The Startup Street has become bustling again.
Recently, I've heard two pieces of news: First, there are approximately 20,000 investment promotion personnel from other regions active in Haidian, a district covering an area of over 400 square kilometers every month. Second, Wang Xingxing decided to open a branch in Haidian because he wanted to be closer to a certain robotics unicorn company in the district.
These two stories point to the same fact: everyone has their eyes on the enterprises in Haidian. Investment promotion personnel are targeting mature projects, while top unicorn founders are eyeing the most competitive enterprises in Haidian. Ultimately, people's attention is fixed on Haidian because it is the cradle of technological innovation. Since Peking University Science Park sowed the first seeds in 1992, followed by the establishment of Tsinghua Science Park in 1994, and the subsequent rise of Dongsheng Building and Zhongguancun Intelligent Manufacturing Street after 2000, over the past three decades, university laboratories, corporate R & D centers, and cafes in Wudaokou have intertwined to form a rich innovation ecosystem in Haidian.
The deep integration of "industry, academia, research, and application" has not only given birth to numerous unicorns in Haidian but also continuously nurtured disruptive source technologies. In the past, many internet giants such as Wang Xing, Su Hua, and Cheng Yixiao started and grew here. Today, a new generation of AI unicorns, including Zhipu, Darkside of the Moon, Mianbi Intelligence, and Galaxy Universal, have become the new calling cards of Haidian. Zhipu AI is located on Zhongguancun East Road, the Beijing headquarters of DeepSeek is in Rongke Information Center, and the headquarters of Darkside of the Moon is in JD Technology Building on Zhichun Road. The AI Origin Community Exhibition Hall on the first floor of Dongsheng Building shows that there are over 100 AI companies, making it the core of the "Zhongguancun AI Large - Model Industry Agglomeration Area".
Data provides a more straightforward view. The "Haidian: The Origin - 2025 Haidian AI Innovation Handbook" released by Zhongguancun Science City shows that with only 2.6% of Beijing's land area, Haidian generates over a quarter of the city's GDP, and is home to over 70% of the country's AI companies and over 80% of the world's top AI scholars. As of June 2025, 89 large models have been registered in Haidian, accounting for one - third of the national total and nearly 70% of Beijing's total.
As large models become the new infrastructure, AI applications are seen by investors as a market opportunity worth hundreds of billions or even trillions of dollars. With the upsurge of AI application startups, the once - quiet Zhongguancun Startup Street has become bustling again. You can attend a roadshow by Miracle Plus and then chat with entrepreneurs at the AGI Bar, just like the days before the explosion of the mobile internet. Whether on the model side or the application end, Haidian is the epicenter of this wave of AI innovation, with unique advantages in building an ecosystem.
The AI Wave Starts from Haidian
In fact, every wave of the AI industry in China is closely related to Haidian. If we trace the origin of the industry wave, it began with rule - based and logical - reasoning artificial intelligence, where humans write rules and machines make judgments based on logic and reasoning. Typical products of this stage are "expert systems" and "knowledge graphs". This period was mainly driven by national projects and infrastructure construction. As the region with the most concentrated scientific and technological resources in China, Haidian played a key role in source research, system piloting, and national engineering platforms. It was the first to establish a number of national key laboratories and theoretical research teams, and also accumulated a large number of talents in the process.
For example, the Institute of Automation of the Chinese Academy of Sciences, founded in 1956, was one of the earliest scientific research institutions in China to conduct theoretical research on artificial intelligence. In the 1980s, it undertook major national projects on artificial intelligence under the "863 Program" and made significant contributions in the fields of expert systems, automatic reasoning, and image recognition. Another example is Professor Zhang Bo, a pioneer in the field of artificial intelligence in China. In 1979, under his leadership, Tsinghua University was the first to establish a major in "Artificial Intelligence and Intelligent Control". Many of his students have also joined the current startup wave, including Zhu Jun, the chief scientist of Shengshu Technology, and Yuan Jinhui, the CEO of Yiliu Technology.
Of course, the first wave of AI popularity among the general public was led by perceptual intelligence represented by deep learning. In the 2012 ImageNet competition, AlexNet significantly improved the accuracy of image recognition, marking the rise of deep learning. The core breakthrough of this wave of technology lies in neural networks, shifting from "making machines think" to "making machines perceive". Fields such as image recognition, speech recognition, and NLP became popular. When AlphaGo defeated the human Go champion in 2016, it triggered a global AI boom, and China also witnessed a wave of AI startups. The "Four Horsemen of Computer Vision" (SenseTime, Megvii, CloudWalk, and Yitu) emerged during this period.
Although the "Four Horsemen of Computer Vision" are registered in different places, their core R & D teams, early - stage capital, and application scenarios are deeply connected to Beijing, especially Haidian. For example, Megvii's founders, Yin Qi, Tang Wenbin, and Yang Mu, all graduated from the "Yao Class" at Tsinghua University. In the early days of their startup, they were based in Tsinghua Science Park and the surrounding incubators, which is a typical example of the closed - loop ecosystem of "Haidian universities - parks - industries".
Later, in 2022, with the global explosion of ChatGPT, the era of general artificial intelligence and large models arrived. The emergence of general capabilities based on pre - trained large models has shifted the focus from perception to content generation and complex interactions such as reasoning. AGI has become the goal of entrepreneurs. There has been a boom in large - model startups in China, with companies like Zhipu, Baichuan, Darkside of the Moon, and MiniMax rising rapidly. As early as July 2023, the Haidian District Government's press conference reported that there were about 50 large - model companies, universities, and research institutions in the district, releasing a total of 35 large models, ranking first in the country and forming the first "AI Large - Model Agglomeration Area".
Many of these large - model companies bear the mark of Tsinghua University. For example, Zhipu AI originated from the Knowledge Engineering Group (KEG) Laboratory at Tsinghua University, established in 1996. The laboratory has published a series of internationally leading research results in the fields of knowledge graphs, graph neural networks, and cognitive intelligence. Tang Jie, the chief scientist and the soul figure of Zhipu AI, is the director of the KEG Laboratory. He led the development of China's first open - source large model with one trillion parameters, "WuDao 2.0", and designed the GLM series of model architectures, promoting the autonomy of domestic large - model technologies. In April this year, Zhipu AI filed for IPO guidance, becoming the first large - model company to officially start the IPO process.
Looking at the recently booming embodied intelligence industry, companies like Galaxy Universal, Xinghai Map, Xingdong Era, Songyan Power, Qianxun Intelligence, and Acceleration Evolution all carry the "Tsinghua" label. The recently - funded Guangxiang Technology is also a technology transfer project from the School of Vehicle and Mobility, School of Artificial Intelligence, and the AI Industry Research Institute at Tsinghua University, and even has official shares from Tsinghua University.
Ultimately, Haidian is home to 37 universities, covering all 985/211 universities in Beijing. It has the largest number of AI doctoral and master's graduates in the country and has gathered a large number of top - level research institutions such as the Institute of Automation of the Chinese Academy of Sciences, the Institute of Theoretical Computer Science, and Microsoft Research Asia. It has established a complete AI basic research platform, forming a complete ecosystem from academia to education to research.
In addition to universities and research institutions, cooperation with industry leaders has injected industrial genes into Haidian's source innovation. As reported in the article "Humanoid Robots Flock to Haidian" by China Venture Capital News, as early as 2015, Tsinghua University began to cooperate with Ubtech. They established the Beijing Research Institute, which is divided into two branches: the northern branch focuses on software algorithms and technology pre - research, while the southern branch in Shenzhen focuses on engineering and manufacturing. The Beijing branch has provided a large number of scientific research talents for the embodied intelligence industry in Beijing.
The Combination Punch of "Source Innovation"
"More than thirty years ago, many Zhongguancun companies moved south. Although Beijing established the Shangdi Information Industry Base and the Yongfeng Industrial Base, the manufacturing of hardware equipment still went to Shenzhen, Dongguan, and Huizhou. When the Zhongguancun Science Park was first established in 1999, the 0.5 - square - kilometer core area, 'Zhongguancun West District', did not develop as expected, while the Yuehai Street in Nanshan District, Shenzhen, developed rapidly and was hailed as China's 'Silicon Valley' by The Economist. It was truly a case of 'planting willows casually but getting a shady grove'."
Professor Wang Jici from Peking University, who has been closely following domestic and international industrial parks and innovation clusters, recalled this piece of Zhongguancun history in a recent article. Today, Zhongguancun and Haidian have moved beyond the traditional thinking of simply building physical industrial parks. Instead, they are leveraging their geographical advantages to empower technology startups through a systematic and ecological approach.
Take AI applications, which are currently attracting the attention of investment institutions, as an example. It is well - known that large models are driven by three key elements: data, algorithms, and computing power. For AI applications to be truly implemented, they need to meet six hard conditions: computing power, data, models, application scenarios, talent, and capital. Any weakness in these areas will keep the technology on paper. Haidian has played these six cards in a way that highlights its unique advantages by focusing on source innovation.
In terms of computing power, the Beijing Artificial Intelligence Public Computing Power Platform was launched in Haidian in 2024. Its first - phase capacity is 3,500 P, which will be expanded to 10,000 P by the end of the year. It is also the first to connect to the "East - West Computing" node, forming the country's first three - level computing power network of "city - district - enterprise". Leading companies such as Baidu and Sugon have placed their latest GPU/ASIC clusters in Haidian. Through the Haidian node, startup teams can more easily obtain low - cost, short - term GPU/ASIC computing time, significantly reducing the cost of trial - and - error.
In terms of data, the Beijing Data Basic System Pilot Zone was established in Haidian. The first - phase 1,100 TB of high - quality industry datasets are open - sourced for free and are continuously updated through a point - sharing mechanism. The Haidian Urban Brain has a computing power of 2,028 trillion operations per second, which can collect real - time data from more than 200,000 sensors across the district to build a spatio - temporal information map, providing real - time, annotatable live data for computer vision and multi - modal large models.
On the model side, Haidian has a significant advantage in quantity. As of June this year, 89 large models have been registered in Haidian, accounting for one - third of the national total, covering all fields such as language, vision, scientific intelligence, and embodied intelligence. Models such as WuDao by BAAI, ERNIE by Baidu, GLM by Zhipu, and Darkside of the Moon were all first open - sourced in Haidian. Enterprises can "call with one click + fine - tune on - site" and launch customized industry versions within two weeks.
In addition, as early as June 2023, Haidian District issued the "Several Measures to Accelerate the Innovation and Development of Artificial Intelligence Large Models in Zhongguancun Science City", providing policy support for the research and development of large - model underlying technologies and computing power. The measures focus on issues such as the shortage of computing power resources and high costs in large - model research and development, and aim to build public computing power to reduce the cost of large - model research and development. Instead of just talking, Haidian provides up to 10 million yuan in computing power subsidies for large models with strong technological innovation and rich application ecosystems.
On the application scenario side, in 2024, Haidian District released the first list of "AI Panoramic Empowerment Open Scenarios", covering 10 core scenarios such as government affairs, healthcare, education, industry, and household robots. The government acts as the "leading customer", allowing enterprises to directly test their technologies on - site. At the same time, physical test platforms such as the Zhongguancun Embodied Intelligence Innovation Industrial Park, the Robot Data Collection Factory, and the 3C Electronics Flexible Production Line are also open, helping to solve the "last - mile" problem for robots.
In terms of talent, Haidian has unparalleled advantages in China. 80% of the world's top 2000 AI scholars, 101 highly - cited scientists, and 12,300 AI scholars are concentrated in Haidian. In terms of academic disciplines, 21 universities such as Tsinghua University, Peking University, and Renmin University of China have established undergraduate programs in artificial intelligence. The "20 Measures for the AI Talent Special Zone" have raised the upper limit of annual salaries for core employees by more than 1 million yuan and provided a package of policies such as flexible computing power, rent reduction, and children's education, ensuring that talents can stay and thrive.
Beyond computing power, data, models, application scenarios, and talent, capital is the final piece of the puzzle for promoting AI application innovation. Using capital as a link, Haidian's strategy is to cooperate widely with market - based investment institutions to establish a portfolio of funds covering the entire growth cycle of technology companies.
In February this year, Haidian launched the Zhongguancun Science City Technology Growth Fund Phase III with a scale of 10 billion yuan. So far, the total scale of the Zhongguancun Science City Technology Growth Fund series has reached 20 billion yuan. In March, the technology growth fund signed cooperation agreements with five leading investment institutions: CPE Yuanhe, Lenovo Star, BlueRun Ventures, Shunwei Capital, and Huashan Capital. A public notice in June showed that the technology growth fund plans to invest in five subsidiary funds, including CASC Venture Capital, Meituan Longzhu, Qiming Venture Partners, Lisi Capital, and SEE FUND. This shows that Haidian's capital circle of friends is constantly expanding, and its investment reach is extending.
Data from the "Recruitment Wisdom Investment Promotion System" of China Venture Capital News shows that last year, 38 robotics companies and 79 AI companies moved to Beijing (or established branches), of which 20 robotics and 54 AI companies settled in Haidian. This is just the tip of the iceberg. In the long run, Haidian is using this AI combination punch to upgrade the "source" into an "engine", boosting an AI application market worth hundreds of billions or even trillions of dollars.
This article is reprinted from the WeChat official account "China Venture Capital News" with the author Liu Yanqiu, authorized by 36Kr.