Is Embodied AI at a "ChatGPT Moment" or Heading for a Cold Winter? Founders Weigh In

Embodied intelligence is growing rapidly, but it still has a long way to go before reaching the real "ChatGPT moment." Multiple investors in the embodied intelligence sector told the Economic Observer that a "winter" may come next, with tightened financing and other situations. After the hype, there will be a comprehensive screening of business models and technical routes. How to survive may become a common challenge for embodied intelligence enterprises.

By the end of 2025, the embodied intelligence industry is being pushed forward by two forces simultaneously.

On the one hand, there is a long - awaited capital boom. Humanoid robots are regarded as the "next trillion - dollar terminal," and the capital enthusiasm continues to heat up. Data from IT Juzi, a third - party data institution, shows that the domestic robot financing amount in the first three quarters of this year has reached 38.624 billion yuan, 1.8 times the total amount in 2024. Since November, companies such as Yuanli Lingji, Xingchen Intelligence, and Jiasu Jinhua have successively completed a new round of financing. Investors are shifting from investing in concepts to paying more attention to the commercialization speed and verifiable value. This enthusiasm has also extended to the enterprise side. New products are emerging one after another, and leading manufacturers frequently announce that they have obtained thousands of "commercial large orders." The term "the first year of mass production" appears repeatedly.

On the other hand, there are risks that cannot be ignored. Silicon Valley star company K - Scale Labs suffered a broken capital chain, and more than a hundred pre - ordered orders had to be refunded; domestic startup company Yixing Robot was dissolved. Both companies were established less than a year ago and had received multiple rounds of financing. These phenomena have made the industry realize that "the first year of mass production" is not equivalent to an inflection point in the industry. Instead, it is more like a collective life - and - death test.

These two forces jointly point to a core fact: embodied intelligence is growing rapidly, but it still has a long way to go before reaching the real "ChatGPT moment." Multiple investors in the embodied intelligence sector told the Economic Observer that a "winter" may come next, with tightened financing and other situations. After the hype, there will be a comprehensive screening of business models and technical routes. How to survive may become a common challenge for embodied intelligence enterprises.

Before mass production, distinguish between real and false demands

In the past year, the capabilities of humanoid robots have improved almost "visibly." Robots that could only stand stably in 2024 can now do continuous flips, run marathons for several hours, and even drag vehicles on the spot.

However, all these amazing actions do not necessarily mean usability. Almost all interviewees expressed a similar judgment: embodied intelligence still has a long way to go before truly entering high - intensity, long - cycle, and repeat - purchaseable real - world scenarios.

At the Zhiyuan Embodied 2025 Open Day, Wang Zhongyuan, the dean of the Zhiyuan Research Institute, gave a typical example. They purchased 10 units of a certain type of robot, and 5 of them broke down within just one or two months. "The hardware stability is still at the research stage," he added. In the laboratory, the robotic arms frequently shut down due to over - heat protection, and some robots even need a fan beside them, "like having a nanny for them." In his opinion, the scene of "robots running all over the street" will not appear in the next two or three years. This is the reality on the hardware side and the most direct obstacle to large - scale deployment.

The bottlenecks on the model side are also obvious. Embodied large models are still in the early stages in terms of control accuracy, cross - environment generalization, and operation consistency. The toolchain is incomplete, and the deployment standards are not unified. It is much more difficult to let a robot continuously execute a two - hour process in a real - world scenario than to complete a one - time action on a booth.

Therefore, Wang Zhongyuan warned that the industry must distinguish whether the current mass production comes from real demands or false demands driven by policy subsidies and investment enthusiasm. If the scenario providers find that the robots cannot meet the expectations, the enthusiasm will quickly fade, and there may even be periodic bubbles and troughs. "A sentence I often say to founders is 'Survive first, get through the possible winter, and then you can embrace the real future of embodied intelligence.'"

In 2025, with the joint acceleration of capital, technology, and enterprises, the industry has witnessed a boom. Many companies have released models, complete machine hardware, and offered high salaries to recruit talents.

Tang Wenbin, the co - founder of Yuanli Lingji, called this year the "year of emergence" - companies, technologies, and funds have emerged in large numbers at the same time, and the progress has far exceeded expectations. But at the same time, he observed another side of the industry, where there are many "magical orders." Some projects involve large amounts of money and seem like commercial large orders, but upon closer inspection, it is difficult to explain what real problems they solve, whether they can form repeat purchases, and whether they can truly help enterprises reduce costs or improve efficiency. He warned that mass production cannot rely on simply piling up equipment but on solving problems.

When judging whether a scenario is worth entering, he proposed three criteria: First, the technology should not be locked in prematurely. To seize the first wave of orders, some enterprises are eager to adapt robots to a highly customized vertical scenario. However, this path often sacrifices the model's generalization ability, making it difficult to expand to other applications later.

Second, start with scenarios with high fault tolerance and low time sensitivity. Early - stage robots cannot achieve "zero errors." Cross - environment operation requires more time for optimization. Therefore, the industry should choose scenarios with high tolerance and gradually improve the usability from 90% to 95% or 100% through real - world deployment.

Finally, the demand must be large and strong enough to verify the value, spread the costs, and form a real commercial closed - loop. Otherwise, no matter how large the order volume is, it may only be a periodic accumulation and difficult to sustain.

On the demand side, the judgment of purchasers is more straightforward. A person in charge of an embodied intelligence project in a large enterprise said that they only look at three indicators: whether the robot can solve problems with high complexity, high risk, and high cost. In addition, the robot must be able to operate stably 24/7 and have basic capabilities such as continuous work, waterproofing, dust - proofing, and environmental adaptation. "I've asked many manufacturers, and many of them haven't really considered these indicators," he said. In his opinion, these seemingly basic engineering indicators are the key to whether a robot can enter the repeat - purchase stage.

High - quality data is extremely scarce

From demos (demonstration videos) to large - scale applications, they all point to the same problem: high - quality data is still extremely scarce.

Different from the pre - birth era of large language models, which had a large amount of text and image data, embodied intelligence follows a different path: every piece of key data comes from the interaction between robots and the real world, and this kind of data is much scarcer and more costly than text.

At the beginning of this year, many people judged that the embodied intelligence industry was roughly at a stage similar to that between GPT - 1 and GPT - 2 of large language models. Now, as it approaches the "robot GPT - 3 moment," the model's capabilities have improved significantly, but it still has a long way to go before truly understanding the world.

Wang He, the founder of Galaxy Universal, presented a key fact: there may be fewer than 1,000 robots truly operating in human work scenarios globally today. This number is far from sufficient to support an action - first model system. Galaxy Universal's strategy is that in the short term, simulation and synthetic data will still undertake more exploration tasks; in the long term, the real - world deployment volume of robots must increase by hundreds or thousands of times.

Luo Jianlan, the partner and chief scientist of Zhiyuan Robot, mentioned that the future data ecosystem should rely on robots to generate data themselves. His vision is to deploy large - scale robots in the real environment, let them continuously interact with the world, generate real physical data with wide coverage, long time - series, and complex structures, and then use this data to feed back into model training, forming a self - evolving closed - loop.

Wang Qian, the founder of Zibianliang, warned that the industry's understanding of data is changing. It's not that the more data, the better, but "the more effective, the better." The era of language models has proven that high - quality, structurally consistent data is often more effective than simply piling up large amounts of data. This is especially true in the physical world, which is full of details such as contact, friction, and collision, and is difficult to describe with language or pictures. If the model cannot understand these basic physical processes, it cannot establish a reliable expectation of the world. Therefore, he believes that in the future, the new type of basic model of the physical world driven by embodied intelligence may dominate the multi - modal direction. This will be a different development path from that of large language models.

Looking back at the development of large language models, we can see that three conditions are indispensable - algorithms, computing power, and data must reach the critical point simultaneously. In the field of embodied intelligence, these three conditions are not yet fully mature. The algorithms are still in the stage of exploring usability; the computing power system for training embodied intelligence is still weak; and the data scale and data quality are far from reaching the critical point.

In other words, the future of embodied intelligence is approaching at an accelerating pace, but it is still in the pre - dawn stage. The hype is high, and the capabilities are strong, but there is still a long way to go before the real golden age of industrialization.

This article is from the WeChat official account "Economic Observer". Author: Zhou Yue. Republished by 36Kr with authorization.