Author | Lin Dengwan Editor | Shi Jie Editor | Wang Yue
When an entire industry pivots collectively in the same autumn, it is usually not because everyone sees an opportunity at the same time, but because everyone loses other options simultaneously.
On October 9, the China National Intellectual Property Administration published a design patent for an embodied intelligent robot, with BYD Company Limited as the patentee.
The patent drawings show a black-and-white color scheme for the body, a rounded curved head without an external display, and densely distributed joints on the limbs. The description of its use in the specification is plain, simply stating it is an intelligent robot with walking, balancing, and interactive functions.
This is "Xiao Di", initially rumored to be codenamed "Yao Shun Yu", BYD's independently developed embodied humanoid intelligent robot. In early August, it made its real-machine debut at Di Space in Zhengzhou. Official data shows it stands 1.61 meters tall, features 31 degrees of freedom (DOF) in motion, and its dexterous hand has a repeat positioning accuracy of ±1 mm. BYD plans to deploy 2 to 3 units in each of its stores.
Peers are taking action as well. XPeng's Iron has entered trial production at its Guangzhou factory, Xiaomi's robot has entered the Beijing Yizhuang auto factory, and SAIC's Nengzai No. 1 has taken up posts on the Buick battery production line.
The "humanoid building" business is booming, but the financial reports for the same period tell a different story. In the first half of 2026, BYD's revenue was CNY 344.815 billion, a year-on-year decrease of 7.13%; net profit attributable to shareholders was CNY 12.325 billion, down 20.54%; sales volume was 1.8085 million units, down 15.72%. Calculated by sales volume, the profit per vehicle is about CNY 6,200. Wang Chuanfu stated that the industry is undergoing a brutal elimination race.
Domestic OEMs are adding a side business under the premise of safeguarding their main business; robots serve to sell cars rather than replace car sales. Tesla can afford to let its robot business remain unprofitable in the long term and bet on valuation, but domestic automakers do not have this luxury. The cash burn in their main business is much faster than the cash generation from robots.
The question is what these robots can solve. In 2025, China's humanoid robot shipments reached about 14,400 units, with a market size of approximately CNY 1.55 billion. In the same year, the automotive industry's revenue was CNY 11.18 trillion and profit was CNY 461 billion. The former accounted for 0.014% of revenue and 0.34% of profit. In contrast, from January to August 2026, the automotive industry's profit had dropped to CNY 253.4 billion, a year-on-year decrease of 16%.
An industry with annual shipments of 14,400 units and a market size of about CNY 1.55 billion simply cannot fill the gap within the cash burn time window of automakers. Therefore, to understand why so many OEMs are crowding into "humanoid manufacturing", the real question to answer is not whether robots work, but what exactly automakers are aiming for by rushing in at this point.
01 A Brand New Choice Under the Dilemma
The difficulty in the auto industry is not that selling prices are too low, but that revenue, costs, and asset efficiency are deteriorating at the same time, and this state has persisted to this day.
Data released by the China Passenger Car Association (CPCA) on September 28 shows that from January to August 2026, the automotive industry's revenue was CNY 7.0062 trillion, a year-on-year increase of 2.9%; costs were CNY 6.237 trillion, up 4%; profit was CNY 253.4 billion, down 16% year-on-year; and the sales profit margin was 3.6%. This figure is lower than 4.3% in 2024 and 4.1% in 2025, and also lower than the average level of 5.66% for industrial enterprises above designated size during the same period.
The divergence between production and sales is even more direct. Data released by the China Association of Automobile Manufacturers (CAAM) on October 12 shows that in the first three quarters of 2026, automobile production and sales reached 23.246 million and 23.271 million units respectively, both down 4.5% year-on-year. In September, production and sales reached 2.99 million and 2.956 million units respectively, down 8.7% and 8.4% year-on-year.
The domestic market is hit the hardest. From January to September, domestic auto sales were 15.066 million units, down 22.4% year-on-year. In September alone, sales were 1.904 million units, down 26% year-on-year, of which domestic sales of traditional fuel vehicles were 6.599 million units, down 33.6% year-on-year. Chen Shihua, Deputy Secretary-General of CAAM, stated that the market generally shows a trend of being cold domestically and hot externally, and expects domestic sales to continue to decline in the coming months.
The decline in capacity utilization rate blocks the path to recovery by spreading fixed costs. Data from the National Bureau of Statistics shows that this indicator for the automobile manufacturing industry was 70.3% in the first quarter of 2026 and 70.8% in the second quarter, below the 75% healthy line usually adopted by the manufacturing industry, while in 2021 this value was still 74.7%. Structural mismatch is more troublesome than total volume, because the large-scale idling of ICE vehicle capacity and the shortage of high-quality NEV capacity exist simultaneously.
Exports are the only segment currently still growing at a high speed. From January to September, auto exports reached 8.205 million units, a year-on-year increase of 65.8%. In September alone, exports were 1.053 million units, up 61.4% year-on-year. NEV exports reached 4.006 million units, a year-on-year increase of 1.3 times. CAAM expects full-year exports to exceed 10 million units.
But the problem is that the US resolutely refuses to open its market, while the EU is shifting towards trade barrier policies. Exports are currently facing increasing policy risks. Going global is shifting from complete vehicle exports to local production and supply chain compliance. The growth rate is still considerable, but it no longer naturally equals profit.
Relying on scale to spread costs, relying on payment terms to occupy funds, and relying on exports to find incremental growth—these three paths that previously supported automakers' profits are narrowing at the same time. Enterprises do not have a more direct and effective path to solve current problems by simply selling more cars.
02 The Capability Boundaries of OEMs
Robots are pushed to the forefront at this position. The reason for this operation is actually easy to understand, as it can simultaneously meet three demands.
First, to give the capital market a story independent of the vehicle model cycle; second, to provide a new path for final assembly automation that bypasses dedicated equipment; and finally, to give the highly overlapping auto parts capacity a new order outlet. However, there is a problem: none of the above falls on the current income statement. Valuation re-rating is a one-time event, cost reconstruction is measured in decades, and new supply chain orders are first recorded in the suppliers' financial statements.
Therefore, the reason why so many OEMs are intensively "building" and "showcasing" humanoids is driven by two other forces.
Externally, there is policy heat. Embodied intelligence has been written into the Government Work Report for two consecutive years, and the outline of the 15th Five-Year Plan lists it as a future industry. On June 9, 2026, the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) jointly issued a notice to launch a special action for real-scene practical training, aiming to distill over 100 high-value application scenarios by the end of the year. Complete absence would be interpreted as having no position in physical intelligence. Internally, there is valuation pressure; when the main business is under pressure, a story independent of the vehicle model cycle is needed.
Automakers dare to rush in also because they have ready-made things at hand. Multi-sensor fusion, domain control, end-to-end models, and data closed loops in autonomous driving can be used for the robot's environmental perception and task planning. The motors, power electronics, and precision bearings of NEVs highly overlap with robot joints and dexterous hands in the supply chain.
He Xiaopeng previously stated at an earnings call, "Iron is fully self-developed, and over 85% of its supply chain can overlap with XPeng Motors." Meanwhile, SAIC's Nengzai No. 1 handles cell grabbing and loading on the Buick Zhijing E7 battery mass production line, Chery's Mojia has completed the signing of 1,000 intelligent police robots and the delivery of 100 units, and GAC established Huilun Technology in February 2026 and completed a financing of over CNY 100 million.
But what we need to point out here is that reusability explains the low threshold that causes many OEMs to cross over into this field, but it does not mean that commercialization will inevitably succeed. The reasons are at the technical level—
Autonomous driving deals with planning safe trajectories on open roads, while robots deal with continuously controlling a high-DOF body to interact with unknown objects. The latter requires whole-body dynamics modeling, bipedal balancing, gait planning, dual-arm coordination, and torque control. Cars can rely on lanes, rules, and redundant systems to maintain safety boundaries, whereas robots must maintain balance in unstructured spaces.
In addition, data requirements are not entirely the same.
The data closed loop for autonomous driving relies on large fleets, fixed sensor layouts, and clear traffic rules. Robot learning requires human-in-the-loop demonstration, teleoperation, force and tactile signals, and cross-embodiment transfer. Simulation can reduce some costs, but the sim-to-real gap caused by friction and sensor noise remains.
Earlier this year, when the author had the honor of communicating with an expert in this field from the China Academy of Information and Communications Technology (CAICT), the expert mentioned that robots have much stricter requirements for positioning and operation accuracy than cars. Mobile positioning requires centimeter-level accuracy, and hand-eye coordination requires millimeter-level accuracy. Therefore, the data advantage of OEMs only forms a flywheel after the embodiment, sensors, and task library are unified—which is usually not that easy.
Since the differences in technical detail requirements are obvious, the internal friction caused by cross-industry transition is unavoidable.
OEMs are accustomed to long-cycle platform development, supplier nomination, and large-scale production lines, with highly hierarchical management. In the early stages, robots require more rapid trial and error, co-iteration of algorithms and hardware, and on-site engineering residency. Given this obvious problem, several companies have already chosen to establish independent companies and introduce external capital, indicating that the industry has realized that internal organizations may not be the most suitable fit.
But even with such problems, OEMs are still stumbling and have managed to "build" the "humans".
03 Not Quenching Immediate "Thirst", but Unable to Let Go
Starting from 2025, China's humanoid robot industry has gradually moved from the entertainment field towards the path of true commercialization.
For example, UBTECH's main model Walker S2 delivered over 500 units last year, with order amounts approaching CNY 1.4 billion. After small-batch procurement and trials, some industrial customers believe that when performing tasks like handling and simple tooling, the work efficiency per unit time reaches 60% to 80% of that of human labor.
Due to the very positive market response in 2025, sales of this model increased to 921 units in the first half of 2026, but the total sales amount dropped to the scale of CNY 590 million due to increased production and product optimization. However, the gross profit margin was as high as 66.8%. Nevertheless, even with such results, UBTECH's financial report for the first half of this year still showed a loss of CNY 339 million.
At this stage, we are standing on the eve of the explosion of humanoid robots, but when to step over the threshold still requires solving some technical problems. What is worth being alert to is that, given the more intense competition in the industry currently and various new solutions emerging, the main product prices of some enterprises have already dropped to the CNY 100,000 level. In a cycle of declining gross profit margins, R&D amortization, on-site deployment, and maintenance will turn hardware gross profit into service investment. How enterprises and even the entire industry grasp the input-output ratio has become the key.
There is another layer, which is the consideration from a capital perspective. The image of the humanoid robot industry in the public eye currently possesses characteristics of high-tech/high-technology. Compared to the increasingly "cabbage-priced" main auto business, if an enterprise has such a side business, it can obviously enhance its image and be helpful to the valuation of listed companies. The most typical example in this regard is Tesla.
In the fourth part of the Master Plan in September 2025, Elon Musk shifted the focus from electric vehicles to AI and robots, stating that about 80% of the value comes from Optimus. The Model S and Model X production lines in Fremont were dismantled and transformed into dedicated lines for Optimus, with a planned annual capacity of 1 million units. One of the unlocking conditions for his CNY 1 trillion (Note: context implies USD 1 trillion, but strictly following rule: if not explicitly stated as USD, default to CNY. Wait, the original text says "1万亿美元", so it IS explicitly USD. I will use USD 1 trillion) compensation plan is deploying 1 million units. Although judging from the execution so far, this is likely another typical Musk-style overpromising, the combination of the robot business and the Cybercab business can at least support Tesla's market capitalization in the short to medium term.
Of course, the matter of OEMs building robots actually has a certain obsession in the main direction of car manufacturing—the automation rate improvement demand in the final assembly link of OEMs, due to a large number of complex installation processes, has long been unsatisfied. Humanoid robots can directly use existing human production facilities, which is an obvious intensive improvement, and also the irreplaceable value of self-owned factories as training grounds.
The picture of robots turning 180° and going to work in the factory right after rolling off the production line, although there is still a certain distance with current technology, is hard for any OEM to resist its temptation.
What the money invested by automakers in robots has bought in return needs to be distinguished.
What is bought back is the qualification for entry, not profit. The capabilities accumulated by the auto industry are to scale complex products, and its服务对象 (service objects) are shifting from human mobility to human labor itself. The market magnitude of the latter is far greater than that of automobiles. Whoever first adapts to the cost structure and manufacturing system will have the qualification to participate in the next round of distribution. Valuation is just a byproduct, built on external capital pricing, and has nothing to do with the main business cash flow.
The establishment of necessity does not equal the establishment of urgency. The mistake most easily made by automakers trapped in a 3.6% profit margin is treating long-term options as short-term tourniquets, paying bills that will only be cashed in ten years with tense organizational attention and cash flow. The brilliance of idle chess pieces lies precisely in the fact that they do not occupy the main line. Once the main business is required to make way, the nature changes from an option to a liability.
Returning to the question at the beginning, why exactly this timing. The answer is actually a bit disappointing, because this timing was not chosen by the "robots" themselves, but forced by the realistic situation of the main business of car manufacturing. Neither the real technological turning point nor the real commercial closed loop has appeared this autumn; what has appeared is the dual pressure of the income statement and the valuation window.
What OEMs can do right now is to keep the cost at the level of idle chess pieces, use self-owned factories to exchange for real-machine data, use existing supply chains to spread trial and error costs, use independent companies to isolate organizational conflicts, and then wait to see which of the two accounts turns around first. The main business account relies on product definition, overseas operations, and cash efficiency, and can only be ground out bit by bit in vehicle models, channels, and payment terms. The robot account depends on when a standard operating scenario with 10,000-level repurchases appears.
Before this, the most pragmatic approach is to keep the two accounts separate. Building humans cannot save this round of dilemma, but giving up building humans may mean not even getting the ticket for the next round. Putting these two sentences together is the complete reason why so many enterprises are taking collective action this autumn.