On the morning of September 11, during the 2026 Inclusion · Bund Conference, industry professionals, investors, and young entrepreneurs in the field of embodied intelligence discussed topics such as industry bubbles, technical routes, and the "ChatGPT moment." Many guests believe that the current "bubble feeling" in embodied intelligence mainly comes from a mismatch between the industry stage, valuation, and commercial value. The industry is still in the "pre-dawn" phase. At the same time, key capabilities of embodied intelligence have begun to emerge, but they are still far from forming large-scale productivity. System engineering, computing power, and data remain the main bottlenecks.
Responding to the "Bubble Theory": Embodied Intelligence is Still in the Pre-Dawn Phase, and the Industry Structure is Far from Finalized
The development of the embodied intelligence industry has only been a few years, yet it has already attracted a lot of capital and startups. Faced with the recent rising "bubble theory," many entrepreneurs believe that the industry is still in its early stage, and the current issue is more of a temporary misalignment between valuation, technological progress, and commercial value.
Jia Peng, co-founder and CEO of Zhijian Dynamics, believes that the "bubble" people feel now is more about a mismatch between stages and valuations. In the long term, embodied intelligence may form an industry scale larger than mobile phones or cars, with a sufficiently long industrial chain, from sensors and components to final applications, which can accommodate companies of different types.
"I think the reason people feel there's a bubble at this stage might be related to the fact that we haven't created much value yet; it's a phase mismatch," Jia said. Embodied intelligence requires more human resources and GPU, and the overall industry investment is still not sufficient.
Xiao Wanggang, founder and chairman of Daxiao Robotics, also believes that part of the current anxiety comes from the high expectations of the market for the development speed, while substantial breakthroughs in models and scenarios are still waiting to appear.
"I feel that we are in the pre-dawn," Xiao said. Compared to last year, when more reliance was placed on real-machine data, VLA, and small model approaches, this year, with the introduction of ego data centered around humans and world models, we have started to see the manifestation of the Scaling Law. He believes that embodied intelligence is closely integrated with vertical industries and will still have strong diversity in the future. Currently, the links such as data, components, models, physical bodies, and scenarios are relatively scattered, and different types of companies may further recombine in the future.
Han Zheng, co-founder and CEO of Sudo Technology, stated that China has a relatively rich supply chain and scenario testing environment, so there are many companies and technology explorations in areas such as physical body, embodied models, and world models. He believes that embodied intelligence is likely to eventually become a complex system, and there are signs of convergence in technical routes. In the future, companies that can provide complete underlying solutions and models may not be particularly numerous, but there are still opportunities in vertical applications, global markets, and model combinations.
When Will the "ChatGPT Moment" for Embodied Intelligence Arrive? System, Computing Power, and Data Remain the Main Bottlenecks
Regarding the question of when the "ChatGPT moment" for embodied intelligence will arrive, several guests gave their judgments from different perspectives, including systems, computing power, and data.
Han Zheng believes that if compared to large language models, embodied intelligence is currently at a stage similar to 2018-2019. Many key technical frameworks have already emerged, and the current bigger challenge is how to integrate different technologies into a commercially viable, highly reliable, and generalized big system.
"Integrating this big system is the most important issue for the entire embodied intelligence industry right now," Han said.
Jia Peng, however, does not like the term "embodied ChatGPT," because the problems addressed by embodied intelligence and digital AI are not entirely the same. A large language model achieving 80 points may already be a good product, but once embodied intelligence enters the real physical world, the requirements for reliability are much higher. He believes that if we only look at model capabilities, embodied intelligence has already shown some ICL (in-context learning) abilities. However, if we consider the formation of real productivity, it may still take three to four years or even four to five years.
If only one biggest bottleneck could be chosen, Jia Peng believes it is GPU computing power, including both training and edge-side inference computing power. Many factory scenarios cannot be connected to the internet, and robots must be deployed on the edge side. Insufficient edge-side computing power directly limits model capabilities.
Xiao Wanggang believes that the current biggest bottleneck is data. In his view, the key to the "ChatGPT moment" for embodied intelligence is whether intelligent emergence, scenario generalization, and zero-shot capabilities can be observed.
"If we always need to re-train for new tasks in various application scenarios, we won't be able to achieve large-scale promotion," Xiao said. Recently, the zero-shot capabilities shown by general models have also given him new hope: the intermediate data output by these models may help accelerate the evolution of embodied basic models in the future.
Young Entrepreneurs on Route Selection: World Models, the Physical World, and Global Competition
At the forum, many entrepreneurs born in the 90s and 00s also shared their experiences and discussed technology routes and global competition based on their own entrepreneurial practices.
"When we started, we said that in the next six months, all embodied intelligence companies would claim to be world model companies," said Li Yiming, founder of Licheng Intelligent. He noted that there are many world model companies, but the key is to see what specific problems they solve.
In his view, startups should first determine what the ultimate outcome might look like and then break down the technical routes needed to achieve that end goal. Based on priority and the team's background and strengths, strategic choices can be made.
Ding Ning, founder of Natural Will, believes that "the GPT-3 moment in the field of embodied intelligence has already arrived, it's just that it hasn't been noticed yet." He said that after 2025, the development of large models has solved some fundamental issues, but the problem of operating in the physical world still needs to be resolved.
He described the industry's recent development as happening at an "unusual pace." "Whether it's code volume, experimental results, or the integration of different intelligences, we can feel a very strong pushing force."
Facing overseas entities that have access to massive computing resources, Ding Ning admitted he felt "very anxious." However, he also believed that China still has opportunities, one of which comes from the hardware and supply chain capabilities behind high-precision data. "High-precision data hasn't scaled up yet, and high-precision data requires better hardware sensors, and better sensors require more intelligent fusion technology. This can only be produced in Shenzhen globally right now."
Chen Boyuan, founder of Nix Matrix Technology, who is 22 years old, said that his generation of entrepreneurs grew up after the emergence of Scaling Law. "We have experienced the rise and peak of large models, but we also see that large models still face many insurmountable challenges in solving real-world physical problems."
In his view, for embodied intelligence to truly enter the physical world, physical AI foundations need to understand causality. The real physical world is full of noise and continuous change, "edge cases are not the tail, but the norm."
Chen Boyuan believes that physical AI is not something that can be completed by 'a few geniuses.' It is a complex system engineering project involving model infrastructure, data, algorithms, and the real physical world loop. "People won't easily trust your chosen path just because you used to work at a big company; nor will they dismiss your ideas because you're an intern." He said that the most important thing to drive technological innovation is to respect the first principles, find the most promising technical routes through small-scale validation and large-scale scaling, rather than relying on the experience of an individual or a few people.
