12 billion, Zhou Zhifeng unveils Kaiming Venture Capital's AI landscape
Where is the tide of the AI industry heading?
On July 19, at the 2026 WAIC "Qiming Venture Partners · Entrepreneurship and Investment Forum", Qiming Venture Partners released the Top 10 AI Outlook for 2026, covering foundation models, embodied intelligence, AI infrastructure, AI applications and other fields.
Qiming Venture Partners has been doing this for four consecutive years, and it has become an industry weathervane. What the outside world may not know is that this leading VC institution has invested in more than 100 projects in the AI sector in total, with a total investment of 12 billion RMB. Qiming Venture Partners' presence can be seen behind star projects such as Zhipu, Biren Technology, StepStar, Kling AI, and Galaxy Universal.
Not long ago, Zhou Zhifeng, Managing Partner of Qiming Venture Partners, had an exchange with PEdaily. "The more frenetic the era, the more we need to learn from historical lessons; the more noisy the era, the more important philosophical thinking becomes." Experiencing the current AI boom firsthand, Zhou Zhifeng's words are filled with restraint and calm.
"We receive embodied intelligence BPs every week"
Several observations from Zhou Zhifeng
"In the past year, embodied intelligence has been the field that 'killed' the most of my brain cells, bar none." Zhou Zhifeng remarked to PEdaily with emotion. As one of the earliest institutions to lay out this track, Qiming Venture Partners has invested in many impressive projects: UBTECH, CloudWhiz, Galaxy Universal, Forcen, RoboXplorer... the list goes on.
Not long ago, Zhou Zhifeng successively witnessed the IPOs of Biren Technology, Zhipu, Axera, and Yunyinggu. Precisely because of long-term deep cultivation in the track, with the number of enterprise samples far exceeding most peers, Zhou Zhifeng's feelings are even more profound.
Witnessing the concentrated attention from the primary and secondary markets pouring into the embodied intelligence track, Zhou Zhifeng believes that the core reason is that this may be the first industry in history that has the market size of two major sectors at the same time — it boasts the shipment scale of smartphones, while also having the unit price of passenger vehicles. He made a simple calculation: assuming the industry matures, there will be an annual shipment volume of 1 billion units, with an average unit price of about 30,000 USD, equivalent to 200,000 RMB. This is undoubtedly one of the largest tracks in the 200-300 year commercial history of mankind.
Regarding the strong expectations of the secondary market for embodied intelligence enterprises, Zhou Zhifeng believes that this is determined by the characteristics of the secondary market. That is, when the first one or two enterprises in each major track go public first, due to the scarcity of targets, they can enjoy super capital dividends, which is intuitively reflected in that their stock prices and market capitalization will be so high that they deviate from conventional logic. "The essence of embodied unicorns rushing to go public is to chase this scarcity dividend."
The grand scene is vivid in memory. According to incomplete statistics from the Qiming Venture Partners team, there are as many as 370 newly established domestic embodied intelligence enterprises in just the past two years, and even now the team can still receive two or three BPs of new projects every week.
Zhou Zhifeng admits that the track has become so crowded that many companies are difficult to distinguish, and their backgrounds are increasingly homogeneous: on the one hand, founding teams are concentrated in several categories: university professors, "child prodigies", senior executives from autonomous driving giants, and AI model researchers; on the other hand, the technical routes are nothing more than VLA or world models, and the implementation scenarios overlap in industrial logistics, commercial services, and bionic robots.
He judges that if there is no key breakthrough in the embodied intelligence track, especially if the technical routes cannot converge, the industry will find it difficult to achieve large-scale scenario implementation, and ultimately can only make some dispensable demo projects. Once the commercial implementation fails to meet expectations, even if the companies go public successfully, their market capitalization may fall to the range of tens of billions, which will probably lead to valuation inversion between the primary and secondary markets. "What we can do now is to continuously and actively capture all new projects, ensure comprehensive information collection, and continuously track the industry landscape."
Looking at the global perspective, we can see that embodied intelligence has now become the main battlefield of competition among major countries. After years of observation, Zhou Zhifeng found that China's advantages are mainly concentrated in three aspects: data, implementation scenarios, and hardware supporting facilities.
According to his disclosure, several leading US giants are purchasing data from Chinese enterprises, which is enough to show that their own data reserves are insufficient. In sharp contrast, a embodied intelligence data platform enterprise invested by Qiming Venture Partners, which was established only 4 months ago, has existing orders at the level of 1 billion RMB.
In addition, China has leading manufacturing enterprises such as CATL and BYD, with sufficient physical factories that can collaborate in R&D, containing a vast range of industrial implementation scenarios. In contrast, leading US enterprises such as Figure AI can only travel far to Europe to cooperate with BMW.
More critically, among all embodied intelligence enterprises in the United States, only Figure AI and Tesla have the ability to develop the entire machine hardware independently. According to calculations, a humanoid robot has about 1,200 components, and more than 90% of the supply chain is concentrated in the Yangtze River Delta and Pearl River Delta regions in China. "As a result, domestic embodied enterprises can rapidly iterate both the robot body and the model simultaneously. Once a mismatch between the model algorithm and hardware execution is found, they can contact suppliers to adjust and optimize within two weeks."
During the conversation, Zhou Zhifeng also talked about the recently popular world models. Roughly calculated, about 30 enterprises have flocked into this track. But he said frankly that the world model is not a brand new track. These newly emerged startups have no essential difference in commercial implementation compared with companies that previously adhered to the VLA technical route. "In other words, the world model has just been hyped into a popular concept in the primary market."
"Half a step ahead"
New heavy investment in Kling
This is just a microcosm.
After sorting out, Qiming Venture Partners' layout in the artificial intelligence field almost covers the entire industry stack: from chips carrying computing power and foundation large models driving innovation, to robots and vertical scenario applications that undertake technology, and then extending to cutting-edge integrated fields such as life sciences, a systematic investment map has taken shape.
In fact, Qiming Venture Partners' systematic layout in the artificial intelligence field started as early as 2013. During that period, it successively invested in future star enterprises such as iFlyTek, Megvii, UBTECH, and WeRide. And before the emergence of GPT-3 in 2020 and the global popularity of ChatGPT at the end of 2022, the team's investment portfolio had already laid out key nodes in the AI industry chain such as Zhipu, StepStar, ShengShu, AxonAI, and Infinite Lightyear.
PEdaily obtained a set of data — up to now, Qiming Venture Partners has invested in more than 100 AI projects in total, with a total investment of 12 billion RMB. The invested enterprises cover the entire AI industry chain, making it one of the most active investment institutions in the artificial intelligence sector in China and even in Asia.
Just as peers perceive, Qiming Venture Partners' AI investments always seem to lay out in advance at every key node, and finally reap the rewards of time when the industry erupts.
This is no accident. Qiming Venture Partners has always emphasized the investment strategy of "half a step ahead" — the golden window for technology investment often appears when the "technical breakthrough point" has emerged but the "market ignition point" has not yet arrived — requiring precise layout between the two.
Among them, Qiming Venture Partners' bet on Zhipu is the most vivid embodiment. As early as May 2020 when GPT-3 was released, Zhou Zhifeng and his team saw that the Scaling Law had been verified, the technical routes were converging, and they had reached the technical breakthrough node. At that time, ChatGPT had not been released, almost no one in China had heard of the term "large model", and the "All in AI" trend in the venture capital circle was nowhere to be seen.
But based on keen judgment, Qiming Venture Partners co-led the B1 round financing of Zhipu in December 2021, and increased its investment in subsequent rounds. Today, Zhipu has become a non-negligible name in global AI, with its market capitalization once exceeding 1 trillion HKD.
Zhou Zhifeng reminds that investing before the technical breakthrough point will be very dangerous for investment institutions, because that stage is more suitable for national laboratories to carry out divergent exploration. Even if the investment institution bets on the route that will succeed in the future, the exit cycle may be 10 or 15 years, far exceeding the normal investment cycle.
He recalled a case he invested in Silicon Valley before. It was a hydrogen fuel cell company whose market value now exceeds 700 billion USD. But unfortunately, this company was invested by Zhou Zhifeng's previous institution in 2002. After accompanying it for more than 10 years, they had to exit all positions, and did not wait for the commercial explosion stage of hydrogen energy.
In the past year, video model technology has achieved a leapfrog growth. Specifically, the new generation of video models, such as the globally popular Seedance 2.0, adopts the MoE architecture, with significantly improved intelligent capabilities, and now supports 4K resolution. For this reason, many Hollywood films, or advertisements of big brands such as Coca-Cola and McDonald's, have many clips fully or partially generated by AI, relying on the high-definition generation capability of the models.
"In particular, world models can empower video generation, realize object movement and collision effects, and restore real physical laws, which was completely unpredictable a year ago."
In Zhou Zhifeng's view, the market size of video models is expanding rapidly, and after scaling up, the division of labor will be more detailed, and the commercial focus of each player has shown obvious differentiation. Among them, the leading players are divided into two categories: there are three global giants, namely ByteDance's Seedance, Kuaishou's Kling, and Google's Veo; then there is a legion of startups gathered from all sides, whose business and revenue have achieved a tenfold growth at present.
In this track, Qiming Venture Partners' "half a step ahead" investment is also vividly reflected. As early as the beginning of 2024, Qiming Venture Partners participated in the angel round series financing of ShengShu, making an early layout, and the latter just completed a 500 million USD B+ round financing not long ago.
Almost at the same time, Kling AI, a video generation large model under Kuaishou, completed financing of nearly 30 billion USD, with a post-investment valuation expected to reach 18 billion USD, and Qiming Venture Partners' presence can be seen behind it. Talking about this investment, Zhou Zhifeng admits that the valuation is indeed very high, and it is not a traditional VC project.
The reason why Qiming Venture Partners placed a firm bet is that Kling has built a capability circle for video generation. "The Kling team's understanding of video generation gave me confidence, and I firmly believe that Kling's future investment returns will exceed everyone's expectations."
Top 10 AI Outlook for 2026
In the torrent of the AI era, some changes have far exceeded Zhou Zhifeng's expectations.
The first is AI computing power. At present, the overall popularity and increment of the computing power market, as well as the transfer speed of computing power paradigm demand from training to inference, are impressive. According to Zhou Zhifeng's disclosure, a domestic technology giant finally allocated more than 50 billion RMB for computing power last year, and this year's budget is 6 times that of last year.
He emphasizes that whether a large number of new-generation AI chip enterprises emerge in the primary market, or the secondary market speculates on the HBM memory and optical communication tracks, the essence of all these hot trends is driven by huge computing power demand, and the underlying logic is consistent.
Secondly, it is the development speed of model technology itself, and the consensus that the market quickly formed around the models. Last year's WAIC, Qiming Venture Partners mentioned in its Top 10 Outlook that coding capabilities are very important. Today, coding capabilities have become the most core competitiveness of large language models.
Correspondingly, the development of AI applications is lower than Zhou Zhifeng's expectations, "The way to open it is different from what I thought last year." In 2025, he was full of optimism that this year we would see AI empowering thousands of industries, giving birth to several 2C applications that are promising to become new Tencent, new ByteDance, and new Alibaba. But today, no exciting 2C applications have emerged in the market.
In response, Zhou Zhifeng conducted an internal review and summary. He pointed out that the first generation of AI applications established a few years ago, mostly represented by products such as dialogue tools and emotional companionship CharacterAI, many of which have now lost development momentum, and the industry has fallen into homogeneous product competition. The user growth rate is no longer as fast as in the previous two years, and the overall growth in the past year has been relatively slow. "The core problem is actually simple and clear: the user growth and traffic logic of the Internet era does not work for making 2C products in the AI era."
"So, where to place the next investment?" As 2026 is halfway through, this is a question that every investment institution will be asked. One of the highlights of this year's WAIC is that Zhou Zhifeng, on behalf of Qiming Venture Partners' AI investment team, once again released the Top 10 AI Outlook, including:
Foundation Models — Outlook 1: In the next 12-24 months, leading models will internalize most of the "add-on" capabilities, including task planning, tool invocation, multi-agent collaboration, and part of the Harness engineering capabilities.
Outlook 2: Multimodal models will further evolve toward interactive world modeling, becoming a key technical path for AI to obtain environmental perception, long-term planning, and large-scale implementation in the physical world.
Embodied Intelligence — Outlook 3: The effective data of leading robot companies will jump from the "10,000-hour level" in 2025 to the "million-hour level" in 2027, with human first-person perspective data accounting for an absolute majority.
Outlook 4: Dexterous hands will see accelerated development. Dexterous hands with tactile perception will gradually mature and continue to reduce costs, forming a high-low matching with the two-finger gripper solution, and gradually penetrate on a large scale in complex operation scenarios.
AI Infrastructure — Outlook 5: The focus of AI computing power demand will shift to inference. The production capacity of storage, advanced manufacturing processes, and advanced packaging will become increasingly tight. In the next two years, AI infrastructure will continue to face structural shortages, and computing power reserve will be upgraded to the core strategy of AI enterprises.
Outlook 6: AI infrastructure will enter the stage of system-level competition, with competition dimensions covering key links such as chips, interconnection, heat dissipation, and power supply. In the next two years, computing power chips and super-node large clusters based on new architectures are expected to emerge, to achieve low-cost, high-efficiency Token production.
Outlook 7: Security and trust will be listed alongside product performance and Token cost as the three key variables affecting the large-scale implementation of enterprise AI. Secure and trustworthy AI will be upgraded from an optional item to a mandatory item.
AI Applications — Outlook 8: In the next 12-24 months, the business model of AI applications will accelerate to break away from the Freemium logic of the Internet era, and shift to pricing based on results and value. The core indicator to measure AI companies will shift from user scale to the commercial value created by per-unit intelligence cost.
Outlook 9: In the next 12-24 months, the commercialization of AI applications will focus on vertical scenarios and high-paying users. The efficiency side (save time) will take the lead in breaking out before the consumer side (kill time); with the continuous decline of Token cost and the innovation of interaction paradigms, phenomenal AI consumer applications will gradually emerge.
Outlook 10: In the next 12-24 months, AI-Native organizations will move from concept to empirical evidence, and a number of enterprises will achieve per capita productivity several times that of traditional organizations.
"This year I have seen many irrational investment behaviors