Billions are pouring into embodied-AI, but slow robots, circular transactions and uncertain demand are raising questions about whether valuations can last.
By Li Zinan, Shen Yuan, Xu Yumeng
At a simulated auto factory earlier this year, a humanoid robot arranged four fingers into a flat surface, lifted a box and walked 20 meters to a shelf. The whole process took about 90 seconds — roughly 30% as efficient as a human.
Last month, at a bearing factory, another robot picked up a bearing from a pallet and placed it in a plastic box. A human could complete the movement in two seconds. The robot took 70 seconds.
The company that built the first robot is valued at nearly 40 billion yuan ($5.95 billion). The second, founded less than a year ago and already through five funding rounds, is worth more than $1 billion.
According to Morgan Stanley, more than 10,000 embodied robots were shipped in China last year; Unitree Robotics(688836.SH) sold 5,215. After its August IPO, Unitree’s market cap briefly exceeded 400 billion yuan. The combined market cap of four new-energy vehicle makers — Li Auto, Nio, Xpeng and Leapmotor — is about 300 billion yuan. Together they sold 1.75 million cars.
That gap between what embodied AI can do today and what investors believe it will eventually be able to do lies at the heart of China’s humanoid-robot frenzy.
Money is pouring into robot makers, dexterous hands, sensors, data collection and components. Valuations are soaring even when products and revenue have barely changed. Orders are proliferating, but some are tied to investors, suppliers, government-backed projects or other participants in the same ecosystem.
The technology may eventually justify the optimism. For now, the industry is still searching for evidence that the money flowing through it represents genuine demand.
If you can’t price it, find the next round
Investment has followed the same trajectory. More than 90 billion yuan flowed into China’s embodied-AI sector in the first half of 2026, according to IT Juzi, with 22 companies valued at more than 10 billion yuan. At least six robot makers had already crossed 20 billion yuan valuations, while three dexterous-hand companies had also reached that level.
One of those companies was founded little more than a year ago. The other two were valued at less than 1 billion yuan in early 2025. Their valuations increased 20-fold in about 20 months.
Ask investors how they calculate those numbers, however, and there is no clear answer.
Venture capital once prided itself on finding opportunities outside the consensus. In embodied AI, the consensus has already formed. Investors are increasingly concerned with getting into the leading deals and making sure the right institutions are investing alongside them, because those investors may help bring the next funding round.
Founders now act like celebrities, only opening allocation when they meet a fund’s senior partner. Over private dinners, executives trade tips: when 10 investors add you, leave them hanging, then pick one or two. Create mystery. The more opaque, the more desirable.
Orders are growing. Sustainable demand is not
The industry’s revenue story is more complicated.
Unitree’s prospectus offered the market its first detailed look at the income structure of a leading general-purpose robot maker. It reported 1.7 billion yuan in revenue in 2025 and sold 5,215 humanoid robots, which generated nearly half its revenue.
But more than 70% of its revenue as of September last year came from research and education. Less than 10% came from industrial applications, and only 15.7 million yuan could be clearly attributed to smart manufacturing, inspection and logistics.
Its customers were also fragmented: its five largest customers accounted for just 10.6% of revenue, with JD.com, its biggest customer, contributing 3.54%. There were still no industrial customers buying its robots on a large scale.
Li Yuanqing, co-CEO of Lexiang Tech, said: no matter what robot you build, you will sell at least 500 units, because 500 competitors will buy one to tear apart. But Unitree’s case is hard to replicate. It has a decade of history, a consumer quadruped line, the Gala halo and a 99,000-yuan price advantage. Only two or three peers sold more than a thousand units.
For many other companies, the industry’s apparent demand is being generated by a much more complicated ecosystem. One example is robot data-collection factories. Operators buy robots and have workers control them while performing tasks such as folding clothes, wiping tables and sorting objects. The resulting visual and movement data is then sold to robot companies.
Some robot makers sell robots to these factories and then promise to buy data from them, effectively returning part of the original purchase price.
Where are the buyers?
The economics look attractive on paper. Real-robot data can sell for more than 700 yuan an hour, while a 100-robot facility could theoretically produce about 10,000 hours of data a month. At a robot price of 500,000 yuan, a fully utilized facility could recover its robot investment in eight months before labor, land and maintenance costs.
In practice, some factories have robots sitting idle and data for which they cannot find buyers. One facility that had spent about 80 million yuan on robots eventually sold some to schools after its expected payback period stretched beyond expectations.
The problem is that the industry has not yet agreed on what data is actually valuable. Different robot designs, sensors and camera positions mean that data collected for one machine often requires substantial conversion before it can be used by another.
Local governments and state-owned enterprises have also become important participants, sometimes acting simultaneously as customers, investors and operating partners.
In one case, four of a robot company’s five largest customers were local governments or state-linked entities, together accounting for more than 30% of revenue. Its largest customer was also a shareholder, and the two jointly established a robot training center.
Manufacturing companies have their own reasons for participating. Suppliers to the auto and battery industries have been eager to present robotics as their next source of growth. Some buy robots, invest in startups, and announce supply agreements.
When one battery-materials supplier announced that it had bought a robot from AgiBot and would supply its core joint modules, its share price jumped 20% in a day.
These relationships can create genuine business opportunities. But they can also blur the distinction between a robot being sold because a customer needs it and one being bought because the buyer has a financial, strategic or investment relationship with the manufacturer.
As one investor put it, the sustainability of robot revenue ultimately depends on four factors: how many units are sold to end customers with no equity or supply-chain relationship, who bears the inventory risk, whether customers actually accept the machines and whether they buy again.
All of those questions eventually come back to one thing: whether the robots work.
Billions raised but ‘brains’ are immature
The industry’s technical progress is real but the “brain” remains at an early stage.
One investment banker recalled visiting a company valued at more than 20 billion yuan in May. Engineers asked its robot to fold a towel. The demonstration lasted 15 minutes and the towel never got folded.
In another demonstration, the robot lifted a specially designed box using slots of fixed dimensions. The investor moved the box slightly and asked whether the robot could still handle it. It could not.
Some leading companies are investing heavily. Ubtech Robotics (9880.HK) spent about 500 million yuan on research and development last year, while AgiBot also spent more than 500 million yuan, according to investors.
But Unitree spent just 145 million yuan. Another startup valued at more than 20 billion yuan after raising about 5 billion yuan had spent less than 40 million yuan on R&D.
One investor joked that if the company put all its funding into wealth-management products, the returns could cover its operating costs, turning it into a “perpetual company.”
The industry has a reason for being cautious. Its technical route has not yet converged, making large, irreversible investments risky. But that also creates a peculiar situation: companies continue to raise money that they cannot necessarily turn quickly into technological advantages.
Increasingly, money is being directed toward computing power and data. But even there, the scale remains modest. The industry leader in first-person data had produced only tens of thousands of hours in the first half of the year. The largest orders received by data companies were worth only several million yuan.
Nobody has yet established which type of data is most useful. The industry is experimenting with first-person data collected from people wearing cameras and sensors, while the relative value of this approach versus real-robot data remains uncertain.
Unlike autonomous driving or large language models, embodied AI has no widely accepted benchmark or milestone that shows how quickly a company is progressing.
Computing power, data and talent can all be bought. Nobody yet knows how reliably those inputs can be converted into a robot that can generalize to unfamiliar environments.
Unitree becomes a valuation anchor
This is why Unitree’s stock-market debut matters beyond the company itself.
The sector’s initial enthusiasm was fueled by Tesla’s Optimus project and Elon Musk’s prediction that humanoid robots could eventually become a huge global industry, with as many as 10 billion units in operation.
But Tesla’s own progress has forced some investors to reconsider. Production targets and timelines for Optimus have repeatedly shifted, and some suppliers that spent years positioning themselves for the robot market are now questioning whether the opportunity will arrive as quickly as expected.
Primary-market enthusiasm in China has not dimmed. Investors treat a robotics start-up as an “industrial option”: if general-purpose robotics breaks through, today’s outlay could yield enormous returns. What sustains that bet is the prospect of a public listing.
Unitree’s IPO has now provided something the private market badly needed — a public valuation for a company with real production, real revenue and a proven ability to manufacture robots at scale.
But Unitree is not representative of most of the sector. It has proven it can mass-produce, generate revenue and maintain margins. Its edge lies in hardware, low-level control and manufacturing — know-how accumulated over more than a decade.
The “brain” is another matter. Unitree is developing embodied models and plans to spend over 2 billion yuan on foundational technology. But these capabilities have not been through commercial validation. Additional model and compute spend would raise costs and could compress margins.
For now, Unitree looks more like a highly capable smart-hardware manufacturer than a proven embodied-AI model company. That distinction could determine the industry’s next phase.
If public markets are willing to pay Unitree a substantial premium for AI capabilities it has yet to prove, private companies will have room to justify their extraordinary valuations on the promise of future breakthroughs.
If investors instead value Unitree primarily on its existing hardware revenue and profits, the many startups that lack revenue, profits and mass production will have a much harder question to answer:
Why are they worth so much?
Source:
LatePost
