CBN News 2026-10-08 13:23:03

Author: Zhou Daohong Editor-in-Chief: Zhang Jian

(The author of this article is Zhou Daohong, the chairman of Shanghai Shengshi Capital Management Co., Ltd.)

In the first half of 2026, the global shipment volume of humanoid robots approached 25,000 units, a year-on-year increase of 432%, with China accounting for over 97% of that figure. The Ministry of Industry and Information Technology expects the annual production to exceed 100,000 units, compared to just 20,000 units in 2025. According to industry research, Yushu Technology's revenue in 2025 exceeded 1.7 billion yuan, with a net profit margin of 35%; Zhiyuan Robot's cumulative shipments of humanoid robots surpassed 10,000 units; AMD acquired World Labs founded by Li Feifei for $8.2 billion, which is regarded as an important capital event in the physical AI sector.

It seems that everything is accelerating.

But beneath the facade of acceleration, the industry is repeatedly torn apart by several irreconcilable contradictions: the pace of hardware iteration exceeds intelligent capabilities, and mass production scales are rising rapidly, yet market orders have not kept up accordingly; China has a prominent advantage in robot body manufacturing, while the United States still holds a head start in underlying world models and software ecosystems. The humanoid robotics industry in 2026 is not lacking in opportunities, but rather in certainty—every seemingly correct direction hides a fundamental question that remains unanswered.

I. Should we strengthen our minds first or build our bodies first?

The deepest divide in the humanoid robotics industry is not between companies, but between different approaches or strategies.

One faction advocates for "hardware first": focus on making the robot stable, affordable, and capable of producing energy first, with intelligence being improved gradually over time. Yushu Technology, UBTECH, and Boston Dynamics follow this approach. As a result, Yushu has become the global leader in shipments, and UBTECH's Walker S robots have entered the factories of BYD and Jikr. However, UBTECH only managed to sell 921 full-size humanoid robots in the first half of the year, resulting in a loss of 339 million yuan – although the production capacity was there, customers did not keep up.

Another faction advocates for "intelligence first": build the brain first, and the body can be added later. Tesla, Zhiyuan, Xingyuanzhi, and Galaxy General are representatives of this approach. Galaxy General secured thousands of orders from companies like CATL and Bosch by relying on its heavy-duty robots Galbot S1 which can work continuously on the production line for over 3 months without rest. They may not be the cheapest, but they are indeed the most capable.

The author believes that both sides have their reasons, as well as their underlying concerns.

The hidden concern with a hardware-first approach is that the overall gross profit margin for manufacturers is generally only between 8% and 12%, while the gross profit margin for the actual manufacturing of the hardware itself ranges from 15% to 25%. Moreover, the intensity of price competition is increasing. Once open-source blueprints and public-domain solutions become widespread, simple assembly will quickly lose its barriers to entry, leading to homogeneous competition and involution. The gross profit margin of contract manufacturing companies' robotics businesses is even negative: Dongfang Precision's intelligent hardware business has a gross profit margin of -1.20%, accounting for only 1.73% of its revenue. The more they produce, the more they may lose.

The hidden concern with intelligence prioritization is that, no matter how advanced the brain is, it still requires a physical body to function effectively. Physical AI does not yet have a "ChatGPT moment," and it will take time for world models to move from the laboratory to industrial applications. During this transition period, pure algorithmic companies without hardware capabilities may fail before dawn.

A deeper confusion arises: these two paths are not a binary choice of either/or, but rather a sequence of priorities over time. The author argues that in the short term, within one to two years, the hardware-first approach will first capture the market; in the long term, over three to five years, the intelligence-first approach will build barriers to entry. In the end, it is very likely that a "integration of software and hardware" strategy will prevail, just like today's smartphones—these require both powerful chips and operating systems, as well as excellent industrial design and supply chains. However, the concept of an "endgame" seems too distant for every company that is currently burning money.

II. Is the humanoid form truly the optimal solution?

The term “human form” is both the biggest narrative in this field and the greatest trap as well.

The logic behind supporting a humanoid form is this: the world is designed for humans—doorknobs, stairs, kitchen countertops, tools, switches, social distancing guidelines—all are tailored to the human body and habits. The greatest advantage of humanoid robots is not their efficiency, but their ability to utilize human-made spaces and tools without the need to modify the environment. In unstructured scenarios such as home use, elderly care, and services, they naturally possess a value as a kind of “universal hardware interface.”

However, the author points out that there is a fatal blind spot in this logic: having a human form does not equate to possessing general intelligence. Smartphones, which do not have a human form, are nonetheless the most successful intelligent devices. In the future, home services may not necessarily be entirely handled by bipedal humanoid robots; rather, it is more likely that a distributed system will be in place, involving collaboration between vacuum cleaners, mobile chassis, robotic arms, voice assistants, sensors, and cloud-based brains. A wheeled chassis with two arms and a lifting mechanism might prove to be more practical, cheaper, and safer than bipedal robots.

In industrial scenarios, humanoid robots are usually not the optimal solution. Industrial environments can be optimized for efficiency: fixed production lines, specialized robotic arms, AGV, collaborative robots, which perform better in terms of cycle time, precision, reliability, and cost-effectiveness. The appropriate applications for humanoid robots in industry include handling multiple varieties in small batches, frequent line changes, dangerous tasks, and remote operations—they serve as a "flexible supplement," not a replacement for efficiency. The author also proposes that in the future, home scenarios might follow the same principle of prioritizing efficiency, emulating past industrial settings by designing homes in a way that combines "human habitation with convenience for AI work," rather than adapting existing human activity environments to fit humanoid robots.

This is the core dilemma of being "both left and right": Industry is not the optimal solution for humanoid forms, but it is currently the only scenario where a commercial closed loop can be achieved; the family represents the ultimate value of humanoid forms, yet it may still take three to five years, or even more than ten years, to realize this potential. Qian Dongqi, the chairman of iRobot, is more pragmatic in his assessment—he is "optimistic for three years, pessimistic for five years." The current manufacturing cost of such robots is about 100,000 yuan, and the additional cost for computing power solutions adds another 20,000 to 30,000 yuan, which is simply unaffordable for ordinary families.

The author believes that a more realistic approach might be as follows: in the short term, to conduct flexible pilot projects for specialized automation and industrial logistics; in the medium term, to develop wheeled semi-humanoid robots for commercial services and elderly care inspections; and in the long term, to return to providing general services for home scenarios. Having humanoid robots is not wrong, but regarding “humanoids” as the “only solution for universal intelligent terminals” and the “ticket to enter the largest market in the near future” is debatable.

Home is the final destination, not the starting point.

III. Scaling Law Hits a Hard Wall in the Physical World

If the debate over "the body" is still merely a dispute over different approaches, then the disagreements regarding "the brain" have escalated to a clash of paradigms.

In 2026, there was a fundamental technological shift in the field of embodied intelligence: from training models to inference models. The Beijing Zhiyuan Research Institute clearly stated in "Top 10 AI Technology Trends of 2026" that AI is shifting from "predicting the next word" ( Next Token Prediction ) to "predicting the next state of the world" ( Next State Prediction ). The industry no longer aims for robots to memorize action trajectories by rote; instead, it requires them to perform physical deductions in their "minds" before taking action—predicting the consequences of their actions, assessing environmental changes, and dynamically adjusting strategies, just like humans.

Professor Su Hao from Fudan University summarized it quite accurately: “The essence of physical understanding is not ‘what I see’, but ‘what will happen if I do something’.”

According to the author's analysis, there are three major flaws that prevent the pure training approach based on Scaling and law (the scaling principle) from being viable:

• Firstly, the data gap cannot be bridged by sheer quantity. The physical world is characterized by high noise and redundant information. Yao Maoqing, a senior executive at Zhiyuan Robotics, pointed out that for robots to perform daily tasks, the amount of data required is more than ten thousand times that of the pre-trained language model corpora. At a minimum, hundreds of millions of hours of qualified data are needed to achieve a basic success rate. However, the amount of compliant data from real-world physical interaction scenarios in China currently amounts to only about 500,000 hours, representing a gap of over 99%.
• Secondly, fitting a trajectory does not equate to understanding physics. The essence of traditional imitation learning is to teach robots ‘how to do’; the models only remember the actions but do not learn ‘why’. Once the scenario changes, or in cases not covered by the training data, such as when the handle of a cup breaks, the robot immediately becomes ineffective.
• Thirdly, the cost of physical trial and error is extremely high. In the digital world, if a model makes a mistake, at most it will produce distorted images. However, in the physical world, predictive errors can lead to equipment damage or even personal injury, and it is not possible to make unlimited attempts at trial and error like with large language models.

Thus, the industry collectively turned to reasoning models. The deeply intelligent PhysBrain 1.0 injects human-first perspective data into physical causality, surpassing industry benchmarks with a success rate of 80.2% in SimplerEnv tests; the Daxiao Robot's Enlightenment World Model 3.1 integrates generative intelligence, physical intelligence, and cognitive intelligence within a single architecture, possessing autonomous reflective capabilities; its AWE 3.5 model is capable of deducing future states in hidden spaces, supporting several minutes of continuous interactive closed-loop reasoning.

In July 2026, the autonomous mobile robot safety assessment standards released by ISO for the first time listed "physical common sense reasoning" and "long-term temporal task planning" as mandatory certification indicators for L4 level embodied intelligent agents. The author stated that the standards have now set a clear direction.

But confusion still persists: the prospects for reasoning models are clear, yet the path forward is not. Each technical approach is still in the verification stage, and whoever manages to make it work first will gain control of the “operating system” for embodied intelligence – however, there is no consensus in the industry yet on what exactly constitutes “making it work.”

IV. The bottom of the smile curve is collapsing.

An irreversible shift is taking place in the industrial value chain: from a "hardware shell" to an "intelligent core."

Industrial data for 2026 shows that the value proportion of traditional hardware shells has decreased from over 60% to around 40%, while the value proportion of intelligent cores—models, data, and dexterous hands—has risen to 60%. The manufacturing segment is becoming the bottom of the “smile curve”: it is large-scale and generates revenue, but the profits are extremely thin.

Large model companies are systematically entering the field of physical AI. OpenAI established a robotics team in June 2026, adopting a strategy of "simulation first, build the brain before adding the body." Google DeepMind's Gemini Robotics covers full-body control and task planning. NVIDIA's Cosmos, GROOT, and Isaac Sim constitute the infrastructure platform for physical AI. AMD's acquisition of World Labs is essentially buying an entry ticket to the "physical AI infrastructure layer" for $8.2 billion.

Domestic efforts are also accelerating. SenseTime's Daxiao Robot has launched the "Enlightenment" World Model 3.0; Zhiyuan's GE 2.0 ranked first in the WorldArena leaderboard with a total score; Jijia Vision's GigaWorld-1 is the first embodied world model to exceed a comprehensive score of 60 points. Yushu Technology received a strategic investment of approximately 140 million yuan from DeepSeek, and the two parties are jointly developing large models and embodied intelligent products.

However, the author also makes a key correction: the giants of the future may not necessarily be “pure large model companies,” but rather “embodied model native companies.” Capital has already voted with its actions; in the financing in the first half of 2026, the top-ranked embodied intelligence companies were Zixing Robot (6.3 billion), Zhifang Square (6 billion), and Qianxun Intelligence (4.5 billion), rather than traditional large model companies.

The manufacturing process will not disappear, but the roles involved will change fundamentally. The author compares this to the smartphone industry: Foxconn's profit margin is not in the same order of magnitude as Apple's, yet Foxconn still exists and is of a huge scale. It is very likely that the manufacturing process for humanoid robots will follow a similar pattern—only the real barrier lies in the data loop: whoever can continuously obtain high-quality, real-world physical interaction data to feed into model iteration will be able to establish a competitive advantage. And this is precisely the strength of companies at the model layer—they can use software to define data collection standards, allowing hardware manufacturers to "work" for them in collecting data.

V. Five Tough Curves to Navigate Through Confusion

The author believes that confusion itself is not the problem; the real issue lies in what to focus on and what to ignore amidst that confusion. When looking at the humanoid robot industry, it is not sufficient to merely consider the actions, performances, and shipment volumes presented at press conferences. Instead, there are five key indicators that deserve closer attention:

• Firstly, regarding the yield curve, it is uncertain whether it can stably progress from prototype development to consistent mass production deliveries. Unlike automobiles which have a standardized chassis, there is no absolute convergence in the hardware design of humanoid robots, making it difficult to ensure consistency in large-scale production.
• Secondly, regarding the cost curve, whether the actuators, dexterous hands, and sensors can continue to reduce costs. Calculations show that only when the overall cost of the system is reduced to less than 400,000 yuan and the operational efficiency reaches over 75%, does it become economically viable to replace human labor.
• Thirdly, regarding the data curves, it is essential to have a real-world scenario for validation, rather than relying solely on simulations. Real device data is becoming the core driving factor, while simulated synthetic data is used only as an auxiliary for pre-training.
• Fourthly, regarding the order curve, whether industrial customers will shift from trial use to repeat purchases. Customers are willing to try the product, but they remain cautious about large-scale purchases. The transition from “Demo” to “Deploy” represents the true commercialization inflection point.
• Fifthly, regarding the scenario curve, whether it shifts from showcasing and performing low-value activities to replacing them with high-value tasks. The entertainment industry can recover its costs quickly, but industrial scenarios represent the long-term value.

Conclusion

The author believes that the field of humanoid robots has never been a binary choice, either left or right.

Should we focus on strengthening our minds first or our bodies first? Both are important, but the time frames for each are different. Is the human form the optimal solution? It depends on the context; it might be suitable for family settings, but not necessarily for industrial applications. Should we train models or inference models? The industry has already provided some directional answers, but the specific paths are still being explored. Is it more valuable to create tangible products or to develop intelligence? The answer is becoming clearer, but the pain of transformation has just begun. Is China leading or is the United States leading? In this case, it's a matter of competitive differentiation, with each having its own strengths.

The article states that 2026 represents a convergence of "the first year of mass production" and "the eve of commercialization" for humanoid robots. The hardware supply chain and mass production capabilities are already ahead, while embodied intelligence, data systems, and the stability of real-world scenarios are still in need of improvement. The key lies not in choosing one path over another, but in finding that narrow path between them: using orders from industrial scenarios to reduce costs, feeding models with data from commercial applications, and accelerating validation through the openness of policy-driven environments. It's about developing skills while "working," rather than waiting until one is fully competent before taking action.

The author concluded that the confusion surrounding embodied intelligence is essentially an inevitable challenge faced by emerging industries after they have completed the “from 0 to 1” phase and are just beginning the “from 1 to 10” phase of development. The direction is clear, but the path is unclear, and the road ahead is winding. It is precisely in times when the path is uncertain that all great industries forge ahead through twists and turns to create new territories.

First published exclusively by Yicai.com. This article represents solely the views of the author and does not constitute investment advice.