After Unitree's IPO, Investors Tighten Their Purse Strings: Why Robot Companies Are Falling One After Another

How rapidly they rise, and how suddenly they fall.

In 2026, robots not only nearly took the center stage of the Spring Festival Gala; the sector's financing heat also reached its peak: in the first half of this year, 322 financing events occurred in China's embodied intelligence sector, with total financing reaching 93.5 billion yuan, up 137% year on year.

According to IT Juzi statistics, this figure is not only about five times the same period last year, but has also already exceeded the whole of 2025. At the same time, embodied intelligence "unicorns" valued at over 10 billion yuan increased to at least 22, whereas last year the number was still 3.

But in this year of frenzied capital inflow, a batch of robotics companies collapsed one after another. According to public reports and court announcements, in the 19 months from February 2025 to September 2026, more than 20 robotics companies went bankrupt, dissolved, entered liquidation, or shut down business lines.

Last March, after Unitree's H1 appeared at the Spring Festival Gala and brought financing enthusiasm to the industry, Zhu Xiaohu of GSR Ventures publicly said he was "exiting in batches" from robotics companies, questioning whether there is a bubble in the humanoid robotics industry.

Zhu Xiaohu may have been right?

Orders that cannot be fulfilled

Among the robotics companies that have fallen, there is no shortage of star companies backed by big tech and unicorns valued at over 10 billion yuan.

In May this year, DeepBlue Robotics, a unicorn valued at over 10 billion yuan that its founder once positioned as "China's DeepMind," entered bankruptcy liquidation proceedings;

In August, the Shenzhen Intermediate People's Court issued an announcement on bankruptcy acceptance and creditor claims, and Inhe Robotics officially entered bankruptcy proceedings. The company had received investment from Meituan, ByteDance, GSR Ventures and other institutions, with a peak valuation of $500 million. But its full-year revenue in 2025 was only 4.33 million yuan, with a net loss of 67 million yuan.

In October, according to National Business Daily, Wujing Dynamics, which had the halo of a "top-student CEO + a star team of top students from Tsinghua and Zhejiang University," was reported to have owed wages since August this year, stopped paying housing provident fund in September, and stopped paying social insurance in October; it owed wages to thirty or forty employees, totaling nearly 8 million yuan.

Big tech investment and a valuation of over 10 billion yuan sound beautiful, but in the end they cannot help these companies get past commercialization. The reason Zhu Xiaohu "exited in batches" from robotics companies was that the commercialization path was unclear.

In 2021, DeepBlue Robotics announced cooperation with South Korea's LG, which planned to purchase tens of thousands of robots within three years, worth more than 700 million yuan; in 2021, Inhe Robotics signed with Shunde, announcing a 500-mu site in the Beijiao Robot Town to build an Industry 4.0 production line of "robots making robots"; in June this year, Wujing Dynamics also claimed that within four months of its founding it had won nearly 1 billion yuan in orders, including nearly 500 million yuan in orders signed with several central state-owned enterprises.

But judging from the results, these orders and construction plans did not fully translate into actual operating results. According to Sina Finance, the DeepBlue-LG order was not executed on a large scale; Inhe Robotics' 10-billion-yuan base has not been seen going into production, and its orders were mostly demonstrations; of Wujing Dynamics' nearly 1 billion yuan in orders, nearly 500 million yuan were framework agreements, and another nearly 500 million yuan were still local market reserve orders.

What is more helpless is that even if orders are ultimately delivered, that still does not mean the commercialization path has been proven.

A UBS Securities analyst once said that even if many manufacturers plan to achieve deliveries of 10,000 units in 2026, humanoid robots do not equate to a commercialization inflection point; at present, a considerable portion of robot products go to universities, research institutions, data collection centers and entertainment performances, and do not directly enter factories or commercial scenarios to undertake production tasks.

There is a lot of evidence supporting this view. Unitree's prospectus shows that in the first nine months of 2025, 73.6% of Unitree's humanoid robot revenue came from scientific research and education;

In 2025, UBTech accumulated nearly 1.4 billion yuan in humanoid robot orders. A review of the major orders shows that most of the buyers were embodied intelligence data collection and testing centers led by local governments.

IDC data shows that among humanoid robot shipments in 2025, commercial performances, scientific research and data collection accounted for 78.4%. Scientific research, commercial performances and data collection also have real demand, but whether this demand can continue to expand and support the industry's production capacity and valuation remains a question.

Therefore, for robots: terms such as mass production, delivery and orders can only prove that they can be sold, but being sold is only the first step.

Sell to whom? Can the robot work stably? Are customers willing to buy again? These questions determine whether robots can become a sustainable business.

Are they all being bought by data collection centers?

Therefore, before entering factories, robotics companies first found a group of buyers: data collection centers.

Such centers are usually equipped with robot bodies, teleoperation equipment and sites simulating real environments such as homes and factories. Operators control the robots to repeatedly complete tasks such as grasping, carrying and folding clothes, and then use the motion trajectories, visual and force-control data for model training.

As embodied intelligence models need large amounts of real-machine data, data collection centers have been rapidly built across the country. According to STAR Market Daily, as of August 2026, the number of centralized embodied intelligence data collection centers built and under construction nationwide reached 106, of which 84 were already in operation.

But the data collection centers themselves have not yet formed a reasonable closed commercial loop.

Data collection centers themselves are not the problem. For the still-immature robotics industry, data collection is indeed an important part of training models and improving operational capabilities.

The real question is whether the revenue generated by the robot-data collection center cooperation model comes from real data demand, or from a continuously circulating set of transactions.

On September 9, Shao Tianlan, founder of Mech-Mind Robotics, publicly questioned on WeChat Moments that some embodied intelligence companies use data collection centers to manufacture fake or unsustainable revenue in order to support valuations and IPO expectations. He believes: "Robots are sold to data collection centers, and then the data is bought back through channels such as leasing companies, and the revenue is created."

In fact, this business operation model is reasonable. Data collection centers bear the investment in equipment and venues, robotics companies obtain revenue from selling the robots themselves, and then train models by purchasing data. In theory, this can form an industrial chain with a clear division of labor.

Whether this model can be sustained depends on whether there is sufficient paid demand to cover the continuous investment in data collection.

The Embodied Data Development Report 2026 also points out that the current growth in embodied data collection capacity does not match the growth in paid demand, and the business model is still in the early stage of exploration.

The report mentions that a single set of real-machine teleoperation equipment costs more than 200,000 yuan, the price of a robot itself can reach 600,000 to 700,000 yuan, and teleoperators have only 2 to 3 hours of effective collection time per day.

When equipment, venues, and labor all require continuous investment, while the data is difficult to sell repeatedly, data collection centers may turn from industrial infrastructure into a heavy cost burden.

For robotics companies, data collection orders can bring early revenue, but they still cannot prove that the product already has the capability for large-scale application.

If equipment procurement continues to increase while paying data customers do not keep up, then how far can the orders brought by data collection centers actually support robotics companies?

Robots: anyone can make them now

The commercialization of humanoid robots has not yet been proven, yet more and more players are entering the field. On the one hand, this is driven by the attraction of financing enthusiasm; on the other hand, the threshold for building a robot has become increasingly low.

In March 2026, Shanghai Pudong launched the Zhangjiang South Embodied Intelligence Supply Chain Platform, one of the country's first embodied intelligence supply chain platforms, providing enterprises with one-stop services ranging from rapid complete-machine assembly and core component support to pilot testing, scale-up testing, and maintenance. The platform claims it can achieve "matching in one day, prototyping in three days, and trial production in one week" at the fastest.

The robotics industry chain is becoming more modular. Startup teams do not necessarily need to develop all motors, reducers, controllers, and mechanical structures from scratch; they can also leverage mature suppliers and external manufacturing services to quickly assemble a robot capable of walking and grasping objects.

This lowers the threshold for hardware development, but it also brings a problem: when more and more companies can produce robots with similar appearances and similar movements, it is already difficult to judge the real technological gap between them based on a demo video alone.

In August this year, a robotics industry strategy report pointed out that as capabilities such as gait, dynamic balance, and basic grasping gradually become widespread, more and more companies are able to complete similar public demonstrations.

What truly sets companies apart is continuous operating time, task success rate, failure rate, and whether the robot can adapt to different working environments. The report believes that it is becoming increasingly difficult to obtain valuation premiums based on technical demos alone.

In other words, making a robot move no longer requires possessing technology that is difficult to replicate.

Beyond hardware, the robot's "brain" has also become another focus of competition. In the past, the industry discussed vision-language-action models (VLA), hoping to enable robots to directly generate actions based on visual information and natural language instructions. This year, concepts such as world models and physical AI have gained more attention.

But updates to model architecture will not automatically translate into actual working capability. Sina Finance's visits to several embodied intelligence companies found that different technical routes are being advanced, from VLA model training and teleoperation data collection to world model research and development, but there is still a clear distance between a model completing a demo and a robot truly entering the production line.

At the same time, financing enthusiasm in the robotics sector has also begun to diverge.

After Unitree's listing, some investment institutions also began to re-examine the valuation of humanoid robots. According to reports in the investment community, an AI investor revealed that after Unitree's listing, they no longer submitted embodied intelligence projects to the investment committee for review. Another investor also said that although the institution has not explicitly announced a halt to investing in robots, actual investment decisions have basically stopped.

At least for these investment institutions, a prototype, a demonstration, or a grand model story is no longer enough. Robotics companies need to deliver more convincing technological and commercialization results.

In September this year, when asked about Unitree's share price halving and his views on financing enthusiasm for humanoid robots, Zhu Xiaohu once again said:

"(Humanoid robot companies) will definitely undergo a shakeout later, and repeatedly."

However, the great wave washing away the sand is not actually a bad thing. As capital begins to cool, the true value of robotics companies will also face stricter scrutiny.

For the entire industry, what is always needed is those robotics companies that truly have substance.