Is there really no entry barrier in the humanoid robot industry?

The current boom is the result of accelerated maturity driven by capital, and the release of premium on technological concepts.

The "Diverse Faces of Robots" at the Auto Show

At the Beijing Auto Show in April 2026, the humanoid robot exhibition area was even more bustling than the new energy vehicle booths.

On BYD's booth, a 1.75-meter-tall silver robot was handing out mineral water to visitors. Its movements were not particularly smooth, but it stood out for its stability. Three meters away, Honor's robot was performing Tai Chi, executing every move precisely, attracting many mobile phone users to stop and take photos. Moving further forward, Unitree Robotics' G1 robot was doing somersaults, slightly bending its knees when landing to achieve perfect cushioning, drawing a burst of exclamations from the crowd. The Expedition A3 from Agibot was dressed in work overalls, tightening screws on a simulated automobile production line, with a sign standing next to it that read: "Adapted for BYD production lines".

A middle-aged investor in a suit stood in the center of the exhibition area, frowning and saying to his assistant: "Look, companies that make mobile phones can do this, companies that make cars can do this, and companies that make robots can do this too. What exactly is the threshold for this industry?"

This question is probably the most reluctant but unavoidable inquiry that all participants in the current humanoid robot track have to face.

If Honor can do it, and BYD can do it, then what exactly is the core competitiveness of these "professional players" such as Unitree Robotics, Ubtech, and Agibot? If it is only integration and assembly, where is the moat? If the embodied intelligence technology is not yet mature, can we infer in reverse that the current robot companies do not actually have such a high threshold?

The Temptation and Illusion of "Assembly"

To answer this question, we must first debunk an industry illusion: humanoid robots do seem easy to make.

A bipedal walking platform, two sets of robotic arms, plus several sensors and an industrial computer, and then a few motion control engineers are recruited to adjust the algorithm, and a "humanoid robot" that can walk and jump is born. After the explosive popularity of ChatGPT in 2023, large models have installed a "brain" for robots, making voice interaction and task planning seem within reach overnight. As a result, we have witnessed an unprecedented "human-making movement" — According to incomplete statistics, nearly 370 embodied intelligence startups have been established in the past two years, nearly 50 companies have entered the listing process on the Hong Kong Stock Exchange or A-share market, the financing of China's humanoid robot industry exceeded 200 billion yuan in 2026, and the valuation of seven leading enterprises has exceeded 10 billion yuan.

But behind the bustle, an awkward fact is: most players in this industry are essentially "advanced assembly plants".

Mechanical joints can be purchased, reducers can be purchased, torque motors can be purchased, vision sensors can be purchased, and even motion control algorithms have open-source solutions. In the era of large models, the "brain" can also call cloud APIs. Putting all these things together, and then putting on a cool carbon fiber shell, a humanoid robot priced at hundreds of thousands of yuan leaves the factory. This model is not essentially different from the logic of assembling smartphones in Shenzhen's Huaqiangbei ten years ago.

Wang Guangxi, Managing Director of Lenovo Capital, said bluntly at the Lenovo Capital Week in July: In the long run, robot body manufacturing may not be suitable for technology-based entrepreneurial teams. Those with strong large-scale manufacturing capabilities are original equipment manufacturers (OEMs), supply chains and manufacturing enterprises, and these industrial chain companies may also take the robot body as a business increment.

In other words, when "making robots" is still in the stage of "assembling parts", technology-based teams do not have comparative advantages. Honor has the supply chain management experience of consumer electronics to make robots; BYD has the precision technology of automobile manufacturing and production line scenarios to make robots. What they lack is not the ability to "make it", but the ability to "make it well, sell it, and iterate it". And it is precisely the latter three things that constitute the real threshold.

The problem is that the smartphone assembly plants eventually died in the red sea of cost performance, and only the players who mastered chips, operating systems and brand channels survived. The humanoid robot track is repeating this story.

The "Fourth Transformation Sorrow" of Cheetah Mobile

To clarify this threshold, we have to mention a "pioneer" — Cheetah Mobile.

In 2016, when most people were still discussing AlphaGo, Fu Sheng had already led OrionStar into the service robot track. At that time, Cheetah Mobile's market value was at a high of 5 billion US dollars, and Fu Sheng's ambition was to "redefine robots with AI". OrionStar launched shopping mall guide robots and hotel reception robots, which were once visible in Wanda Plazas and chain hotels in many cities.

But nine years later, Cheetah Mobile's market value has shrunk from its peak of 5 billion US dollars to less than 100 million US dollars. OrionStar's robot business has suffered large losses for a long time, and eventually reduced to "other business" in the financial report. During the 2026 World Cup, Fu Sheng came out with the AI digital employee "Lobster 30,000" to predict the matches, trying to ride the wave of AI — but this is more like a desperate marketing struggle, rather than a strategic comeback.

What did Cheetah do wrong?

It's not that the direction is wrong, but the path is wrong. Cheetah's robots, with core sensors and motion chassis all purchased externally, cannot form a closed loop of software and hardware. What it does is "integration", not "R&D"; what it sells is "concept", not "product". The shopping mall guide scenario is highly scattered, small and medium-sized merchants have very low willingness to pay, the volume of large customer orders is limited, and the repurchase rate is dismal. More fatally, Cheetah has always been unable to give up the shrinking basic disk of tool business, and cannot fully invest in the AI track. Pulled by two sides, neither of the two business lines can establish competitive advantages.

Cheetah's story is a cruel fable: in the robot track, there is a Mariana Trench between "can do" and "can survive".

Ubtech's "Valuation Dilemma"

If Cheetah is an extreme case of "pioneer becoming martyr", then Ubtech represents the awkward situation of "professional players".

At the end of 2023, Ubtech was listed on the Hong Kong Stock Exchange, becoming the "first share of humanoid robots". But the capital market did not give it the expected premium. As of early June 2026, Ubtech's share price is about HK$110, with a total market value of about 51 billion yuan. For a robot company that claims to master full-stack technology, this valuation is not high.

Why?

First of all, the revenue structure is single and in a transition period. Although the revenue proportion of humanoid robots has reached 41.1%, Ubtech is highly dependent on the Walker series. Once the delivery fails to meet expectations, the overall revenue will face the risk of cliff-like decline. Secondly, there is uncertainty in the technical route and competition pattern. Tesla Optimus, Agibot and Unitree Robotics are all accelerating penetration, and the customer concentration and substitution risk in industrial scenarios have not been fully released.

More importantly, Ubtech faces a fundamental question: is your technical moat really deep enough?

The Walker series can indeed walk and jump, and can "intern" in factories, but how far is it from the real "embodied intelligence"? When car companies and mobile phone manufacturers enter the market one after another, will Ubtech's full-stack self-research advantage be quickly diluted by their supply chain scale and scenario resources? The low valuation given by the capital market is essentially the pricing of this uncertainty.

And this caution is not only aimed at Ubtech. Wang Guangxi judges that there is an obvious bubble in the current embodied intelligence financing market — "A company has three rounds of financing in one month, and its valuation changes three times, which is definitely not right from common sense". He expects that from the second half of this year to next year, with a number of companies going public intensively and their financial data forced to be open and transparent, the secondary market will start a brutal "interrogation"; at the same time, some leading companies will begin to show real large-scale revenue and landing cases, while players who only have Demo (demonstration) capabilities will face unprecedented survival challenges.

Ubtech's market value of 51 billion yuan may be the prelude to this "de-bubbling" reshuffle — the capital market is voting with real money to distinguish between "companies with products" and "companies with only concepts" in advance.

Unitree and Agibot: Why Can They Enjoy High Premiums?

In sharp contrast to Ubtech's "valuation dilemma" are the high valuation expectations of Unitree Robotics and Agibot.

Unitree Robotics is about to be listed on the Sci-Tech Innovation Board, taking only 73 days from application to approval, setting the fastest record since the pre-review mechanism of the Sci-Tech Innovation Board was implemented. The company's revenue in 2025 was 1.699 billion yuan, its net profit was 278 million yuan, and its main business gross profit margin was as high as 60.13%. The total shipment of humanoid robots in 2025 exceeded 5,500 units, making it one of the few humanoid robot complete machine enterprises in the world that have achieved large-scale profitability.

Agibot is even more radical. In July 2026, Agibot officially announced the launch of the listing process in Hong Kong, with a target valuation of 200 billion US dollars (about 1.35 trillion yuan), and its total revenue in 2026 is expected to reach 4 billion yuan.

Why?

The answer lies in two key words: mass production capability and data closed loop.

Unitree Robotics's real barrier is not that it can twist Yangko at the Spring Festival Gala — although that did bring it huge brand exposure — but that it reduced the price of bipedal humanoid robots from the then common hundreds of thousands of yuan to less than 100,000 yuan, and achieved large-scale shipments. In 2025, the total sales of Unitree's quadruped robots have exceeded 30,000 units, and the shipment of humanoid robots also exceeded 5,500 units. What does this mean? It means that Unitree has run through the complete chain from R&D to manufacturing to sales, and established cost advantages and supply chain discourse power.

But more important than mass production capability is the data closed loop.

Peng Zhihui, "Ji Hui Jun", the founder of Agibot, has a famous judgment: the final competition of humanoid robots is not a competition of hardware, but a competition of "brain". The "Three Intelligence Integration" architecture proposed by Agibot — integrating the large model "brain", the motion control "cerebellum" and the physical "body" — is essentially building a data flywheel. The more robots are sold, the more scenario data is accumulated, the faster the model iterates, the better the product experience, and the more robots are sold. Once this closed loop is formed, latecomers will find it difficult to catch up.

Wang Guangxi's observation confirms this trend. He pointed out that leading enterprises including Unitree are all working hard to make up for the full-stack intelligent capabilities. If an enterprise only makes the robot body and does not have the comprehensive capabilities of robot intelligent R&D and commercial delivery, it may face greater challenges than imagined in the middle and late stage of the track. In other words, the future winner will not be "the best machinery factory", but "the robot company that understands intelligence best".

Unitree also knows this well. In 2026, Nvidia announced at the GTC Conference that it selected Unitree Robotics as its first robot system partner for researchers, and the two sides cooperated to launch a new generation of humanoid robots. This is not only a brand endorsement, but also means that Unitree will enter the vision of the world's top AI research institutions, and obtain the most cutting-edge algorithm feedback and massive scientific research data.

In contrast, although car companies and mobile phone manufacturers have money, scenarios and supply chains, what they lack is robot genes and data accumulation. BYD can make a robot that tightens screws on the production line, but it is difficult for it to build an embodied intelligent model covering thousands of robots and millions of hours of operation data within three years. This cannot be bought with money, it is accumulated over time.

Scientific Research and Education: A Sweet Trap

However, high valuation does not mean that everything is worry-free. There is also a hidden trap in this industry — the dependence on the scientific research and education market.

Unitree Robotics' financial report for the first three quarters of 2025 shows that scientific research and education customers contributed 73.6% of the revenue of humanoid robots, and industrial applications accounted for only 9%. Although leading enterprises such as Agibot and Leju have not released detailed data, practitioners generally reflect that orders in industrial scenarios are not yet large-scale.

What does this mean? It means that most of the robots sold now are not going to "work", but to "be studied". Universities buy one back for algorithm verification, scientific research institutes buy one back to publish papers, and technology companies buy one back to build a Demo. This market has stable demand and timely payment, which is the most comfortable "basic disk" at present — but it is a double-edged sword.

The ceiling of the scientific research and education market is visible. There are only so many top universities and research institutions in the world, and each buys several units, the market will soon be saturated. What really determines the ceiling of the humanoid robot industry is the trillion-level scenarios such as industrial manufacturing, commercial services and household consumption. To enter these scenarios, what is needed is not Demo capability, but reliability, cost-effectiveness and scenario adaptability.

In other words, current robot companies sell "possibilities"; future robot companies sell "certainty". From "possible" to "certain", there are countless production line debugging, countless troubleshooting, and countless customer education. This process is the real threshold.

An Industry Without Thresholds Will Eventually Have Thresholds

Back to the investor's question at the beginning of the article: Is there a threshold in the humanoid robot industry?

My answer is: There is none now, but there will be in the future.

The current boom is the result of accelerated maturity driven by capital, and the release of premium on technological concepts. Honor can do it, BYD can do it, and any company with hardware assembly capabilities can do it. Because the "making" at this stage is just assembling ready-made parts into a movable shell, and then putting on a futuristic shell. This threshold is indeed not high.

But the final threshold of the industry is being raised at a speed visible to the naked eye.

The first threshold is mass production cost. When Unitree lowers the price to less than 100,000 yuan, and when Agibot's cumulative mass production crosses the 10,000 unit mark, latecomers who want to cut in by "assembly" have lost the price space. The large-scale manufacturing of hardware is a moat in itself.

The second threshold is data closed loop. Every robot running in real scenarios is contributing data to the company's embodied intelligent model. Once this data flywheel starts to rotate, the models of first movers will become smarter and smarter, and the catching-up cost of late entrants will rise exponentially. This is exactly the same logic as autonomous driving.

The third threshold is scenario binding. Industrial production lines are not test sites, and customers will not pay just because your robot can do somersaults. Only robots that have actually run in factories, served in shopping malls, and accompanied in families can accumulate irreplaceable scenario know-how. BYD has automobile production lines, but Unitree has the experience of falling on hundreds of different terrains — the latter is the real "driver's license" for robots.

From the end of 2026 to the beginning of 2027, it may be the capital inspection period for the industry to screen out the fittest. When the capital dividend brought by IPO is exhausted, when the orders in the scientific research and education market peak, and when industrial customers begin to evaluate products with ROI (return on investment) instead of Demo effect, those enterprises that only have "assembly" capabilities and no "closed loop" capabilities will, like Cheetah Mobile in those years, become the "transformation sorrow" on the track.

Epilogue: The Root System of the Tech Tree

Unitree Robotics originally wanted to be named "Tech Tree", but because the three characters could not be registered for trademark and domain name, it was abbreviated as "Unitree". The company's mission statement reads: "Create the world's tech tree".

This is a very ambitious name. But the growth law