Source: Late Post. Reported by Li Zinan, Shen Yuan, and Xu Yumeng.
The question that hangs over China’s embodied AI boom is blunt: how many robots working on a production line can replace one BYD worker? The report says there is no clear answer yet.
At the start of the year, in a simulated auto factory, a humanoid robot arranged four fingers into a flat surface, lifted a box, turned, walked 20 meters to a shelf, and put the item down. The whole sequence took about 90 seconds, with efficiency estimated at roughly 30% of a human worker. Last month, at a bearing plant, another robot used a claw to pick up a bearing from a tray and place it into a plastic box to its left. A human can finish the same move in 2 seconds. The robot took 70 seconds. Even if it were 10 times faster, the report argues, it still would not make it into the factory.
The companies behind those two machines have been valued at sharply higher levels than those performance numbers suggest. One is worth close to RMB 40 billion. The other is valued at more than $1 billion, is less than a year old, and has already completed five funding rounds.
Valuations climbed first, while a pricing method never really formed
The report says the most advanced working robot may still be Tesla’s. To lift heavy objects, the machine uses 28 high-precision planetary roller screws across its body, each costing close to RMB 1,000. A full-size general-purpose humanoid robot sells for about RMB 200,000, roughly equal to 40 months of wages for a BYD worker, with depreciation in less than 10 months. A pair of high-degree-of-freedom dexterous hands costs more than RMB 150,000, and if used 8 hours a day, they may need repairs after 10 days.
That cost structure makes adoption in Chinese factories hard to justify for now, especially in a labor market where a RMB 5,000 monthly salary can still recruit workers. Yet the top three dexterous-hand companies have each reached valuations of RMB 20 billion.
According to a Morgan Stanley report cited in the article, more than 10,000 embodied robots were actually shipped in China last year, and Unitree sold 5,215 of them. After its Aug. 19 listing, Unitree’s market value briefly exceeded RMB 400 billion. For comparison, Li Auto, NIO, XPeng, and Leapmotor together are worth about RMB 300 billion, while they sold 1.75 million vehicles last year.
IT Juzi data showed domestic embodied AI financing topped RMB 90 billion in the first half of 2026, with 22 companies valued above RMB 10 billion. Capital has spread from robot bodies into dexterous hands, data, joint modules, and tactile sensors. Some companies less than three years old are already valued above RMB 20 billion. In at least one case, a company had lined up a second funding round before it was formally registered.
Founders and investors are watching Unitree because the company’s listing may serve as a public-market measuring stick. The report’s point is not that Unitree solves the sector’s valuation puzzle. It does not. But with technology, revenue, and profit still falling short of what current private valuations imply, Unitree has become the nearest available benchmark.
At the end of June, AgiBot Variable Robot announced a new round at a valuation of roughly RMB 20 billion. Around the same time, Zhifang also crossed that line. Both called themselves the first embodied AI unicorn in the Greater Bay Area, though by then the ranking mattered less than the threshold itself: RMB 20 billion was turning into the new minimum for the leading names.
Before them, six robot companies had already been valued above RMB 20 billion. By July, fundraising for robot-body companies had started to slow, but the valuation rush kept moving into dexterous hands and data. Three dexterous-hand companies tracked by the report were also valued above RMB 20 billion. One was just over a year old. The other two were worth less than RMB 1 billion at the start of last year and had risen 20-fold in 20 months.
Products and revenue did not rise in lockstep with those valuations. The report says multiple primary-market investors were asked what measurable formula supported such pricing, and none could provide one.
One embodied AI investor decided to quit in June. He described the market this way: 「In a field that is still at a very early stage, if everyone says a company cannot raise much money or reach a certain valuation, it loses its seat at the table. It sounds less like entrepreneurship and more like Texas Hold’em.」
That line captures the shift the report sees across venture investing. Instead of trying to find a non-consensus opportunity, firms now care more about entering the top projects, standing with the right institutions, and making sure the next round happens.
At a dinner in the first half of the year, one investor pitched a data collection company to friends at the table. His reasons included a founding team with what he called a "Nobel-level" academic background, defined in the article as people selected for or awarded ACM, IEEE, or related prizes; the company’s place inside a top industrial ecosystem; and a financing round that had already gathered industrial investors alongside leading RMB and USD funds. Little of the discussion focused on technical capability or commercial prospects. The consensus in that room was simpler: first-person data collection was the best trade.
As one person quoted in the report put it, "A student of a founding figure in the field, plus several top industrial investors and VCs — would anyone say no? I don’t think so."
Fundraising windows are often extremely short, sometimes just two days. Several investors told Late Post that embodied AI founders now behave like stars, releasing allocation only when they are meeting the boss of a fund. Executives at some companies even swap tactics over private dinners: wait a few days before accepting requests from investors, meet only one or two of them, and keep the team structure vague to build mystery. The harder the meeting is to get, the more people want it.
The report outlines a familiar startup template. A serial operator with cross-sector experience works behind the scenes. A PhD who studied overseas under a well-known figure in computer science or artificial intelligence returns to China as CEO and stands out front, offering a technically plausible framework and a distinctive narrative around the company.
Over the past year, another path has become even more typical: auto-industry technical executives leaving their jobs to enter embodied AI. In those cases, a non-technical operator with business experience handles the boardroom side, while a former automotive executive serves as CEO or CTO and becomes the company’s public technical face. The pitch is that someone from a carmaker with experience commercializing autonomous driving at scale knows productization better than a professor does.
The most extreme example in the report involves an executive from a new-energy vehicle company. After leaving, his new startup was already discussing a second round before it was formally incorporated. Half a month after registration, with fewer than 20 employees, it was valued at $1 billion.
For investors who get in early, these organized syndicates produce a clearer mark-to-market gain and a more visible exit path. One source summed it up plainly: "When people see this lineup, a whole crowd wants in behind it." If a project’s background is strong enough, getting to a RMB 5 billion or RMB 10 billion valuation is no longer that hard.
One way to push pricing higher is to promise at the current round that the same lead investor will return to lead the next round too. In a matter of months, the valuation can double. The report says that can even produce different prices for the same period. In one blunt line from the article, "People inside the circle make money from people outside the circle." In effect, the investor roster itself becomes part of the valuation.
This kind of club deal appeared in large language model investing too, where a group of investors backed one project together. Back then, top funds still excluded each other more often. In embodied AI, names such as Sequoia and Hillhouse now appear together much more frequently. One investor’s explanation was simple: "If everyone can make money together, why reject each other?"
Old Unitree shares, industrial tie-ups, and layered incentives kept capital moving
Larger financings also came with more elaborate exchanges of access and benefits.
A manufacturing CVC visited Wang Xingxing in 2024 and did not invest in Unitree at the time. After February 2025, the picture changed completely. Unitree’s robots appeared on the Spring Festival Gala and the company became, after DeepSeek, another breakout Chinese tech name. Capital rushed in. Later, even when the chairman of that CVC personally sought a meeting with Wang, he still could not secure an allocation in the next round.
The investor then looked for early shareholders willing to sell old shares in Unitree. According to a person familiar with the deal, one early holder agreed to transfer part of the stake. After that, the same CVC also invested in another embodied AI company that this early shareholder had helped found.
In the middle of last year, the CVC led a new round in that startup. Soon after, it placed an order for quadruped inspection robot dogs for inspection use cases. By the middle of this year, it had purchased the company’s robots again and deployed them in its own factory. That startup plans to list in Hong Kong this year.
Robots were sold, but durable demand still has not been proved
When Unitree published its prospectus, the market got its first full look at the revenue mix of a leading general-purpose robot company.
In 2025, Unitree recorded RMB 1.699 billion in revenue. It sold 5,215 humanoid robots, which accounted for nearly half of total revenue. As of last September, more than 70% of its revenue came from research and education, including algorithm research, teaching experiments, and secondary development. Less than 10% came from industry applications, and more than half of that was used for corporate reception and guided tours. Revenue clearly tied to smart manufacturing, intelligent inspection, and logistics delivery was only RMB 15.7 million.
Its customer base was also highly fragmented. The top five customers contributed 10.61% of revenue in total, while JD.com, the largest customer, accounted for 3.54%. The report says there are still no industrial customers buying and using the robots at meaningful scale.
Li Yuanqing, co-CEO of Lexiang Technology, said at a media session that any kind of robot can sell 500 units because 500 competitors will buy one to study it.
Unitree, though, is hard to replicate. The report notes that the company has spent 10 years building, has a consumer-grade quadruped product line, benefits from the branding effect of the Spring Festival Gala, and offers a price advantage at RMB 99,000. Selling more than 5,000 humanoid robots last year made it the global leader. A Unitree employee quoted in the story said only two or three peers sold more than 1,000 units last year. Most others were still in the hundreds.
Research and exhibitions cannot absorb dozens of robot companies. Another top robot-company CEO told Late Post earlier this year that the most active buyers in 2025 were newly built data collection factories around China. These facilities buy robots, ask human data collectors to wear equipment and operate them through actions such as folding quilts, wiping tables, and sorting objects, then sell the resulting visual and trajectory data to robot companies.
Some robot companies sign repurchase arrangements with those facilities. They book revenue by selling robots to a data factory, then return part of the money by purchasing data from the same facility.
Two data-collection startup founders said such projects are attractive because inputs are relatively controllable, concentrated in robot capex and land, and they create jobs in a predictable way. The projects also fit local preferences because they combine manufacturing and intelligence and involve no pollution. They are often pitched first to local governments that missed the EV and LLM booms, with the claim that this could become an even bigger opportunity.
According to those founders, real-machine data now sells for more than RMB 700 per hour. A data factory with 100 robots can produce about 10,000 hours of data per month. At a robot price of RMB 500,000, and assuming full utilization while ignoring labor, site, and maintenance costs, the factory could recover its purchase cost in eight months.
On the ground, the article says, conditions look worse. Several people who visited data factories saw robots sitting idle for long periods, while already collected data had no buyer. One facility that had bought roughly RMB 80 million worth of robots earlier ended up reselling some of them to schools at the start of this year because the payback period had exceeded expectations.
Most robot companies have not yet begun buying data at scale, because the industry still has not agreed on the most effective data format. Robot bodies differ in structure, sensors, and camera positions, so data reuse across platforms still requires extensive conversion and adaptation. Real transferability remains limited.
Technical routes, data formats, and acceptance standards are also unsettled. Demand on paper does not quickly convert into orders. Data bought today may lose value quickly if the technical route changes.
Some robot companies have instead formed joint ventures with local state-backed capital. These JVs issue procurement projects, build data factories, and then award supply contracts to the cooperating robot companies. Industrial groups with close local ties play a central role in arranging and joining those projects.
A prospectus disclosed by one robot company in the second quarter of this year showed that four of its top five customers last year were linked to local governments or state-backed systems, contributing more than 30% of revenue together. Its biggest customer was also a shareholder, and the two sides jointly invested in a JV to run a robot training center. In other words, local state capital simultaneously acted as customer, investor, and operating partner.
In the middle of last year, a listed company announced a strategic agreement with local state-backed capital and a robot company. The three planned to form a JV to build robot data collection, training, and deployment scenarios. The JV intended to provide purchase orders to the robot company, which would supply products and technology. The listed company would undertake core components, joint orders, and supply-chain support. The same listed company was also an LP in the robot company’s investment fund, while its subsidiary served as a core distributor.
One robot executive told the publication that robotics is a new industry where every layer needs innovation, and that includes the business model.
The article says multiple identities and cross-transactions are already common. Industrial partners can earn in several ways at once: selling parts to the robot company, taking OEM work, investing through funds, running data centers, and then purchasing robots for factories or training sites. Robot companies gain production capacity, scenarios, investment, and orders.
Still, this revenue has become controversial in capital markets. Late Post says that when one leading robot company submitted listing documents in the fourth quarter of last year, auditors required some revenue with characteristics of related-party transactions to be removed.
Another active group is listed manufacturing companies, especially in the Apple supply chain and EV supply chain. When humanoid robot stocks heated up last year, many supply-chain companies wanted a new growth path into what they described as a trillion-yuan robotics industry. That participation went beyond taking equity stakes, sending samples, or buying robots. The robot business also became a new explanation for future growth in investor communications.
Last year, a lithium battery materials supplier bought robots from Zhiyuan and announced a supply agreement for core robot joint modules. The two companies became both customers and suppliers to each other. On the day the announcement was released, the supplier’s stock rose 20%.
The company’s board secretary had earlier told analysts, according to the report, "If you value us using a robot company model, we’re no longer an automotive parts supplier." At the start of this year, after lithium battery supply-chain prices increased and sector stocks rebounded, the same executive said, "We are deeply tied to CATL. We are a leader in LFP."
A senior foreign analyst also recounted that an IR executive at a Tesla automotive parts supplier once showed her photos of taking Optimus procurement staff to karaoke, saying the Tesla buyer had sung more than ten songs and had never enjoyed a China trip more.
Industrial customers paid for robots, placed them on production lines, then recorded promotional videos saying the robots were already working in their factories. That created a deployment story for the robot maker. At the same time, listed manufacturers that were shareholders and parts suppliers could also gain from rising startup valuations and stronger secondary-market stock prices. The report gives one example: after Joyson Electronics announced cooperation with Zhiyuan in August last year and held a robot parts product event, its stock climbed 95% within a month.
The report also says some robot companies, while discussing allocation with investors, hinted or asked them to buy robots or dexterous hands in proportion to their investment stake. Last month, the chairman of a Tesla supplier told Late Post that one dexterous-hand company already in the listing process had cited shipment figures above the true industry-wide total for the first half of the year, adding, "A lot of those may have been bought by investors."
Channel sales and ecosystem partners are helping inflate top-line numbers
Before real demand arrives, the report says, some companies are trying to expand the ecosystem by making the upstream and downstream profitable first. Zhiyuan is presented as a typical case.
In August last year, Zhiyuan introduced a partner system around robot sales at its ecosystem conference. These partners function much like auto dealers: they buy Zhiyuan’s robots, then, with support from Zhiyuan, look for large customers and use cases. Higher sales bring more support from Zhiyuan.
The framework split partners into four tiers — VAP, Gold, Silver, and Certified — with annual sales thresholds of RMB 20 million, RMB 10 million, RMB 5 million, and RMB 2 million.
That year, ten companies including Joyson Electronics, Ningbo Huaxiang, Yushu Intelligence, Xiaolu Intelligence, Swancor, Shandong Zhixing, Henggong Technology, Wolong Electric Drive, and i-city won VAP seats. Some of them were also Zhiyuan’s suppliers, OEM partners, JV partners, or portfolio companies.
Joyson was the strongest among them. In October last year, the two sides announced robot procurement orders worth more than RMB 100 million. Before that, Joyson had also signed on to supply head assemblies, inertial sensors, energy management systems, and other robot parts, while its subsidiary Puzhi Future became one of Zhiyuan’s contract manufacturers. Joyson had disclosed that its embodied-intelligence brain solution had already been deployed in some industrial scenarios and automated charging scenarios.
According to Deng Taihua, Zhiyuan’s revenue had grown 20 times a year over the past three years. In the first quarter of this year, revenue already exceeded RMB 1 billion, equal to the company’s full-year revenue last year. During its July listing roadshow, Zhiyuan guided for shipments of 16,000 robots and revenue of RMB 4.5 billion. By the end of 2027, its goal is RMB 10 billion in revenue.
Thousands of partners, including industrial, city, and channel partners, are expected to become major revenue sources. In Zhiyuan’s RMB 1 billion of 2025 revenue, partner sales rose from close to zero in the first half to 25% for the full year. In 2026, the company plans to increase the partner-sales share to at least 60%.
The article notes that cars, servers, and industrial equipment all rely on distributors to expand markets. But for robots, whether revenue can last depends on different questions: how many end customers bought products without equity ties or supply-chain ties, who is carrying the inventory, whether customers completed acceptance, and whether reorders follow. That comes back to the robots’ actual capability.
More money is coming in, but technical progress still has weak proof points
An investment banker who had attended roadshows for three robot companies recalled visiting one embodied AI company valued above RMB 20 billion in May. Engineers had the robot demonstrate towel folding. The process lasted 15 minutes, and the towel still was not folded by the end.
In another demo, a robot was shown moving boxes. The boxes were special standard parts with fixed slots on the bottom. The robot only had to align claw plates with the slot, insert them, and lift the box the way a forklift would. "I tilted the box slightly and asked the engineer whether it would still work. He said no," the banker said.
Those demos point to a problem investors keep running into. Embodied AI companies have raised more and more money, but they have struggled to show how much better the robots actually became as a result.
Over the past year, the hardware body and low-level motion-control algorithms of leading robot companies have matured significantly. Stable walking is no longer exclusive to just a few players. The robot "brain," though, is still in a very early stage. Many robots can complete tasks only in pre-set environments. Once object position, object shape, or operating steps change, success rates fall sharply. Most companies still have not solved generalization, stability, or heavy dependence on human intervention.
Many founders keep repeating that robotics will become the largest industry in human history. From the angle of R&D and capital expenditure, however, the sector may not yet have entered a phase consistent with the scale of money it has already raised.
UBTECH founder Zhou Jian told Late Post earlier this year that his company spent RMB 500 million on R&D last year, possibly the highest level in China’s robot industry. Two investors said Zhiyuan spent more than RMB 500 million on R&D last year as well.
Many others spent far less. Unitree’s prospectus showed R&D spending of RMB 145 million last year. Another embodied AI startup that had already crossed a RMB 20 billion valuation in the first half and had raised roughly RMB 5 billion in total spent less than RMB 40 million on R&D last year, most of it salaries for research staff. One investor said that if the company simply put the money into wealth-management products, the returns alone could cover daily costs and turn it into a "perpetual company."
Late Post also said that the four robot companies that appeared on this year’s Spring Festival Gala — Unitree, Songyan Dynamics, Magic Atom, and Galaxy General — each spent close to RMB 100 million on that appearance, more than many companies spend on R&D in an entire year.
Most primary-market investors do not see low R&D spending as entirely irrational. The technical route in embodied AI has not converged, and making large, irreversible investments too early carries high risk. That has produced a strange capital pattern in the sector: companies keep raising money, yet the money does not quickly turn into technical advantage and often just sits on the balance sheet.
One investor in Variable and other embodied AI names said companies valued above RMB 10 billion had theoretically already raised enough. But the fundraising race goes on. "What is the purpose of raising more? I can’t figure it out," he said.
As hardware routes gradually converge and supply chains mature, new R&D spending is concentrating in two areas: computing power and data.
At the start of July, the founder of a company just 10 months old told the publication that the biggest embodied AI computing resources in China are held by large companies such as ByteDance and Alibaba. Among startups, only eight have reserves above 1,000 GPUs, and his company is one of them. Unitree is also in that category. According to the report, Unitree spent about RMB 30 million on computing last year and has already assembled a robot-brain research team of more than 50 people.
Even so, enough compute is only a prerequisite. Compared with large language models, embodied models are more constrained by the lack of effective data. Three executives at data collection companies said the highest output in the industry in the first half of this year was only tens of thousands of hours of first-person data. The biggest order any of them had received so far was only worth several million yuan, corresponding to roughly 10,000 hours. The main buyers were still large companies such as Alibaba and ByteDance, which already had stronger compute resources.
The industry has not even settled on which data is most valuable. More companies are trying first-person-view data, in which ordinary people wear hats with cameras along with hand and foot sensors to perform actions, hoping to scale data more cheaply. There is still no consensus on how far this can substitute for real-machine data.
Model routes are no more settled. Over the past year, the narrative moved quickly from VLA to world models. But under the same labels, model structures, training targets, and data requirements differ across companies. The vocabulary is converging faster than the underlying technical route.
Embodied AI also lacks clear delivery milestones. Autonomous driving companies update smart-driving versions and city rollouts every few months. Large model companies keep releasing new models. Robotics has no common evaluation system and no broadly accepted milestone for progress. A company can keep saying it is still in development, and the underlying technical assumptions are hard to falsify in the short term.
Compute, data, and talent can all be bought with money. What no one can answer cleanly yet is how those inputs turn into better generalization in unfamiliar environments. At this stage, financing buys time for exploration and room to test different technical routes. It does not buy a form of competitiveness that can be verified right away.
After Tesla, Unitree has become the next pricing anchor
Capital and industrial resources will not stay patient forever. Confidence also needs to be supported by visible technical and commercial progress.
The embodied AI frenzy was first lit by Tesla’s Optimus project. Elon Musk said it could become the largest industrial opportunity in human history, reshape the social and economic structure, and eventually reach a global installed base of 10 billion humanoid robots.
That long-term story was then simplified into an easier line: robots will eventually have the price of cars and the installed base of smartphones.
The first to put money behind that idea were Tesla suppliers and public-market investors. Several people told the publication that betting against Musk over the past 20 years had not worked. At peak enthusiasm, hundreds of supply-chain companies were still sending samples to Tesla, and consulting calls with Optimus engineers were priced at RMB 9,600 per hour. If a Tesla supplier said it planned to shift into robotics, its valuation could jump sharply. The market was even willing to believe makers of car doors and headlights could make robot parts for Tesla. It was only in the second half of last year, when the field became too crowded and there were rumors that six or seven suppliers were trying to supply some joint actuators to Tesla, that someone finally pushed back: "Is Tesla sick enough to ask a car-door maker to build actuators?"
Over the past few years, Optimus production timelines and output targets have been adjusted several times, and the wait for real orders has dragged on. Tesla suppliers and A-share robot investors began to reassess that progress. In July this year, the chairman of another Tesla supplier told Late Post, "The past three years were wasted. The valuation of the robotics business has fallen away."
China’s private market, however, has not lost interest. Investors can still treat an embodied AI startup as an industrial option: if general-purpose robotics eventually breaks through, the returns on today’s capital could be huge. That logic rests not only on technical progress, but also on the expectation that one day these companies can list and receive a public-market price.
Unitree’s listing answered one major question: public markets are still willing to price long-dated expectations in embodied AI. Even so, Unitree is unlike most startups in the field. It does not rely on fundraising to survive. It has already proved it can mass-produce robots, generate real revenue, and keep relatively healthy gross margins.
Its current commercial edge is concentrated in the body, low-level motion control, and engineering mass production. Several founders and investors told the publication that Unitree’s real advantage today lies in PCB routing design, motor parameter tuning, and supply-chain management. Those capabilities come from more than a decade of accumulated know-how by Wang Xingxing and his team, and are hard to measure through current-period R&D expense alone.
The "brain" is a different matter. Unitree is developing embodied models such as WMA and VLA and plans to invest more than RMB 2 billion in underlying robot brain and low-level control technologies. But those capabilities have not yet been tested across full application scenarios and commercialization paths. More spending on models, compute, and data would also lift R&D costs and could weigh on short-term profit.
Based on existing revenue and product capability, the report says Unitree still looks more like a strong intelligent hardware company with real engineering and manufacturing strength. It has strong hardware manufacturing, clear brand recognition, and healthy gross margins. Whether it can become an embodied model company remains unproven.
Investors in large model companies know they are betting on the idea that model capability will keep improving. Hardware bodies do not iterate nearly as fast as large models. If public markets are willing to pay a high enough premium for Unitree’s still-unrealized "brain" capability, private-market embodied AI startups will gain new room for valuation expansion.
If the market prices Unitree mainly on hardware-company revenue and profit, then many startups that still have no revenue, no profit, and no mass-production capability will have a much harder time explaining why they are worth so much.
One embodied AI investor summed up the past few years with a short rhyme: "Question the bubble, understand the bubble, embrace the bubble, enjoy the bubble." He did not say what comes next.