Embodied Intelligence Funding Cools: IPO Pressure Amid Valuation-Revenue Mismatch

“Zhiyuan Robotics' pace of listing in Hong Kong has already slowed.” “Selling old shares at a 50% discount, and no one is willing to take them.” “To restrict the outflow of old shares, some embodied intelligence companies have already refused shareholding reform.” “The real financing pressure on embodied intelligence companies may become concentrated after November.” “Investors are neither paying attention to world models nor looking at embodied intelligence projects anymore.”

Since the second half of 2026, controversy over the valuation bubble in embodied intelligence has continued to ferment, and financing data intuitively confirms this change: from October 1 to 15 this year, the sector saw 19 financing deals totaling 4.934 billion yuan, a sharp contraction in heat compared with the peak period's single-month total financing of 33.98 billion yuan. If one also considers that many disclosed transactions are actually transfers of existing equity, the shrinkage of truly new capital in the sector is far more severe than the headline numbers suggest.

Image source: IT Juzi

But wasn't the capital market still frantically scrambling for embodied intelligence in the first half of the year?

In the first half of 2026, investment institutions feared missing the trend, and embodied intelligence got the same hot script that autonomous driving had back then. Angel rounds of hundreds of millions of yuan, multiple financing rounds completed in succession within months, and rapid growth into unicorns within half a year—embodied intelligence became the market's “money magnet.”

IT Juzi data shows that in the first eight months of 2026, financing in the robotics sector reached 124.51 billion yuan, twice the full-year total for 2025. The same frenzy also played out overseas. PitchBook data shows that in the first half of 2026, robotics and physical AI companies raised $33.4 billion globally, already exceeding the full-year 2025 total of $28.7 billion.

From fearing missing the trend to now becoming cautious, why has the capital market undergone a 180-degree reversal? After the boom fades, what new pressures will the embodied intelligence industry face?

4.33 million yuan in revenue vs. $500 million valuation: the mismatch between revenue and valuation

Over 36 trading days, Unitree Technology's stock price fell from an opening price of 1,100 yuan per share to 404.03 yuan per share, a maximum drop of more than 63%, wiping out more than 280 billion yuan in market value. Unitree Technology's listing became the trigger for the capital market's reversal in attitude.

Image source: Xueqiu

Beyond Unitree, robotics components manufacturer Mech-Mind's public offering won about 3,835 times oversubscription, setting a new subscription record for the Hong Kong stock market's robotics sector. But it broke issue price on its first trading day, and its share price has continued to weaken since.

Image source: Xueqiu

Successive broken issues have made the capital market rethink: was capital previously paying for the halo of embodied intelligence's star teams, or was it optimistic about the sector's real, deliverable long-term growth?

Currently, the founding teams of embodied intelligence unicorns and future unicorns fall roughly into two categories:

One is academic teams represented by top university research teams and professor-experts. The Xinor Future team brings together talent from Tsinghua, Zhejiang University and other top global universities, while the core team of Jijia Shijie comes from institutions including Peking University, Tsinghua, the Chinese Academy of Sciences and USTC.

The other is industry-migration entrepreneurship led by technical leaders, business executives or core R&D personnel from big tech companies. The founding teams of Yuanli Lingji and Sharpa come from Megvii Technology and Hesai Technology respectively, while XPeng Robotics was spun out of XPeng Motors.

Star teams + demos + technology narratives are the key support for unicorns' multi-billion-dollar valuations and future unicorns' multi-billion-yuan valuations. The industry has even seen “PPT financing” phenomena relying on star-team endorsements.

But with the slow pace of commercialization in embodied intelligence, can these high valuations be sustained?

“A robot being able to perform and compete is one thing; a robot being able to stably work in a factory is another.” Zhang Lei (pseudonym), a quality inspection manager at a manufacturing enterprise in Anhui, told Physical AI Grand View. In his view, large-scale deployment of robots in factories requires getting three accounts clear.

First, the economic account: only if it is cheaper than human labor and more efficient will factories be willing to repurchase. “Suppose a production line worker's annual salary is 70,000 to 80,000 yuan. If the robot is cheaper and yields higher output efficiency, the boss will naturally be willing to repurchase and gradually transform the production line. Otherwise, it is very hard to advance.”

Second, the stability account: robots need strong reliability and adaptability to complex working conditions. “When customers are pressing for delivery, workers can work overtime to rush the job. Once a robot suddenly breaks down and takes a day or two to repair, who bears the economic loss caused by delayed delivery?”

Third, the responsibility account: the boundaries of rights and responsibilities must be clear. “A lot of manufacturing production emphasizes fixed posts, fixed people and fixed responsibilities, while embodied intelligence robots won't ‘take the blame.’ If robots cause more production line failures, responsibility ultimately falls on the department head—at best a pay deduction or performance penalty, at worst possible dismissal by the company. This naturally makes them resist deployment, and frontline workers, out of worry about being replaced, may also make things difficult for the robots. Can robots really adapt to this complex reality of human relations and responsibilities in industrial scenarios?”

Zhang Lei's concerns are not isolated. Data from the Humanoid Robot Scenario Application Alliance shows that in the first half of 2026, there were 218 publicly disclosed humanoid robot tender projects in China, with a total scale of 1.723 billion yuan; among them, educational and research institutions accounted for 55.5% of procurement, government and state-owned capital platforms 20.6%, and orders from industrial enterprises only 21.1%. And among industrial orders, the vast majority remain at prototype testing, showroom display and training pilots, rather than routine production line procurement.

Total orders are limited, and the market also shows a highly concentrated pattern. Statistics from SAG for the first half of 2026 show that Zhiyuan Robotics and Unitree Technology together account for more than 70% of the global market share; other small and medium-sized manufacturers all shipped fewer than 1,000 units in the first half of the year, and the scale gap between top-tier and mid-tier companies has already opened up.

Li Yuanqing, co-CEO of LeXiang Technology, likewise said that no matter which robot is launched, at least 500 units can be sold, and a large number of competitors will buy them to disassemble and reverse-engineer. But Unitree has a decade of accumulation, a consumer quadruped product line, and the price advantage of 99,000 yuan, making it difficult for outsiders to replicate; only two or three peers in the sector have shipped more than 1,000 units.

Insufficient shipment volumes have caused a severe mismatch between embodied intelligence revenue and valuation. Public data shows that CloudMinds Robotics had revenue of only 1.4 million yuan in the first seven months of 2025 and a net loss of as high as 84.25 million yuan, yet at its peak its valuation reached 22.3 billion yuan. Yinghe Robotics, with full-year 2025 revenue of only 4.33 million yuan, had a peak valuation exceeding $500 million. This mismatch has gradually made the capital market more sober about embodied intelligence projects.

Behind this is the fact that a huge gap still exists between the maturity of embodied intelligence technology and the real needs of enterprises. On one side, the current embodied intelligence industry chain—from upstream core components and complete machine manufacturing to downstream scenario adaptation, deployment and operations, and customer payment—still faces multiple difficulties such as high costs, fragmented scenarios, and long investment return cycles.

On the other side, however, manufacturing requires embodied intelligence to overcome perception, decision-making and fine manipulation challenges in complex environments, and also to achieve high reliability, low deployment costs, and stable positive economic returns, and moreover to establish clear fault responsibility allocation and smooth human-machine collaboration mechanisms. The three thresholds of technology, cost and management leave embodied intelligence facing long-term commercialization pressure.

Valuations under pressure, IPOs encountering obstacles

Long-term commercialization anxiety has given rise to two special transaction models in the embodied intelligence sector:

One is the circular transaction model represented by data collection centers. Shao Tianlan, founder of Mech-Mind, publicly criticized this model: "Many 'assembled' embodied intelligence companies now, including some in Beijing and Shanghai that are very famous, highly valued, and have appeared on Spring Festival Galas, make extensive use of so-called 'data collection centers' to conduct related-party transactions with local governments, investors and suppliers, creating false and unsustainable revenue."

The Financial Times dissected how this model operates: local governments and robot companies contribute funds to build data collection centers → the data collection centers purchase robots from the companies for data collection and training → the collected data is then sold back to the robot companies, forming a circular transaction of "local government builds center → center purchases robots → collected data sold back to companies."

Another is the "both shareholder and customer" transaction model. Large industrial capital both participates in investment or joint ventures and becomes a customer of embodied intelligence companies, while also possibly supplying core components to the company, forming a relationship in which capital, products, supply chains and application scenarios are intertwined.

When the market shifts its focus to the commercialization and implementation of embodied intelligence, a series of questions follows: How many real orders do companies actually hold? Is there mutual price bidding to push up valuations? Are there benefit transfers and related-party transactions? After robots enter enterprise scenarios, what level of stable operating hours can be achieved? After detaching from shareholder customers, can companies continue to develop new external customers?

These doubts have begun to subject embodied intelligence companies to valuation correction pressure. Reuters reported that if revenue related to data collection centers is excluded, the valuations of some robot companies may fall by 60%–70%, and current private equity project valuations have already been cut by about 30%–50%.

Behind the valuation cuts is the market's difficulty in seeing clearly the underlying quality of companies' operations. If share prices continue to weaken after listing, the risks will be exposed in the secondary market. Regulators have therefore tightened IPO reviews for robot companies, and the pressure on companies to go public has risen significantly.

Cailianshe reported that revenue authenticity, profit stability and commercialization implementation capability have become core review orientations for robot IPOs. Huang Lichong, president of Huisheng International Capital, proposed that when embodied intelligence companies go public, the hardest thing to explain to the capital market is: how much of the value created by technology can ultimately remain on the company's books. Improved model capabilities and improved customer efficiency do not equal synchronous growth in corporate profits. "Are customers willing to pay more? Does the company have pricing power? After deducting computing power, data processing, implementation and manual review, how much is left? These all need to be calculated clearly."

Wind data shows that more than 50 embodied intelligence companies are queuing for listing on the Hong Kong Stock Exchange. Although it is still difficult to judge how many companies will ultimately complete issuance successfully, cases of companies actively adjusting their listing pace have already appeared in the market: Weiyi Zhizao, originally scheduled to list on October 28 and aiming to become the "first Hong Kong-listed industrial embodied intelligence stock," temporarily suspended its listing process; Zhongjian Technology also announced the termination of its H-share listing plan.

With tighter IPO reviews and high risk of breaking issue price hanging overhead, investment institutions have become more sober about the embodied intelligence sector. At present, embodied intelligence projects still rely heavily on IPO exits. Suppose a target enters at a valuation of 10 billion yuan, an IPO issues a 10% free float, and after listing the market value surges to 30 billion yuan, then falls back to 15 billion yuan after the six-month lock-up period ends. The paper return appears to be 50%, but when time costs, liquidity discounts and operational uncertainty are added, the actual return attractiveness is greatly reduced.

The capital market's caution, in turn, constrains the IPO process of embodied intelligence companies. Take Hong Kong IPO subscriptions as an example: cornerstone investors must accept a six-month lock-up period after subscribing for shares, and UDI Robotics' lock-up period is even as long as nine months.

Embodied intelligence companies have relatively high valuations, and the cornerstone funds required for issuance range from tens of millions of Hong Kong dollars to hundreds of millions of Hong Kong dollars. Public data shows that Mech-Mind introduced a total of 9 cornerstone investors, who subscribed for about $186 million in total.

Data from CVSource shows that as of October 2026, among the 13 robotics companies listed in Hong Kong, 10 introduced cornerstone investors, with total subscriptions of approximately HK$4.861 billion. But among these 10, the cornerstone investor lineups of Jingfeng Medical and Huayan Robotics were sitting on unrealized losses of HK$152 million and HK$457 million, respectively, for a combined loss of HK$609 million.

Image source: CVSource

With other institutions' unrealized losses as a cautionary tale, a lock-up period of up to six months, and post-listing share price uncertainty, will institutions be willing to enter in large amounts in subsequent embodied intelligence IPOs? Will investors be willing to bear the risk of internal accountability brought by a broken issue? And amid all this uncertainty, will there be a situation where old shareholders are unwilling to backstop and new shareholders dare not enter? The pressure on embodied intelligence companies' IPO financing and subsequent capital continuation has been further amplified.

As the boom fades, the cyclical test for embodied intelligence

The key to resolving financing pressure remains for embodied intelligence companies to return to R&D itself and accelerate commercialization. In 2026, most of the sector's capital flowed to the midstream and upstream of the industrial chain. According to incomplete statistics from Physical AI Grand View, among 33 robotics unicorns, 29 are concentrated in the midstream and upstream, accounting for about 88%; among 36 future unicorns, 31 are in the midstream and upstream, accounting for about 86%.

Data from iResearch shows that in 2026, financing for the global robotics software AI foundation and embodied intelligence brain segments reached RMB 167.093 billion and RMB 162.552 billion, respectively, far ahead of others; financing for downstream industry applications, public service robots and other implementation ends was only more than RMB 1 billion. Capital has poured heavily into the algorithm layer, while the complete-machine and implementation tracks responsible for mass production delivery and scenario validation are short of funds.

Image source: iResearch

But when GPT-6 Astra can already complete tasks such as 3D modeling and physical AI simulation; Kimi K3.1 Pro can directly write Blender and CAD scripts to build 3D models, and Astra is continuously challenging the industry's original technical barriers, will capital continue to bet on the upper-layer model track of embodied intelligence?

For companies focused on embodied intelligence models and world models, if their core advantage remains only in upper-layer capabilities such as environmental understanding and task planning, they may be covered and replaced by general large models. Yuan Rong Qixing CEO Zhou Guang mentioned that after the advent of GPT-6, robot body companies like Unitree will instead have more advantages, and companies focused on embodied intelligence brains and robot large models will face huge challenges.

In contrast, enabling robots to stably complete complex operations in real physical environments and continuously improve task success rates through real interaction data is the competitive barrier that is harder to replace, and this further raises the R&D threshold for embodied intelligence companies.

Compared with the impact brought by the iteration of external general models, Physical AI Grand View believes that deeper R&D risks come from within companies. Currently, embodied intelligence companies fall roughly into two categories: one is rooted in real scenarios for the long term, steadily polishing products; the other is companies rapidly forced to maturity by capital, maintaining a high-cost, high-salary model for a long time. For example, Yinghe Robotics was reportedly exposed for spending RMB 28 million to renovate offices and paying out RMB 5 million in year-end bonuses at once, triggering strong dissatisfaction among investors.

English text is being prepared.

Not only on the R&D side, but the commercialization implementation stage also faces talent adaptation difficulties. Zhu Yang (pseudonym), a biology PhD from USTC, mentioned that his senior fellow student previously wanted him to join an embodied intelligence company and offered very considerable pay, but he ultimately chose to stay in academia.

He explained that many university professors carry master's and doctoral students with them year-round, hold research projects and funding, and can even influence students' graduation. They are in the Party A position, with clear boundaries of pressure and responsibility. But once they enter a company, setting aside environmental differences, the biggest change is identity:

Doing research at school, you are Party A; doing embodied intelligence at a company, you have to become Party B serving clients, and all kinds of demanding requirements must be accepted. As a professor, the core is applying for funding and spending money to do research; as a company, the core is making money and also bearing business risks. This gap in identity and mindset is very difficult for many people from a research background to adapt to.

Embodied intelligence remains an important long-term technology trend. This round of capital cooling-off period is only accelerating the industry's return to its industrial essence: technology cannot stop at the laboratory, products must withstand real-scenario testing, and companies ultimately must prove their value through commercialization capability. The boom will eventually fade, and only companies that truly create industrial value will have the chance to survive the cycle.