"Foreman" of Embodied Intelligence.

Author丨Wang Manhua

Source丨Investment Circle

In late August, "the first humanoid robot stock" YuTree Technology officially landed on the Science and Technology Innovation Board. However, the extremely high attention brought by Huang Renxun’s personal appearance and Lenovo’s emergency investment did not allow YuTree Technology's stock price to continue the wealth creation myth. On the 11th trading day after the listing, YuTree Technology's stock price fell below 550 yuan during the session, "halving" from its peak of 1100 yuan.

At the same time, another group of companies behind "YuTree" are quietly making profits. They do not manufacture robots, nor do they research models; they only do one thing: hire people to "feed" data to robots.

90 billion hot money, who is "eating the meat"?

Despite the huge fluctuations in stock prices, the quick listing path of YuTree Technology over 73 days is undoubtedly a microcosm of the embodied intelligence industry so far this year. According to IT Juzi statistics, in the first half of 2026, there were 322 financing deals in the domestic embodied intelligence track, totaling 93.5 billion yuan, a year-on-year increase of about 5 times. As the capital market strongly pays attention, hot money has also begun to flow into leading companies in embodied intelligence.

In March this year, Galaxy General announced the completion of a new round of financing of 2.5 billion yuan, breaking the record of 300 million dollars (about 2.1 billion yuan) for a single round of financing in China's embodied intelligence set last December. In April, it was Stone Zhihang which secured 455 million dollars (about 3 billion yuan) in a single round of financing, refreshing China's financing record for embodied intelligence. By the end of June, Zhifang announced the completion of nearly 5 billion yuan in a new round of financing, with a post-investment valuation exceeding 20 billion yuan, becoming the first 20 billion-level embodied intelligence unicorn in the Guangdong-Hong Kong-Macao Greater Bay Area.

And just before YuTree Technology's listing, Zhiyuan Robotics, established in 2023, formally confirmed in July that it had started the IPO process for Hong Kong stocks, rumored to be valued at least 40 billion Hong Kong dollars. According to statistics from Yicai, in the first half of this year, there were already 22 embodied intelligence companies in the country valued at over 10 billion yuan.

However, the capital is betting on the "future," and hot money does not equate to profit. YuTree Technology's prospectus (meeting draft) shows that in the first quarter of this year, its net profit after deducting non-recurring gains and losses was 40.2536 million yuan, a year-on-year decrease of 52.55%.

In March this year, Liu Yang, co-founder of Unlimited Force, said in an interview with the Daily Economic News that "in 2025, a large number of so-called 'commercial orders' in the industry will essentially be public relations-style display purchases and data collection collaborations, rather than true productivity replacements." Liu Yang's statement also subtly reveals a fact: within the embodied intelligence industry chain, the first to reap the benefits is not the robot companies, but the data collection companies.

Data collection service provider—Sun Xiao, project head of Hangzhou Shuxi Technology, told Jingzhe Research Institute that previously his company specialized in intelligent driving data labeling. In May of this year, it unexpectedly learned that there was a surge in data collection demand in the embodied intelligence track, leading to a decisive entry into data collection business.

According to Sun Xiao, although his company has formally engaged in data collection business for less than two months, it has already set up three bases in Suzhou, Suqian, and Nanyang and is currently expanding the team of data collectors vigorously. "Currently, the gross margin of this business can reach around 30%, which is higher than many traditional industries. Moreover, the orders are so numerous that we simply cannot finish them."

Why are most robot companies still in the red while data collection companies can eat first? The answer lies in the underlying logic of embodied intelligence. Embodied intelligence can be simply understood as a combination of hardware and software. The hardware includes the robot body, joints, dexterous hands, etc., while the software is an intelligent system driven by a large embodied model that allows robots to understand the physical world and generate action commands.

In recent years, hardware technology has made rapid advancements, but software technology's development is stalled by data. Not long ago, a set of frequently cited data at the World Robot Conference stated: currently, the compliant data of real physical interaction scenarios in the country is only 500,000 hours, while the commercial implementation of robots requires tens of millions of hours, creating a gap of over 99%. Behind these numbers lies the structural contradiction faced by the embodied intelligence industry and the phased opportunities released by the development of the industry chain.

The China Academy of Information and Communications Technology published the "Research Report on Embodied Intelligence Training Grounds (2026)" which shows that as of the end of June this year, over 70 robot training grounds have been built and put into use nationwide, with 46 more under construction or planned. Besides the officially disclosed training grounds, more data collection service providers have also found their own ecological niches.

Embodied Intelligence also has "foremen"

The explosive increase in data demand has caused orders to fly to data collection companies like snowflakes, but the essence of this business is a labor allocation that involves layers of subcontracting.

"Many of our businesses are not directly connected to upstream robot companies, but rather, contractors first take orders from robot companies and then subcontract to us," Sun Xiao told the Jingzhe Research Institute. "The bids given by robot companies for claw data collection are usually over 100 yuan per hour. After a middleman takes their cut, our price is about 70 to 90 yuan per hour, and then our data collection company settles with frontline collectors at a price between 40 and 70 yuan." Clearly, the profit of data collection companies comes from the price difference between upper-level contractors and the settlement prices with frontline collectors.

"When we first started doing data collection in Suzhou, we found that the cost calculation for personnel salaries and employee social insurance was very unprofitable. So we later began to move to third and fourth-tier cities, or even more remote areas. Places like Suqian and Nanyang have cheap rent and lower wage levels. We can hire collectors at around 160 yuan a day, with social insurance paid after two months of work. If we set our headquarters in Beijing, Shanghai, Guangzhou, and Shenzhen, the same kind of manpower configuration would cost more than double."

Regarding the ecological niche of data collection companies, Sun Xiao is quite frank, "We are essentially like foremen in the industry chain. Although this track is still very early, from a historical perspective, we can definitely get a piece of the cake from the growth of the embodied intelligence industry."

Of course, the premise for the "foreman" to make money is that there are people willing to work. From what Sun Xiao shared, the recruitment threshold for data collection companies is surprisingly low: just literacy, no educational or gender requirements, and novices can get started in three days. "Collecting 4 to 5 effective hours a day easily earns around 160 yuan per day, with four days off in a month, making at least 3,800 yuan a month. For young people in third-tier cities, this income is more attractive than working as a cashier in a supermarket or shaking milk tea in a milk tea shop."

However, Sun Xiao also pointed out that although this job is not as tiring as piecework on assembly lines, it can sometimes feel monotonous and tests patience. "Because the essence of data collection is to provide training data for robots, robots (the big model behind them) cannot understand what the collector's action is at the beginning, so they require collectors to perform actions slowly. This leads to even a simple action taking a long time or being repeated many times, and some people lose patience and cannot continue."

The performance of frontline collectors also affects the revenue of data collection companies. Since upstream contractors use effective hours as the settlement standard, most data collection companies will expand their workforce to increase production capacity. However, increasing the number of workers does not automatically increase production capacity; differences in personnel skills, project requirements, and data collection equipment can all influence collection outcomes, especially "human issues."

"We roughly estimate that a frontline collector works 8 hours a day, but can actually produce only about 4 hours of data. The best teams can achieve a production factor of 0.6, equivalent to producing 4.8 hours of data in a day. There are variations among individuals – some treat it like a job, while others see it merely as a part-time gig. We previously allowed home collection, and some people could produce 6 hours of data in a day, while others couldn't even reach the basic 3 or 4 hours."

Sun Xiao confessed that on the surface, data collection seems to rely on human heads to build capacity but is exactly constrained by "people." Lower thresholds for hiring imply that scaling up needs to be premised on managerial abilities. Therefore, to complete orders on time, the company Sun Xiao works for ultimately chose to concentrate personnel together and has two shifts a day.

"If we only work 8 hours a day, according to the upstream data gap, it could take at least 5 to 10 years to meet the demand, and robot companies cannot wait. So if the effective hours of a piece of equipment do not reach 4 hours a day, they will reclaim the data collection device. JD.com previously said it would set up a data collection base in Suqian for 100,000 people, but the local progress has not been as fast as expected. I suspect they have also encountered problems with data capacity and quality." Sun Xiao said.

Waiting for a "GPT moment"

To enable robots to have a "brain" comparable to that of humans, foremen like Sun Xiao’s company began to work hard, while the focus of competition in embodied intelligence is also gradually shifting to the scale of data, "just like GPT, the quality and scale of training data also have a significant impact on robotic models."

In August this year, Dyna Robotics, founded by three Chinese, pre-trained the DYNA-2 model with over 1 million hours of human first-person perspective videos. The result showed that the success rate of tasks in high-precision manufacturing scene tests increased from 20% in the previous generation model to 90%. This means that physical AI now possesses its own Scaling Law (which refers to the law that model performance improves exponentially with increases in parameters, data volume, and computing power), and the "brain" of robots can evolve like large language models by "stacking data."

Sun Xiao said, "Currently, the embodied intelligence industry has formed an almost paranoid consensus: the collection time must be absolutely sufficient; it can be redundant but must not be lacking. The ego (first-person perspective) data demand from our upstream contractors has reached the level of millions of hours, with claw-type data also reaching over 100,000 hours. So I believe that compared to external rumors of a 'tens of millions of hours data gap', the real data demand will only increase."

He also mentioned that the upstream industry's demand for high-quality data from the real world has spawned an interesting phenomenon of inversion – "the smarter your robots are, the more manual labor they require."

"What we are currently doing is merely quantitative work; after submission, the actual data still needs to go through cleaning, labeling, and other processes before it can be genuinely useful. Therefore, the comprehensive cost of one hour of data provided to model training can range between 300 yuan and 500 yuan. I have spoken with upstream algorithm engineers, and one usable piece of data for real machines costs between 1,200 yuan and 1,300 yuan, with some even reaching 1,600 yuan." Sun Xiao stated.

"However, you should not just look at the high price, as they have very high requirements for real scene applicability. For robots to enter thousands of households, they need credible real-world scenarios as raw material. But which data company can suddenly provide dozens of ready houses for you to do data collection? Nevertheless, I know that there is now a new model similar to group buying where several robot companies connect with the same data collection training ground, and everyone uses their devices to collect data in the same scene."

Additionally, there are communication difficulties, high costs, and execution challenges in real-world scenarios like small supermarkets, pharmacies, and breakfast shops located at the entrance of communities where data collection is needed. Sun Xiao believes that currently only standardized scenarios like factories can be replicated in large volumes, while the demands for long-tail scenarios like home life and food retail remain undecided.

The surge of hot money contrasts with the lingering "internal circulation" disputes in the industry. "In some places, units purchase robots, have data collection companies collect data, and then sell it back to robot companies. This way, robot manufacturers obtain training data and can boast bigger stories to capital markets, while local units also meet assessment targets, creating a sense of mutual order brushing."

Sun Xiao also cited the previous example of Nvidia and OpenAI, "I took your investment and then bought your graphics cards to train even stronger large models, mutually enhancing each other's valuations." Currently, the embodied intelligence track exists in a realm where the primary market tells stories, and the secondary market buys expectations, while data collection companies are genuinely "moving bricks."

From the perspective of industrial evolution, the embodied intelligence industry is still far from its "GPT moment," facing three major turning points: technically, it still needs tens of millions of hours of data; operationally, it is transitioning from small-scale trials to large-scale implementation; economically, most companies are still in the red, and the business loop has not yet been completed.

But it is undeniable that the influx of hot money has brought robots into the real world in a tangible and vivid way. Those "data collection factories" continuously emerging in third and fourth-tier cities are laying the foundation for the comprehensive prosperity of embodied intelligence. Perhaps robots are still burning money, but those who "feed" them have already taken a bite.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。