Behind 14,400 units, who is footing the bill for humanoid robots?

1.44万 units shipped, 5.7 billion yuan in commercial orders, accounting for 84.7% of the global total shipped……In 2025, China’s humanoid robot industry turned in an impressive performance report.

The shipping volumes are real, and the order amounts are real. But in these orders, how many are being placed by governments and state-owned enterprises because of industrial policy? How many are paid-for test runs that automakers carry out under the name of “training”? And how many are procurement decisions made by private enterprises based on the logic of “robots are cheaper and easier to use than people”?

A year ago, Zhu Xiaohu of Sequoia China Venture Capital asked a question—one that has since been repeatedly cited: “Where are your potential commercial customers? Who would spend hundreds of thousands of yuan on a robot to do these jobs?”

A full year has passed, and this question still hasn’t been answered directly.

Orders generated by attracting investment

When you open up the biggest humanoid-robot orders of 2025, the presence of state-owned enterprises is always there.

Two domestic leading embodied intelligence companies jointly won a humanoid biped robot procurement project under China Mobile—totaling 124 million yuan, then the largest single humanoid robot order in China.

A research report from Galaxy Securities disclosed an order of 1.3 billion yuan for UBTECH. Among it, four data-collection-center-type orders from Guangxi, Zigong, Fangchenggang, and Jiujiang totaled nearly 700 million yuan, and all the procurement parties were institutions with local state-owned capital backgrounds.

Procurement by state-owned enterprises comes with internal pressure to align with industrial policy. Procurement by local governments, however, is often embedded in the exchange of interests that comes with attracting investment. Buying robots means completing the task of “supporting domestic technology.” As for what production problems the robots actually solve, that is not the core consideration of this procurement.

Wang Han, a partner at Shanghai Pudongchuan Equity Investment Fund, described the specific operating model: “Many robotics companies, riding on industry hype, when they negotiate landing conditions with local governments, will also casually require the government to support with procurement orders. They bind attracting investment with procurement together. Demand is created, not naturally grown.”

When attracting investment is swapped for procurement commitments, this chain severs the supply-side capacity numbers from the demand-side commercial validation entirely. The orders are real, but what they prove is that the policy-conduction chain is working properly—not that robots have created commercial value.

Zhu Xiaohu of Sequoia China Venture Capital said bluntly before: “Previously, the main market demand was to do research……Today there’s another new customer, and an SOE buys them back to use as a front-desk display.”

Chen Weiguang, managing partner at BlueRun Capital, is a staunch bull on the embodied intelligence track for leading investment investors. He described a more specific case: Ziyuan and automakers such as Chery cooperated to deploy humanoid robots in overseas 4S stores—“Chery thinks that when they put robots in 4S stores abroad, overseas customers will think they’re high-tech and the car is better. More often, they’re used in closed scenarios, like factories, logistics, warehousing, pharmacies, and some fun scenarios, such as 4S stores—mostly demonstrations.”

These orders typically come without service-level agreements, without strict fault-response clauses, with rare renewal records, and with the procurement party never disclosing the robots’ actual usage data. The procurement party’s KPI is “complete the procurement,” not “the robot creates value.” This distinction is the simplest standard for judging whether an order is a policy signal or a commercial signal.

The market created by such procurement has its own inherent fragility. Real industrial demand won’t fluctuate because of the viewing ratings of a gala, nor will it arise because a city leader visits.

Interns on the production line

The term “training” runs through robot company announcements, management interviews, and sell-side research reports. The wording is highly consistent, yet the meaning is rarely questioned.

In a report titled “The Year of Mass Production: A Hundred Schools Contend” released by Orient Securities, it describes UBTECH Walker S1 entering companies including BYD, Geely, and Foxconn. The modifier used is “for training,” and it points out the need to “gradually achieve mass production through 18 to 24 months of production-line training.” A bullish report titled “The Year of Mass Production” uses the core verb for the automaker scenario as “training,” not “deployment.”

Training and deployment are entirely different in nature. Deployment means delivering productive capacity from Party B to Party A; training means the robot is trained as the subject in the factory, while the factory provides real scenarios and operational data—only then does the robot company become the actual beneficiary.

From a business logic standpoint: in the paid training stage, the factory is helping the robot company do R&D. It’s just that this R&D cost is sometimes borne by the factory, sometimes reflected in a nominal contract amount, and sometimes packaged as “strategic cooperation.” There’s a saying in the industry that directly supports this judgment: “What’s signed with the automaker isn’t a sales contract, but a strategic cooperation agreement.”

Take Dongfeng Liuzhou Automobile, for example. An institutional research report labels the order for 20 Walker S1 units as “most already delivered,” and the figure is plainly included in the shipping-statistics breakdown. But break it down: the scale is 20 units, the scenario is an automobile factory, and what’s being used is the previous generation S1, not the latest S2. Every dimension points to the same conclusion: this is paid training, not large-scale deployment meant to replace labor.

What’s even more telling is that, among all publicly disclosed automaker cooperation cases currently available, none discloses the actual operating duration, fault rate, or task-completion rate. In real commercial deployments, customers would not keep such data secret.

Why can’t training be converted into deployment for a long time?

An industry observer close to actual factory operations pointed out: “Even achieving a 99% success rate on specific actions might not be enough, because machines do大量 repetitive work every day, and a 1% failure rate will keep compounding; each failure could cause the production line to stall. In simulated scenarios, if success reaches 90%, on-site it might only be 60%.”

The fault-tolerance logic of industrial production lines is completely different from that of a laboratory. The former pursues zero-fault stability, not average success rates.

Han Fengtao, founder of Qianxun Intelligent, described this difficulty more bluntly: “The embodied intelligence industry is just starting, and humanoid-robot hardware is also just starting. Two very early-stage technology streams are being fused to do a very complex task—that’s extremely hard.”

The motive for automakers running such pilot projects is also worth scrutinizing. Companies like BYD and Xiaomi themselves are laying out their robot businesses; in large part, procurement competitors are doing competitor research. Some automakers, under the pressure of industrial policy in their cities, include local robot companies’ products as part of the industrial-chain coordination. None of this is driven by the commercial logic that “robots can solve production problems.”

There is a standard for judging whether an order is a pilot or real procurement: who is the decision-maker. The decision-makers in automaker pilots are usually the innovation department or strategic investment department, with assessments focused on technical feasibility. For real industrial procurement, decision-makers are the production and manufacturing departments, with assessments focused on cost-replacement ratio and capacity improvement. If the first fails, the project ends; if the second fails, they must assume commercial responsibility.

By this standard, none of the publicly disclosed automaker cooperation cases to date has cleared the threshold.

Buyers have gone into hiding

To measure whether an order constitutes real commercial validation, the standard is not complicated: the procurement party makes an independent decision based on a cost-replacement logic, with clear task definitions and acceptance criteria, commercial consequences if it fails, and willingness to publicly endorse. All four conditions are required—none can be missing.

Using this standard to review all currently disclosed cases, it’s almost impossible to find a complete closed-loop.

Wang Qian, founder of the robot business, has made the most direct characterization of the current situation in public: “There’s only one standard for measuring commercialization: generating a positive ROI for the customer. When customers buy robots to replace labor—whether it improves efficiency or improves productivity for a longer time—if it can achieve that, it counts. But today, in the market, none of the systems can do that.”

He also further pointed out that those companies that have claimed commercial rollout and revenue over 100 million yuan, in essence, are still doing R&D and education markets and hosting welcome/entertainment performance markets—going into factories to do simple repetitive work is “actually a PR-type activity.”

This judgment has also echoed within the capital circle. An investment partner at Sequoia China, Yuan Geng, said: “A company can’t rely solely on doing demos and demonstration projects. It must find a true commercialization path that creates value… The total cost for a robot to complete tasks must be lower than the local labor cost, and the completion quality must be better. Only then will someone actually pay.”

The subtext is very clear: in his view, what many companies rely on to survive today is demos and demonstration projects, not real commercial value. Such a judgment comes from someone who has continuously invested in this industry, which implies that investment is not based on a clear commercial logic.

This can also be confirmed by the way companies enter factories themselves. As described by industry observers close to multiple vendors, factories will not proactively spend money to bring a batch of “robots that can’t do anything and will slow down production” onto the production line.

Therefore, “who gets you into the factory” is the primary question, not “what the robot can do.” The current paths that get you into factories mainly include: shareholder relationships, technology endorsement needs from strategic cooperation partners, interest exchanges across the upstream and downstream of the industrial chain, and governments acting as matchmakers. Among this list, the path “because robots can replace labor and factories autonomously procure based on economic rationality” currently does not exist. The logic for entering factories is driven by relationships, not by demand pull.

At the current stage, the application of humanoid robots is still in the exploratory early phase, mainly piloted in some companies that have demand for automation and intelligence… Their large-scale adoption is still constrained by factors such as cost, technology, and market acceptance.

The industry’s number-one shipped-volume company uses “exploration” and “pilots” in its listing documents—not as modesty, but as the most accurate official description of the current “quality” of commercialization.

The closest case to a truly real commercial order is that a certain leading robot brand provides in-room services at the Marina Bay Sands Hotel in Singapore, and provides traveler navigation in Arabic at Dubai Airport. But neither of these cases disclosed any subsequent data: number of deployed units, contract value, fault records, or renewal situation. Neither the hotel nor the airport has made any public statement.

This kind of silence itself is a signal. Only customers who have truly made the commercial logic work would have the motivation to publicly say, “I bought how many units, used them for how long, and saved how much money.”

The absence of data means validation hasn’t happened.

Technical capability is the hard constraint

The problem with the order structure ultimately comes back to technical capability.

If a product can’t sell real commercial orders, it’s because it’s either too expensive, or not useful enough, or both at the same time. The current predicament of humanoid robots is that both conditions are simultaneously true.

All parties’ assessments of humanoid robots’ efficiency in today’s industrial scenarios are highly consistent. UBTECH’s Chief Brand Officer Tan Min gave a figure of 30% of labor efficiency, and expects it to exceed 50% by early 2026. Leju Robotics says industrial-scenario efficiency is close to 50% of labor. Xingdong Jiyuan claims that some real industrial scenarios have already reached above 70%. Morgan Stanley’s independent research report puts the figure at 30%.

There’s a clear gap between vendor self-reporting and third-party evaluations—which in itself is already common. But even if you accept the most optimistic vendor figures, the conclusion isn’t encouraging: if 70% efficiency means robots need 1.4 times the human time to complete the same tasks; if you believe Morgan Stanley’s 30%, then it needs 3.3 times. None of these numbers supports the procurement logic that “robots are cheaper and easier to use than people.”

Beyond the numbers, on-the-ground reality is even more direct. Observers close to robot demonstration operations described a detail: the so-called “robot takes care of people,” but in reality it’s often “3 people taking care of 1 robot.” In the demo backstage, you can frequently see a row of robots lying on the ground waiting for manual intervention. After watching a project demo, Wang Han remarked: “Just organizing shoes took one minute for identification and computation. If it can’t even do basic household chores, how can you talk about reaching ordinary households across the country?”

Insufficient efficiency combined with high costs produces a rather grim ROI calculation.

A typical procurement price for a humanoid robot is currently around 100,000 yuan. If it replaces one worker earning a monthly salary of 5,000 yuan, the payback period based only on hardware procurement costs would be about 20 months. And the fact that efficiency is only 30% to 50% of human labor means that to achieve the same output, you actually need 2 to 3 humanoid robots. After conversion, the full investment payback period would exceed 40 to 60 months. This hasn’t included maintenance expenses, loss from faults, and additional integration costs caused by the industry’s very low current level of standardization.

The latter is often underestimated. UBTECH’s Chief Brand Officer Tan Min described it directly: “Current industry standards for humanoid-robot hardware and software aren’t fully clear, and they’re nowhere near truly standardized production. There are no standardized interfaces across the different components, which makes integration costs and process complexity quite high. If volumes are high, it’s unrealistic to rely on engineers ‘handcrafting’ it.” This means that even if the cost of the robot itself declines, the hidden cost of system integration remains high.

When will costs come down?

Research data shows that by 2035, global robot component prices are expected to drop by about 70%. This implies that a cost-rationality-based labor-replacement procurement decision will only become viable at scale in the mid-2030s. Before then, who is buying robots, and what they’re buying them for—the answer is most likely not “factories autonomously procure based on cost-reduction logic.”

This judgment is already affecting the configuration logic of some investors. Wang Han candidly said that in the embodied intelligence sector he deliberately avoids the complete machine and instead focuses on the upstream: “From the perspective of suppliers or spare parts. Because this demand is the most real: what kind of spare parts the downstream actually needs is pretty clear.” Choosing the upstream and avoiding the complete machine is a rational judgment: the commercial logic of the complete machine hasn’t worked yet, but demand for components is real.

He has a specific reference for “real demand”: “Industrial robots—like Mecamand—truly improve efficiency and solve specific problems, such as sorting, and therefore they have value. For embodied intelligence, aside from some flashy demo functions, can it solve real needs? If one day it can, then customer recognition will likely show up. The core issue right now is that it’s still in the technology development stage; the underlying demand can’t really be seen.”

Fixed scenarios, single tasks, and measurable reliability are precisely what humanoid robots currently lack the most. Generality is the selling point of humanoid robots, but it is also their burden: the more general it is, the harder it is to meet industrial-grade stability requirements in any specific scenario.

An internal narrative in the industry is quietly changing. An investor described it this way: “This wave of grand narratives about building a big general embodied model has basically already come out. Starting this year, you’ll gradually see more embodied efforts focused on landing in scenarios… a gradual wait for the day embodied intelligence matures.”

When the “general humanoid robot will overturn everything” narrative becomes hard to sustain for financing, the industry shifts toward smaller, more specific deployed scenarios—logistics sorting, pharmacy dispensing/pick-up, and warehouse handling and transport. These scenarios share characteristics: the environment is relatively structured, tasks are relatively single, and there is comparatively more room for error.

But the narrative that supports current valuations has never been “humanoid robots dispense medicine in pharmacies.” It has been “humanoid robots enter factories and fully replace human labor.” The former is approaching reality, while the latter is still far away. The gap between the two is exactly the thickness of the bubble in today’s valuations.

Source of this article: NashNova

Risk warning and disclaimer terms

        The market involves risk; investments should be made cautiously. This article does not constitute personal investment advice, nor does it consider any individual users’ special investment objectives, financial conditions, or needs. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at your own risk.