Humanoid robots are about to produce big results!

AI is growing arms and legs. The United States’ advantage lies in its smarter brain, while China has the strongest and most complete body.

Text | Huashang Taolue Huaxin

On October 21, Tesla announced in its third-quarter financial report that the Optimus humanoid robot has started production on the mass production line of the Fremont factory in California. This production line originally belonged to the Model S and Model X. In July this year, Tesla dismantled the old line in just 46 days and replaced it with the Optimus production line.

At almost the same time, Nvidia founder Huang Jenxun shouted at the GTC conference in Berlin: Physical AI and humanoid robots are the next great industry. Wang Xingxing said that the biggest mistake he made was not trusting AI enough, and he advanced the timetable for the large-scale application of humanoid robots. Yushu Technology, whose market value was cut in half, began a stock price counterattack.

【From knowledge to action】

If the large model of the past three years is compared to a person, it is more like a “know-it-all” trapped in a computer: it can create, program, answer all your questions, and chat with you about life, but it never leaves the computer, even to pick up a cup for you. What physics AI needs to make up for is this lesson——

Let “know everything” become “omnipotent”.

Huang Renxun stated at CES in January this year that AI is moving beyond the stage of understanding digital information and is beginning to understand gravity, friction, inertia and cause and effect. Translated into vernacular, physical AI is to transform AI from “being able to speak” to “being able to do”. It will come out of the screen and into factories, warehouses, streets and homes.

The core of Physics AI is robotics.

A car that can drive itself is a robot with wheels; a drone that can fly by itself is a robot with wings; a robotic arm that can find parts and tighten screws on its own on a production line is a robot that grows on the workstation…

Humanoid robots are just the most difficult and the most highly anticipated of them all. The reason is: Stairs, door handles, tools, workstations, the human world are all designed according to the human body.Whoever can look like a human can take over all human tasks more seamlessly and directly.

Physics AI is targeting the largest market, the human labor force itself.

This is a big deal.

It is becoming increasingly difficult to recruit workers in the global manufacturing industry, and the aging of the population has left the manpower gap in elderly care, nursing, and logistics growing larger year by year. Citigroup has estimated that by 2050, the number of humanoid robots worldwide will reach 648 million units, equivalent to more than 40% of today’s global car ownership.

Making a humanoid robot work is like making a person work. It seems simple, but behind it is a series of complex disassembly.

Suppose you say to it: Help me get a bottle of Coke from the refrigerator.

It first needs to use its “eyes” – cameras and 3D vision sensors to see clearly where the refrigerator is, where the door handle is, and which layer of Coke is on it.

Then there’s “brain.” The large multimodal model breaks this sentence into a series of actions: walk over, open the door, find the Coke, pick it up, close the door, and hand it over. In the industry, this type of model is called VLA, which is the “visual-language-action” model.

Next comes the “cerebellum,” the motor control system. Walking on two legs is essentially on the verge of falling with every step. It has to adjust its center of gravity in milliseconds to ensure that the body does not fall.

Finally, “muscles and sinews.” A humanoid robot usually has about forty joints, and each joint is filled with motors, reducers, screws and sensors. You also need to have touch and force sense in your hands: you can use a little force to grab a Coke bottle, you have to use extra force to grab a steel ball, and you have to hold back your strength to grab an egg.

It sounds like every step is not complicated, but for robots, every step is a hurdle.

There is a famous “Moravec Paradox” in the field of artificial intelligence: It is relatively easy for a computer to win the world championship of Go, but it is even more difficult for a robot to fold a piece of clothing like a three-year-old child.

The difficulty lies in the data.

The reason why large language models have advanced by leaps and bounds is because humans have accumulated decades of text on the Internet. But in the physical world, there is no ready-made “Internet”. Every grab and every fall of the robot has to be tried over and over again in the real world, which is slow and expensive.

There are two ways to break the situation. One is to let real machines work in large quantities and collect data while working; the other is to “open a driving school” in the virtual world.

The latter is like a pilot’s advanced simulation cabin: first fall a million times in the virtual world, and then take the first step in the real world. NVIDIA announced a set of data in March last year: using its tools, it can generate 780,000 robot synthetic action trajectories in 11 hours, which is equivalent to the amount of data from 6,500 hours of human demonstration.

Even so, today’s physics AI is still far from “universal”.

Wang Xingxing, founder of Yushu Technology, said at the World Robot Conference in August: In a fixed scene, after sufficient data collection and special training, the success rate of robots completing tasks can be almost 100%; however, as long as an item is changed, or the environment is slightly moved, the success rate will drop significantly.

Wang Xingxing’s standard for the “ChatGPT moment” of embodied intelligence is: take a robot to any unfamiliar environment, such as your home, and it can basically complete about 80% of the tasks. On that day, the industry will truly reach the critical point of explosion. At that time, his expectation for this time was: two to three years if fast, five or even ten years if slow.

But on October 12, at the roundtable dialogue at the opening ceremony of the 26th China International Industrial Expo, Wang Xingxing advanced this prediction. He believes that the rapid advancement of model capabilities will soon break down the barriers. If this critical point is fast, it may arrive within one or two years. He even said that when GPT-7 comes out next year, it may directly solve the problem of embodied intelligence.

in other words,A few months ago, physical AI was still a distant future, but now it’s on the home stretch.; And the manufacturing, logistics and service industries it wants to transform are precisely the home fields that China is most familiar with.

[Heroes fall]

There are many players on the global physics AI poker table, and their playing styles are completely different.

Musk is the most high-profile evangelist and the heaviest bettor.

As early as the Tesla shareholders’ meeting in June 2024, he made a big calculation: the world may build 1 billion humanoid robots every year in the future, and Tesla’s share may be far more than 10%; when the scale is large enough, the cost of each unit can be reduced to about 10,000 US dollars, and it can be sold for 20,000 US dollars.

He called Optimus Tesla’s “biggest product in history” and imagined that with it, everyone could have a top surgeon.

But Musk’s schedule has always been faster than the factory. He had expected that thousands of Optimus would enter the factory in 2025, but in January this year, he changed his mind: Optimus has not yet been put into use in the factory in a substantial way. However, Musk really put his property on the line.

He stopped the Model S and Model X and converted the Fremont production line into the Optimus factory.In the third quarter of this year, Tesla’s R&D expenses jumped 69% year-on-year to US$2.76 billion, and net profit fell 43% year-on-year. The money was mainly invested in new businesses such as super intelligence and robots that Tesla said.

As early as July 1, Musk took a precaution when he posted a photo with the Optimus production team: the initial mass production will be extremely slow because everything is new, “it’s not like building a car.”

On October 21, Optimus finally started production on the mass production line. Musk said that most of the equipment for this production line, which is designed to have an annual production capacity of 1 million units, has arrived; around April next year, Optimus will be replaced with a self-developed AI5 chip. His judgment is still grand:

Optimus will become the largest product on earth, and the number of robots will exceed humans in the future.

Huang Renxun is taking another path: not to build robots, but to make all robots around the world grow on NVIDIA’s platform.

Nvidia’s style of play is a complete set of “family buckets” from training to upper body:The Cosmos world model is responsible for understanding and generating the physical world, the Isaac simulation platform is responsible for “opening a driving school”, GR00T is a general brain prepared for humanoid robots, and Jetson Thor is a chip installed on the robot.

At the GTC conference in March this year, industrial robot giants such as ABB, Fanuc, KUKA, and Yaskawa, as well as Figure from the United States and Zhiyuan from China, all appeared on Nvidia’s partner list.

At the GTC conference in Taipei at the end of May, NVIDIA released Cosmos 3, which is claimed to be the world’s first fully open source, full-modal model that can compress the training and evaluation cycle of physical AI from months to days.

At the GTC conference in Berlin on October 21, Huang Renxun was accompanied on the stage by several small 25-centimeter-tall, duck-like robots. He asked them to do somersaults and “sleep”, but one of them turned its back, so Huang Renxun had to tell it: focus. There was a burst of laughter in the audience.

After teasing the duck, he brought out something serious: a reference humanoid robot based on Jetson Thor and GR00T built by Germany’s Agile Robots. Agile Robots plans to build 1,000 to 2,000 units next year. Huang Renxun’s judgment that day was that the robot’s ChatGPT moment was rapidly approaching.

What Huang Renxun wants to do is to bring the “chip + software + ecosystem” that makes a lot of money in the data center into the physical world: no matter whose robot wins in the end, their brains will have to grow up on NVIDIA’s platform.

Other players in Silicon Valley are betting the most money on “brains.”

At CES in January this year, Google DeepMind announced that it would cooperate with Boston Dynamics to install Gemini Robotics models into a new generation of Atlas humanoid robots. According to Hyundai Motor’s plan, the new Atlas will enter its factory in Georgia, USA this year.

What is quite dramatic is that in 2017, Google sold Boston Dynamics to SoftBank due to the high investment and long return of robots. After nine years of going around in circles, Google came back as a “brain supplier”.

The valuations of startup companies have skyrocketed. Figure AI was valued at US$39 billion in September last year; Physical Intelligence, which was co-founded by former Google DeepMind researchers and Stanford and Berkeley scholars, and only makes general models for robots, was revealed in March this year to be negotiating a new round of financing of about US$1 billion at a valuation of about US$11 billion, doubling its valuation in four months. Co-founder Sergey Levin positions the company as a “ChatGPT for robots.”

On September 28, AMD announced an all-stock acquisition of World Labs founded by Li Feifei for approximately US$8.2 billion. This is AMD’s second-largest acquisition in history after Xilinx. Li Feifei’s World Labs was founded in 2024. It does “spatial intelligence”: generating and simulating interactive 3D worlds from text, pictures, and videos. It is regarded as the upstream of physical AI. Su Zifeng said: To build the next generation AI computing platform, we must have a deep understanding of how models evolve. Li Feifei will serve as AMD’s executive vice president and chief scientist. This deal also changed the competition among chip giants from selling cards to grabbing “world models”.

Nvidia open sourced Cosmos and allowed the entire industry to do physical AI on its platform; AMD simply bought a top laboratory home and used models to define chips.

However, Silicon Valley has the smartest brains and the most expensive valuations, but the real body capable of mass production is not here, but in China. China’s robots may not be smart enough and have limited capital, but they are far ahead in mass production.

The basic pattern of China’s physical AI with humanoid robots as its core is: Ontology is running the show, car companies are turning around, and giants are setting up the stage.

Let’s look at the new forces of Ontology first.

According to statistics from market research firm Smart Analytics Global,In the first half of this year, approximately 19,100 humanoid robots were shipped worldwide, with Chinese manufacturers accounting for more than 97%.Among them, Zhiyuan shipped about 8,400 units, with a global share of 44%, ranking first; Yushu shipped about 5,900 units, with a share of 31%. Together, the two companies account for three-quarters of the world.

The founding team of Zhiyuan is quite interesting: Chairman Deng Taihua is the former vice president of Huawei; President Peng Zhihui is the “Zhihui Jun” on Station B, who joined Huawei as a “genius boy” that year. Two Huawei people took the company from 300,000 yuan in revenue in the first year to more than 1 billion yuan in 2025 in three years.

On June 28, Chiyuan’s 15,000th G2 robot rolled off the production line, less than three months after the 10,000th one. This robot was not in the exhibition hall, but was sent directly to the tablet production line of Longqi Technology’s Nanchang factory to undergo quality inspection together with eight other G2s. In the previous six days of live broadcast, these eight robots completed more than 17,500 quality inspections, with a success rate of 99.99%.

▲Elf G2 live broadcast at Longqi Technology production line

Yushu has a different temperament. Wang Xingxing started his business with four-legged robots in 2016 and has been doing it for ten years.

At the end of May this year, a reporter from “Daily Economic News” walked into the Yushu headquarters in Binjiang, Hangzhou: at the door, R&D staff were debugging a four-legged robot; on the roof, a wheeled robot dog was patrolling unattended; the first-floor exhibition hall had been vacated, and the sales team was moving into a new building that had not yet been renovated. The company grows faster than the house.

Smart cars and humanoid robots are highly homologous at the bottom level in fields such as sensors, chips, motors, batteries, and algorithms. According to estimates from the Gasgoo Automotive Research Institute, the supply chain overlap between the two reaches 60 to 70%. A slight extension of China’s new energy automobile industry chain is the robot industry chain. In fact, Chinese car companies are also actively deploying robots.

Among them, Xiaopeng is the one who has turned around most actively and thoroughly. He Xiaopeng said in an internal letter on June 10 that this is a key step in Xiaopeng’s transformation from a “smart car company” to a “physical AI company.” In August, Xpeng’s robotics business completed a first-round financing of over US$900 million, with a post-investment valuation of over US$6.3 billion. Tencent and Alibaba entered the game as strategic investors. In addition to Xiaopeng, Ideal’s Li Xiang announced in March that 2026 is a critical year for Ideal to evolve into an “embodied intelligence” enterprise.

BYD, which has surpassed Tesla in new energy vehicles, still focuses on manufacturing when it comes to understanding robots. Its executive vice president Li Ke said:The competition among robots is based on manufacturing capabilities, software capabilities and hardware capabilities, and the AI ​​capabilities accumulated in cars can be transplanted to robots.

Chinese technology giants are also increasing their deployment of robots.

Huawei Cloud’s embodied intelligence platform CloudRobo was released in June last year, upgraded and launched into public beta in June this year. It provides a complete pipeline from data, models, simulations to deployment, and has cooperated with more than 30 industry partners and more than 20 ontology manufacturers. Xiaomi’s Lei Jun defines 2026 as “the first year of large-scale embodied intelligence models.”

But China’s ability to run fast in humanoid robots relies not just on one or two star companies, but on the industrial chain and ecology. According to the Ministry of Industry and Information Technology’s disclosure in July, there are currently more than 400 complete models of humanoid robots in China, accounting for more than half of the world’s total; the global sales share of quadruped robots is close to 70%. Motors, reducers, batteries, sensors, structural parts, almost every link can be found with suppliers within an hour or two of driving in the Yangtze River Delta and Pearl River Delta.

In the Robot Valley in Nanshan, Shenzhen, about 10 kilometers of Liuxian Avenue is a complete system for the research and development of key robotic technologies and parts manufacturing. From algorithm software such as large and small brain models, perception systems such as tactile sensors and lidar, to execution systems such as servo motors and power modules, to core components such as robotic arms and dexterous hands, you can find high-quality suppliers in the Robot Valley to achieve rapid delivery.

Here, “publishing pictures in the morning, proofing in the afternoon, and iterating the next day” is the norm. A tactile sensor for dexterous hands used to cost 100,000 yuan to import, but now domestically produced products start at 199 yuan. From January to August this year, robot freight orders in Nanshan District increased by 94.4% year-on-year.

The United States is telling the grandest stories and building the smartest brains; China is building the most bodies, running the fastest mass production, and entering the most realistic scenes.When China and the United States join forces, the entire humanoid robot will be able to run faster, and AI will be able to get out of the screen faster and enter all walks of life and every household.

[The Way to Nuggets]

When the industry becomes popular, money will come.

According to statistics from industry organizations, in the first half of this year, China’s total humanoid robot production exceeded 40,000 units, with a month-on-month growth of 50% in the second quarter; total track financing reached 104.1 billion yuan, which has exceeded the entire year of 2025. The Ministry of Industry and Information Technology predicts that the output of complete machines this year is expected to exceed 100,000 units. Citi even predicts,By 2050, the global humanoid robot market will reach US$7 trillion.

The demand is also changing from “performance” to “job placement”. The Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission of the State Council jointly issued a notice in June to launch a special action for real-life training of humanoid robots and embodied intelligence in 2026. It requires that by the end of 2026, key products such as humanoid robots will be the first to complete application verification and normal deployment in a number of representative scenarios, and start the “operation mode.” The “Shanghai Action to Accelerate the Development of New Generation of Intelligent Manufacturing” released by Shanghai on October 12 made it clear that 8,000 humanoid robots will be promoted into factories.

Data from IDC shows that in the first half of this year, global shipments of humanoid robots were approximately 25,000 units, a year-on-year increase of 432%. Among them, the proportion of “non-working” scenes such as scientific research, education, and exhibitions has dropped to 69% from 83.8% last year. This means that the robots are moving from the booth to the workstation one by one.

But the more exciting it gets, the more we need to ask calmly: Whose pocket will the money from physics AI end up in?

Looking back at the history of new energy vehicles, there is a rule worth remembering:When the entire vehicle is in trouble, the first thing to make money is often the core components. The Ningde era is the best footnote. There is a high probability that physics AI will repeat this scene.

The overall machine structure is far from finalized, and today’s champion may not be on the poker table tomorrow; but no matter whose robot wins, joints, motors, reducers, screws and sensors cannot be avoided.

The industries and investment opportunities that physics AI will generate in the future focus on three lines:

First line:Let’s look at “body” first, then “brain”.

Taking A-share companies as an example, most of the physical AI companies that have been verified by orders today are concentrated in the actuator link. The reason is very practical:The value of the brain is currently mainly taken away by Nvidia and unlisted model companies; China holds the strongest manufacturing base in the world, so China’s opportunities are first hidden in the body of physical AI.

The green harmonic is representative of this line. Harmonic reducers have been monopolized by Japanese companies for a long time. Green Harmonic was the first to break the gap. Now it has achieved large-scale mass production of planetary roller screws and fully penetrated into the core supply chain of embodied intelligence.

Sanhua Intelligent Control and Tuopu Group represent another type of player: the auto parts giant has entered into robot actuators with large-scale manufacturing and cost reduction capabilities, and is regarded by the market as a core player in Tesla’s supply chain.

In addition, the precision gears and reducers of Double Ring Transmission, the screws of Hengli Hydraulics, the coreless motors of MOONS Electric, the 3D vision of OBI Zhongguang, and the force sensors of Keli Sensing all have strong skills in their respective subdivisions.

Second line:Let’s see if the main business can support it.

Robot volume expansion is a marathon, and there will be bumps along the way. Companies that can survive to the end are often companies with sufficient cash flow from their main business.

Inovance Technology is a typical example. Revenue in the first half of this year was 24.675 billion yuan, and the domestic market share of general servo systems rose to 35.2%, ranking first. Humanoid robots are not a transformation for it, but a natural extension of the capabilities of servo, motor, and controller. This type of company may not have the greatest flexibility, but they are least afraid of subject matter fading.

Third line:Look at the order, not the concept.

Lesai Intelligence is an example of using financial reports to speak for itself. It focuses on the two major categories of joint modules and dexterous hands. In the first half of the year, its revenue was 1.247 billion yuan, a year-on-year increase of 39.92%; its net profit attributable to the parent company was 194 million yuan, a year-on-year increase of 62.79%. Its products have received bulk orders from domestic mainstream robot manufacturers.

At the machine level, Yushu Technology is already the purest humanoid robot target in the A-share market and the valuation anchor of the entire sector; Zhiyuan has its own capital platform in the A-share market through its acquisition of Shangwei New Materials.

In turn, there are three types of companies that require extra caution.

One type is those who are enthusiastic about “cooperation”, “shareholding” and “strategic agreement”. On June 8, home furnishing store Juran Smart Home was hyped as a physical AI concept due to its cooperation with UBTECH, and its stock price hit the limit. That night, the company issued an announcement to clarify: it was just a framework agreement, with no actual orders, and it had not contributed a penny to revenue.

One category is where valuation is far ahead of performance. As of early July, Green Harmony’s stock price has risen by more than 150% this year, but its revenue in the first quarter was only 140 million yuan. The market has bought several years of growth in advance.

One type is to create blueprints and provide concepts, but they have no diamonds but have to do porcelain work. For example, a company with neither strong capital nor top technical talents claims to be the brain of the machine and even competes with NVIDIA.

A further test lies in price. Once robots increase in volume, the unit price of components will inevitably be continuously depressed, just like photovoltaics and power batteries back then.The final winner is either the one with unique core technical barriers, or the one who can minimize costs and maximize scale.

In the final analysis, when investing in physical AI, you don’t have to bet on who can build the best robot, but rather find something that no one can live without.

This change is coming faster than most people expected. Wang Xingxing’s self-reflection at this year’s Industry Expo is worth remembering by everyone who is staring at this track:

“The biggest mistake I’ve made in the past decade or so is that I don’t believe enough in AI.”

——END——

Welcome to follow [Chinese Business Tao Lue], get to know the famous figures, and read the legend of Tao Lue.

All rights reserved. Private reproduction is prohibited.