200 billion yuan in hot money has poured in, and embodied intelligence is stuck on the "gloves".
At the beginning of 2026, the financing pace in the embodied intelligence track was so rapid that it made people feel dazed.
On March 10th, Lingchu Intelligence first publicly dissected its financing situation: After being established for one and a half years, it has cumulatively raised more than 20 billion yuan, and the list of investors ranges from China Development Financial Corporation and CCTV Industrial Fund to Yangtze Optical Fibre and Cable Joint Stock Limited Company and SAIC's industrial capital. This previously low - key company's valuation has increased by six to seven times within a year.
This is not an isolated case. According to incomplete statistics, since 2026, the total disclosed financing in China's embodied intelligence and robotics track has reached approximately 20 billion yuan. Galaxy General's new financing of 2.5 billion yuan has refreshed the industry record. Qianxun Intelligence has secured nearly 2 billion yuan in two consecutive rounds of financing. Independent Variable Robotics, Zhifang, and Xingdong Jiyuan have successively crossed the threshold of a valuation of 10 billion yuan. In just two months, seven new unicorns have emerged in the industry.
However, having a lot of money does not mean having a clear story. A partner of an institution that has participated in multiple investments recently admitted that everyone is in a hurry now. "The secondary market will definitely be enthusiastic about the concept of embodied intelligence, but the primary market needs to figure out what exactly these funds are betting on."
The answers are diverging. Some are betting on the shipment volume of the robot body, some on the generalization ability of the "brain", and others on an even more fundamental aspect - data. The two founders of Lingchu Intelligence explained to the media very straightforwardly: The core reason why embodied intelligence has not been implemented so far is the data problem, and the solution to the data problem lies in human hands.
The Flow of 20 Billion: Seven New Unicorns and Two Routes
If we spread out the financing map at the beginning of 2026, we can clearly see the capital's flow preference.
Unitree Technology and Zhipu Robotics are still advancing at their own pace. The former has managed to reduce the price to the 30,000 - yuan level through extreme cost - reduction, and more than 6,500 units were mass - produced and launched in 2025. However, more newly - emerged companies with a valuation of 10 billion yuan are focusing on the "brain".
Independent Variable Robotics has launched the native multi - modal architecture WALL - A, emphasizing the integration of VLA and the world model; Qianxun Intelligence's self - developed Spirit v1.5 model has become the first Chinese open - source model to surpass Pi0.5 in performance; Zhifang has created the all - domain and whole - body VLA large model GOVLA, emphasizing zero - sample generalization ability.
The founding teams of these companies mostly have backgrounds in AI or autonomous driving. Their consensus is that the ultimate competition of humanoid robots is the competition of the "brain".
Capital's acceptance of this is increasing. Zhang Zhikuan, a director of the investment department of Cornerstone Capital, has observed a change: As technological paths such as VLA gradually mature, the generalization ability of robots in cross - scenario tasks has significantly improved. "Brain - focused companies are beginning to show stronger technological certainty."
However, the embarrassing reality is that this group of newly - emerged brain - focused companies with a valuation of 10 billion yuan do not have outstanding commercialization data. IDC data shows that their shipment volumes in the past year were all within a thousand units. In contrast, Unitree and Zhipu have achieved an annual revenue of one billion yuan.
The capital market is not betting on the current shipment volume. The American humanoid robot company Figure AI was established in 2022, and its shipment volume is still in the pilot test stage, but its valuation has reached 39 billion US dollars. The key variable is the launch of its self - developed end - to - end VLA model "Helix", which has evolved the robot's functions from moving boxes to folding clothes. In China, Independent Variable Robotics' model can complete the entire process of takeaway delivery without human intervention. This is the underlying logic for Meituan to make consecutive investments.
Some are also betting on another direction. The two founders of Lingchu Intelligence clearly stated that they will not engage in self - research of the whole - machine hardware - the supply of wheeled chassis is already in excess, but they must conduct self - research on dexterous hands and data collection devices. The core route of this company is to train the embodied model through the method of "body - less data collection": humans wear data gloves to collect operation data, and then use these data to train the basic model.
Whether this solution can work has not been concluded yet. However, its emergence has precisely hit a collective anxiety in the industry: Where does the data come from?
The Data War: The Ebb of Simulation and the Entry of Human Data
"There is a consensus on the data problem, but the more fundamental problem is that there is currently no mechanism that combines a technological path with a business model to enable data to flow back on a large scale and at low cost." Wang Qibin, the CEO of Lingchu Intelligence, once analyzed. There is no "Tesla model" in the embodied field yet.
The Tesla model was able to be launched because it started to lay the foundation in 2013. After the Model 3 was released in 2017, its annual sales quickly reached one million units. Relying on a large - scale deployment, data flowed back rapidly, thus promoting the iteration of FSD. However, humanoid robots today are facing a typical "chicken - and - egg" problem: without data, large - scale deployment is impossible, and without deployment, there is no low - cost data flow - back.
Simulation data was once highly anticipated, but the gap is too large. Chen Yuanpei, the co - founder of Lingchu Intelligence, reviewed the team's changing perception of simulation: In the early stage, for some demos - like playing mahjong - a lot of data was collected from simulation because of its high parallel efficiency and the ability to quickly accumulate a large amount of data. However, as the industry has developed and the scale of human data has become a feasible solution, the weight of simulation in model training has gradually decreased.
This judgment is spreading in the industry. A startup founder told the media that what really restricts the further development of embodied intelligence is real and reusable data. "In 2026, the data problem has become a core constraint that cannot be avoided when expanding the ability boundary of embodied intelligence. 2026 will be a year for the large - scale accumulation of embodied data and the market explosion."
Lingchu Intelligence has chosen to start with human data. The team is deploying a set of multi - modal data collection gloves on a large scale in Beijing to collect information such as tactile sense, multi - perspective vision, and joint angles, with the goal of long - range operation data in logistics and supermarket scenarios. According to the company, the cost of this solution can reach one - tenth of that of the whole - machine remote operation after March.
However, some investors hold a reserved attitude. People who have invested in multiple embodied intelligence companies told the media that the current orders for embodied intelligence data collection mainly come from local governments.
The data war has just begun. Who can collect high - quality data, who can train the data into the model, and who can use the data to feed back the business closed - loop - these three questions are replacing "whether the robot can do a somersault" and becoming the new competition points in the industry.
Implementation and Hype: The Real Battlefield Behind 800 UPH
Beyond the exciting financing figures, another piece of data is worth noting.
According to incomplete statistics from the Humanoid Robot Scene Application Alliance, in 2025, there were more than 292 publicly disclosed winning bids for humanoid robot projects in the Chinese market, with a total disclosed contract value of more than 1.81 billion yuan. However, 235 of these projects had an amount of less than 5 million yuan, and only four projects had a single - order value of over 100 million yuan.
Many investors and industry insiders told the media that the orders of embodied intelligence companies are "inflated". Many orders in the industry often end up with only a deposit paid and then nothing more, but when companies promote themselves externally, they still package these intended orders or framework agreements as actual delivery volumes. This is also a very embarrassing point in the industry: A large number of so - called 'commercial orders' are essentially for public relations - style display purchases and data collection cooperation, rather than real - sense productivity replacement.
Lingchu Intelligence is quite cautious in choosing implementation scenarios. The team spent nearly half a year sorting out a large number of scenarios and finally selected three niche scenarios: clothing supply, in - box inspection, and sorting wall. Wang Qibin explained that the logic for choosing these scenarios is not only whether the business can work, but also a more important dimension: Whether data can form an increment in the real scenario and whether data flow - back can feed back the model.
He revealed that in the clothing supply scenario, the team has been able to achieve generalization for more than a thousand pieces of clothing, and the highest beat can reach 800 UPH. This beat is achieved through a training process mainly based on reinforcement learning. Chen Yuanpei explained that there is a stage of "self - exploration + acceleration" in the reinforcement learning training process, and the action speed has the opportunity to exceed the upper limit of human remote operation. "The actions trained by reinforcement learning are often cleaner, more efficient, and more dexterous."
However, an objective fact is that even if the 800 UPH is achieved, there is still a long way to go to truly replace human labor. The more common situation in the industry is that robots still cannot achieve millimeter - level precision, only centimeter - level, and it is difficult to complete relatively delicate tasks such as screwing.
Miao Rentao, the deputy dean of the School of Labor Economics at the Capital University of Economics and Business, once pointed out that China's intelligent robots still lag behind international giants in core technologies such as motion control algorithms, precision reducers, and high - precision sensors. The proportion of core components of humanoid robots that rely on imports still exceeds 40%.
The Elimination Round: Whose 2026?
In this reality, what does 2026 really mean?
A startup founder's judgment has resonated in the industry: "2026 is not the 'year of commercial explosion', but a combination of the 'year of commercial verification' and the 'year of the elimination round'."
There are currently more than 200 embodied intelligence enterprises in China, among which more than 100 are humanoid robot companies. After the leading enterprises quickly obtain funds, the living space of the second - tier enterprises is being squeezed. Research institutions have observed that "some companies are actually no longer viable, but are still operating at a low speed".
After the hustle and bustle, what is truly important? Some believe that "Embodied intelligence is taking off its technological finery and putting on a 'battle robe' to enter the battlefield. The excitement of the Spring Festival Gala will eventually fade away, and the enthusiasm of capital will also return to calm. In the end, the robots that remain on the playing field must be those that are truly working in workshops, warehouses, and various industries."
For Lingchu Intelligence and all the startups that have received huge financing, the test questions for 2026 are already there: Can the data form a closed - loop? Can the model be generalized? Can the scenarios work? The answers are not at the financing press conferences, but in the customers' workshops.
This article is from the WeChat official account "Caixinquan", author: Sijiali, published by 36Kr with authorization.