 11 million MAU. Zhipu's attempts to acquire users were met with limited success, as Chinese consumers were generally unwilling to pay for premium chatbot features. Recognizing that their academic background might not suit the aggressive consumer market, Zhipu doubled down on its "model-as-a-service" enterprise business. Zhang noted that while price-cutting exists in enterprise, it is less extreme than in consumer markets, as clients prioritize product quality over excessively low bids.
Zhipu gained international prominence in January 2024 with the release of GLM-4. This 200 billion parameter model (dropping "Chat" to signify general intelligence) featured agentic capabilities like Python, web browsing, and image generation. It was a closed model, accessible via API to developers. Evaluations estimated its performance at about 90% of GPT-4, with particular strength in Chinese-language tasks and alignment, though some criticized its closed-source nature and censorship.
In May 2024, DeepSeek launched its V2 model with significantly lower pricing, prompting Zhipu to cut prices for its GLM-3Turbo API. Alibaba Cloud, Baidu AI Cloud, and Tencent followed suit, initiating the anticipated price war.
The Path to Going Public
2025 began with significant events. In January, the United States added Zhipu to the Entity List, reportedly due to links with the Chinese military, restricting its access to certain US-origin software and products. A week later, the "DeepSeek crash" impacted markets, causing a reevaluation of Nvidia's stock and highlighting Jevon's Paradox.
DeepSeek's global prominence influenced Zhipu's subsequent model development. Doubts about whether a mixture-of-experts (MoE) structure could achieve GPT-4 level performance were dispelled by DeepSeek. Zhipu studied the model, made modifications, and released its first MoE model, GLM 4.5, on July 28, 2025. Notably, this was Zhipu's first open-weight frontier model.
Initially, GLM 4.5 was met with mild interest, but its quality surprised observers, quickly reaching first place in Hugging Face's rankings within 48 hours. Other companies, including American ones, began integrating it into their products. Zhang realized that new models no longer required grand announcements. Zhipu continued to release new models throughout 2025—GLM-4.6 and 4.7—via simple social media posts. These models were well-received for their coding ability and low cost.
With revenue growing at 100-150% annually (albeit from a low base), Zhipu went public in January 2026. Its stock price soared tenfold, reaching a peak valuation of $170 billion in June 2026, surpassing Xiaomi and BYD, and more than half of Alibaba's market cap. The market cap has since settled to around $50 billion. In July, the company sold shares to institutional investors to raise more capital.
Zhipu has continued to release highly regarded models, with GLM-5.3 and its Flash version garnering significant attention, though DeepSeek remains the leading Chinese AI lab. In September, Zhipu reported a 400% increase in six-month revenue to $140 million, with an annual recurring revenue (ARR) of approximately $1.6 billion on a monthly basis. The company is transitioning its revenue model from largely on-premise model deployment to cloud API, which is perceived as more scalable.
However, Zhipu faces a significant financial challenge: its R&D spending, largely on compute for training new models, is more than double its revenue. This immensely dilutive business model, where models are often given away for free, raises questions about its long-term sustainability.
Takeaways
- Zhipu originated from Tsinghua University’s Knowledge Engineering Group, spinning off the AMiner research platform into a commercial venture in 2019 while retaining strong academic ties.
- The company built the GLM series, open‑sourcing the 130‑billion‑parameter GLM‑130B model to gain credibility and ecosystem support despite its $600 K training cost.
- ChatGPT’s 2022 launch sparked a Chinese “hundred‑model” race, prompting Zhipu to release ChatGLM‑6B and later consumer chatbot Qingyan, though it lagged behind rivals in user adoption.
- Zhipu shifted focus to enterprise Model‑as‑a‑Service, launching GLM‑4 with agentic features and later the MoE GLM‑4.5 model, which quickly topped Hugging Face rankings and drove price‑war dynamics.
- The firm went public in early 2026, achieving a peak valuation of $170 billion, but its R&D spend now exceeds revenue, raising concerns about long‑term financial sustainability.
Frequently Asked Questions
What is the GLM framework and how does it differ from GPT and BERT?
GLM (General Language Model) is a pretraining architecture that fills in blanks token‑by‑token using surrounding context, combining GPT’s left‑to‑right generation with BERT’s bidirectional masking. Unlike GPT’s unidirectional output and BERT’s purely masked prediction, GLM’s hybrid approach yields more stable training and better intent understanding, as demonstrated by Zhipu’s early models.
Why did Zhipu open‑source the GLM‑130B model even though training cost exceeded $600 K?
Zhipu released GLM‑130B as open‑source to build credibility, attract ecosystem adoption, and align with the pre‑ChatGPT open‑AI culture, believing immediate revenue would be limited. By sharing the model, they gained international recognition, benchmark citations, and a stronger developer community, offsetting short‑term financial loss.
Who is Asianometry on YouTube?
Asianometry is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.
Does this page include the full transcript of the video?
Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.
Helpful resources related to this video
If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.
Links may be affiliate links. We only include resources that are genuinely relevant to the topic.