Embodied intelligence financing cools down, multiple robotics companies fall into work stoppages and bankruptcy

In mid-October Shanghai, the osmanthus is in full bloom, and the warm autumn sun already carries a hint of chill.

This is not an isolated case.

Some embodied intelligence companies are also consciously preparing to weather the winter: laying off staff, scaling back operations, pausing expansion, while at the same time securing one last round of financing before winter sets in. Big money is better; if they can't get the big checks, they take whatever they can get. The reasoning is consistent: next year's money may not be as plentiful or as hot as this year's.

Some companies did not manage to wait for the next round of money.

Nanjing Wujing Dongli was reported to have halted work and owed wages just eight months after its founding, with both employees and the CEO saying they could not reach the actual controller; Shenzhen Yinghe Robotics held its first creditors' meeting on October 19; and Shanghai Shenlan Robotics had its bankruptcy liquidation accepted by a court in May...

IT Juzi data shows that in the first half of 2026, the embodied intelligence industry raised 93.5 billion yuan across 322 deals, with the number of financing events up 137% year on year, and the number of unicorns valued at over 10 billion yuan rising from 3 to at least 22.

Another tally based on publicly disclosed amounts recorded 288 financing rounds involving 226 companies, with a disclosed total exceeding 46 billion yuan; among them, the top 20 companies raised about 33 billion yuan combined, accounting for more than 70%, while the remaining 200-plus companies shared about 12.4 billion yuan.

Image source: IT Juzi

Large financing rounds came one after another.

In March, Galaxy General completed a 2.5 billion yuan financing led by the National Artificial Intelligence Industry Investment Fund; Qianxun Intelligence completed four consecutive rounds totaling nearly 5 billion yuan, with its valuation reaching 20 billion yuan; Zhipingfang's valuation rose from 10 billion yuan to 20 billion yuan in four months; in April, Tashi Zhihang raised $455 million in a single round, setting an industry record at the time; in June, the Global Embodied Intelligence Alliance for Humanoid Robot Scenario Applications tallied monthly global embodied intelligence financing at 31.5 billion yuan.

New entrants are also getting younger and younger.

Huang Yi, born in 2004, founded RoboParty in 2025 and had raised nearly 500 million yuan by July 2026; Yuanche Taichu, founded by Tsinghua doctoral student Qin Shentao, raised more than 500 million yuan within five months of operation; Inverse Matrix, founded by Chen Boyuan, a graduate of Peking University's Yuanpei program, raised more than $100 million in its seed++ round at a valuation of over 5 billion yuan.

The peak of the frenzy was August 19.

Unitree Technology listed on the STAR Market at an issue price of 150.80 yuan, opened at 1,100 yuan, up 629.44% from the issue price, and its market value briefly reached 444.9 billion yuan. With a 0.0181% allotment rate, corresponding to a maximum paper gain of about 475,000 yuan per winning lot, its issue price-to-earnings ratio of 219.23 times was also far above the 38.56 times static P/E ratio of the general equipment manufacturing industry to which it belongs. One investor put the logic bluntly: "Don't agonize over valuation, just get the money."

Unitree Technology goes public

The first voices of a turning point came from those who had invested.

Wang Huadong, a managing partner at Matrix Partners China, was the sole lead investor in Unitree's 2021 B1 round. In the video podcast "Xiangtan," he recalled walking into a robotics lab in Silicon Valley in March 2024, where eight Unitree robots were still usable while products from other brands were basically broken. He believes that among Chinese companies capable of achieving absolute global leadership in a field, DJI is one and Unitree is another.

But in the same interview, he also offered two judgments: angel investments running to hundreds of millions of yuan and valuations in the billions are unprecedented in the history of Chinese business, including the craziest period of the mobile internet; and the concentrated listings of leading companies are very likely the high point of this cycle.

Forty days later, the second milestone arrived.

On September 28, AMD announced it had signed a definitive agreement to acquire World Labs, the world model company founded by Fei-Fei Li, in an all-stock deal valued at about $8.2 billion; the transaction still requires regulatory approval and is expected to close before the end of 2026; after closing, Fei-Fei Li will serve as AMD executive vice president and chief scientist; World Labs had a public valuation of about $5 billion to $5.4 billion after its February financing, representing a rise of about 50% to 60% over seven months compared with the deal price.

AMD Chair and CEO Lisa Su (left), World Labs founder Fei-Fei Li (right)

One embodied intelligence investor put it more bluntly: after more than two years in existence, it had only produced demos, and the path for world models was still unclear; this deal looks more like cashing out and heading for shore.

In the eyes of some domestic embodied intelligence founders, the world model field seemed to have suddenly "lost its standard-bearer," with panic outweighing envy.

On one side, a domestic humanoid robot company set an astonishing first-day IPO price; on the other, a star scientist chose to sell her company at a high point.

The two milestones, forty days apart, together remind the market: at the hottest moment, some people were already arranging their exit.

How the money cooled

The money hit the brakes in September.

In early September, The Information reported, citing people familiar with the matter, that regulators had slowed the listing process for some humanoid robot companies through informal "window guidance"; Reuters followed up, noting that regulatory concerns included excessively high valuations and whether revenue tied to government-supported projects reflects real market demand; the CSRC did not respond to a request for comment.

On September 21, the Shanghai Stock Exchange terminated the STAR Market listing review of Huandong Technology after the company and its sponsor voluntarily withdrew the application. The company, which specializes in high-precision reducers for robot joints, became the first robot IPO project this year to have its review terminated on the STAR Market. In Hong Kong, as of mid-August, more than 50 embodied intelligence companies had filed but not yet listed.

On October 15, IT Juzi's data further showed that in the first half of October, the industry saw 19 financing deals totaling 4.934 billion yuan; in August, there were 91 deals totaling 31.323 billion yuan. October is counted for only half a month and spans the National Day holiday; even if the amount for the first half of the month is simply annualized to a full month, it is still less than one-third of August's.

The real temperature may be lower than the disclosed data.

Most of the financings announced now are deals that were already finalized months ago. As one industry insider put it: "The real changes in the market often take some time before they show up in the data."

The money has not left the sector; it has concentrated further at the top.

A report by Zhongtai Securities dated October 13 shows that in the first half of the year, embodied intelligence financing totaled about 43.8 billion yuan, of which companies with embodied brains as their core business received 22.253 billion yuan, while full-body robot companies received 5.6 billion yuan—the former nearly four times the latter.

The top players are still grabbing money.

According to Caijing, AgiBot has launched a new round of 1 billion to 1.5 billion yuan in financing at a valuation of about 50 billion yuan; as of the report, AgiBot had not responded. Stardust Intelligence announced on the same day that it had completed a 1 billion yuan B+ round, its fourth round in half a year; XPeng's robotics business completed its first round of more than $900 million in August.

XPeng Robotics

The mid-tier players are starting to lower expectations.

The old-share market is even more direct: according to a 36Kr report, old shares have already been offered for sale at a valuation of 12 billion yuan, but few investors are willing to participate in a capital increase at a valuation of 22 billion yuan.

Concentration is inherently the underlying nature of this business. R&D costs run into the hundreds of millions, and talent, supply chain, scenarios and capital are all indispensable; only a minority of teams are destined to make it in.

When things are hot, mid-tier players can still get a share; once money tightens, the sector retreats faster into a game for the few.

The secondary market is also revaluing primary-market prices.

Mech-Mind listed on September 1 at an issue price of HK$101.7, broke below its issue price on the first day, and then fell as low as HK$78.05, about 23% below the issue price. After falling from its first-day high, Unitree's market value rebounded to about 200.6 billion yuan on October 19, still more than 240 billion yuan below its peak. UBTech's share price once fell from a high of HK$328 to a low of HK$40.8, a maximum drawdown of about 87%.

Image source: Fin$cope

For a group of companies still raising money at a valuation of 20 billion yuan, an IPO may no longer mean an honorable exit; it may also mean repricing.

Huang Lichong, president of Huisheng International Capital, put it bluntly: a robotics company obtaining a post-money valuation of 50 billion yuan does not mean the secondary market is willing to take over at the same market value.

The first to fall

Before winter arrived, some had already gotten frostbite. The latest case did not even survive this summer.

Nanjing Wujing Dynamics was founded only on February 11 this year, making industrial-grade quadruped robots and selling robot dogs to power, petrochemical, and energy mining industries. In June, the company disclosed that just four months after its founding it had signed nearly 500 million yuan in framework orders with several central state-owned enterprises, plus another nearly 500 million yuan in local reserve orders, and announced that 110 million yuan in financing had been secured.

Its core team of about 20 people mostly has backgrounds at Tsinghua, Zhejiang University and Shanghai Jiao Tong University, with some members from Zhijiang Lab, AgiBot and Robot Era. In September, it released the Panshi P25-Ex explosion-proof wheeled-legged robot, claiming it reached the IIC T6 explosion-proof rating; another quadruped robot for substation inspection was originally scheduled for delivery in October.

Wujing Dynamics quadruped robot

On October 23, employees told the media that the company began owing wages in August, stopped contributing to the housing provident fund in September, and stopped paying social insurance in October; three to four dozen employees had been owed nearly 8 million yuan in total, the company had halted work, and some employees had already applied for labor arbitration. Both the employees and CEO Yuan Haihui said they could not reach the actual controller, Cui Xiaofei; reporters called Cui Xiaofei multiple times but could not get through.

The orders and financing the company previously disclosed now all need further verification.

Of the nearly 1 billion yuan in orders, half are framework orders and half are local reserve orders, and there are still multiple steps before formal procurement, delivery and payment collection. The actual receipt and use of the 110 million yuan in financing also cannot be confirmed from public information alone.

A person in charge of an embodied intelligence company told Pencil Road that there is a long distance between signing orders and actual procurement and delivery: "If you lack a sense of crisis about costs and commercial operations, even hundreds of millions of yuan in financing can be burned through quickly."

The ones that fell earlier were the previous batch of star companies.

On October 19, the Shenzhen Intermediate People's Court held the first creditors' meeting for Yinghe Robotics. The company was founded in March 2020, incubated by He Jianfeng, son of Midea founder He Xiangjian; ByteDance's strategic investment arm, Meituan's strategic investment arm, GSR Ventures and Panda Capital were all on the shareholder list; it reached a valuation of $500 million in ten months and raised more than 600 million yuan in total.

At its peak, Yinghe ran five business lines at the same time.

In March 2021, it signed an agreement with Foshan Shunde to invest 10 billion yuan to build an R&D and industrial base, less than a month from negotiation to signing.

The key factor that crushed it was delivery and payment collection.

For the Foshan Lighting bulb intelligent manufacturing workshop project undertaken in 2021, final acceptance was still not passed two years past the deadline, and the court's final judgment ordered Yinghe to return 26.904 million yuan and dismantle the equipment at its own expense. Its equipment was judicially auctioned with an appraised value of about 49.74 million yuan; it was put up for auction twice, in April and May this year, and both auctions failed.

The last operating data it left behind was: 2025 revenue of 4.33 million yuan and a net loss of 67 million yuan. In June this year, Li Lun, founding partner of shareholder Panda Capital, posted a video questioning: the company has no stable cash flow, some core executives have after-tax annual salaries exceeding 3 million yuan, and the renovation of a 2,600-square-meter office cost 28 million yuan.

Yinghe is not an isolated case.

Shenlan Robotics was accepted for bankruptcy liquidation by a Shanghai court in May; its parent company Shenlan Technology once claimed a valuation of 14 billion yuan. The five batches of employee claims published by the court total more than 23 million yuan, involving 102 people, and include unpaid wages, economic compensation and housing provident fund, among other items, so they cannot simply all be classified as unpaid wages.

Shenlan's most "viral" product, the Panda smart bus

Zhicheng Dynamics received investment from Li Zexiang's Qingshuiwan Fund and Ecovacs; its pool-cleaning robots have cumulative sales of nearly 100 million yuan, yet it fell into a cash flow crisis due to difficulties in collecting accounts receivable and other reasons. In September, it was petitioned for bankruptcy reorganization by creditors and subsequently entered pre-reorganization proceedings; the company said at the time that it was still operating normally. This is a different legal status from having been ruled into bankruptcy liquidation.

Datao Robotics was once valued at more than 20 billion yuan and was ruled to be accepted for bankruptcy liquidation in 2025; Silicon Valley's K-Scale Labs disbanded after its accounts were left with only about $400,000, issuing refunds and open-sourcing its technology; Longhui Medical, Chuangze Intelligent, Lesheng Robotics, Guangdong Hongda Xingye Robotics and other companies also entered bankruptcy proceedings of varying degrees this year...

Based on public reports and court announcements, TMTPost statistics show that in the 19 months from February 2025 to September 2026, more than 20 robotics companies underwent bankruptcy, dissolution, liquidation or shutdown of business lines, covering at least 11 tracks.

It should be emphasized that bankruptcy review, pre-reorganization, bankruptcy liquidation and voluntary dissolution are not equivalent, but they all point to one thing: after the financing window tightened, delivery, payment collection and cash flow began to face the same stress test.

Liu Qigui, a professor at the Business School of Ningbo University, believes this is not the accidental death of a single company but the cyclical clearing of an investment cycle. A person close to Yinghe's Series A financing recalled that the company's valuation had already reached $500 million before its product was launched. "Crowding in doesn't necessarily mean believing how big the market is; it means believing someone will take over later."

Why the pessimism: robots still can't get off the bench

Why did these companies fail?

On the surface, the money ran out; looking deeper, there is only one common internal cause: capital has entered on a large scale, but robots have not yet entered real production scenarios on a large scale.

This awkward fact was briefly brought to the table by a debate in September, when Shao Tianlan, founder of Mech-Mind, fired off consecutive posts on WeChat Moments. The argument pushed an industry question to the fore: when robots have not yet formed sufficient end demand, how much of the revenue provided by data collection centers, joint ventures and industrial funds can actually be converted into sustainable market revenue?

According to Caijing, some data collection vendors have participated in transactions where contracts were signed but physical delivery was not completed; some local investment promotion officials have even described some data collection projects as a "new type of Ponzi scheme."

These industry phenomena need to be verified item by item, and certainly cannot be directly equated with the illegal acts of any one company.

Li Feifei said something very calm in "Zhao Hejuan Talk" last December: robotics research is still in its early stage, and robots really lack data. The automotive industry has run for decades and accumulated massive amounts of real driving data; robots lack mature commercial applications in everyday scenarios, so valid data is difficult to collect at scale. If this problem is not solved, research will find it hard to move beyond the early stage.

The demand structure provides another set of evidence.

Statistics from the Humanoid Robot Scenario Application Alliance show that in the first half of 2026, more than 218 publicly disclosed bidding projects for embodied intelligence and humanoid robots in China were won, with contract amounts exceeding 1.723 billion yuan; educational and research institutions accounted for 55.5%, government and public platforms 20.6%, and industrial and technology enterprises 21.1%.

Image source: Humanoid Robot Scenario Application Alliance

Unitree's prospectus and responses to inquiries show that in the first three quarters of 2025, scientific research and education accounted for 73.6% of its humanoid robot revenue, commercial consumption 17.39%, and industry applications 9.01%; revenue actually used in scenarios such as intelligent manufacturing and intelligent inspection was 15.702 million yuan, accounting for 2.64% of total humanoid robot revenue.

The two sets of data correspond to different statistical scopes and cannot be spliced together, but they point in a similar direction: at this stage, the main buyers of robots are still not large-scale production lines.

Practitioners' feelings are colder than the data.

Before resigning, Shi Xiaoxin, the No. 1 product lead for XPeng Robotics, asked a question that the industry has repeatedly quoted: "If 10,000 robots are built, where do you put them?" Wang Peng, who works on embodied pretraining, judges that domestic embodied intelligence has entered a technology plateau period, and no new paradigm-level breakthrough has been seen in the short term. His suggestion is to first advance deployment and accumulate scenario data, preserving resilience for the next shock from large models.

Wang Xingxing's changing wording was recorded by Chaoyong AI in "Wang Xingxing's Remarks Today Reveal That Embodied Intelligence Faces Two Hard Years": in August 2025, he said the ChatGPT moment for embodied intelligence would come "if things go smoothly, in one or two or three years"; on August 20, 2026, the second day of listing, it became "two to three years at the fastest, five to ten years at the slowest"; on October 12 at the Industry Expo, he said "possibly even within one or two years at the fastest"; on October 22 at the World Physical AI Conference, it returned to "one to two years or two to three years," and he said signs of a breakthrough had already been seen.

2026 World Physical AI Conference and Unitree Partner Ecosystem Conference

The timetable keeps changing with progress in large models, but the bottleneck has not disappeared.

He laid out the sticking point very clearly: for embodied intelligence to pass the ChatGPT moment, the core challenge is aligning the AI model's inputs and outputs with real robots. "Large models can plan and navigate, but they have not solved the problems that arise after physical contact." On the same day, he also admitted that the biggest mistake of the past decade or so was not trusting AI enough.

Postures for surviving the winter

Contraction is the most common posture.

According to Chaoyong AI, a few leading companies have begun cutting pay and freezing hiring, shifting headcount from expansion to defense, and word has even begun circulating in the market of "star companies' old shares being dumped at a 50% discount," while more companies are cutting business lines and shrinking teams, reserving limited cash for core directions.

Unitree's cost control extends down to the reimbursement level: "Unitree employee says reimbursements over 100 yuan require Wang Xingxing's approval" became a trending topic.

At WAIC in July, Beihang University professor Wang Tianmiao's advice to entrepreneurs was already very specific: let go of fixation on valuation, narrow the battlefield, take small quick steps, focus on cash flow, and survive first.

Another posture is stockpiling provisions.

In June this year, Yuanli Lingji completed a new round of financing, with investors including large-model companies such as Alibaba, Zhipu, SenseTime, and StepFun; the same month, through an equity merger, it merged with Atomix, the logistics robotics company that Tang Wenbin brought out of the Megvii system; Atomix ranks second globally in sales of four-way pallet shuttles, with customers including Uniqlo, Mixue Bingcheng, and CATL, and annual revenue of nearly 1 billion yuan.

Yuanli Lingji merges with Atomix

Co-founder Fan Haoqiang said in an interview in May this year: "There is no AI bubble, but there is an OpenAI bubble."

He believes the industry will go through three rounds of screening, and the first threshold is capital and valuation. Expressing confidence in technology while preparing cash and scenarios in advance is not contradictory. Tang Wenbin and Fan Haoqiang both came from Megvii, have seen the previous AI ebb, and know that confidence cannot replace cash flow.

Looking further out, the cold is synchronized.

On October 15, in McGregor, Texas, at the SpaceX launch event, the second-generation Roadster completed a ramp run, an aerial flip, and a landing; the driver's-side door opened, and sitting in it was Optimus. It did not drive; the entire vehicle stunt was completed remotely by preset programs, and Optimus was only part of the demonstration. In his opening remarks, Musk said a real person needed to be present to prove that everything that followed was not AI or CGI.

Optimus's mass production has also not caught up with the original timetable.

The deployment target of 5,000 units set for 2025 saw actual output stall at several hundred units; for the production line converted at the Fremont factory, external sales will not begin until the second half of 2027 at the earliest.

Optimus

OpenAI has confirmed it is advancing a humanoid robot project, with R&D focused on robot intelligence systems. Altman expects to launch a demonstration prototype in 2027, but true public practical application will still take years.

At the same time, hardware supply capacity is improving.

Along Shenzhen Nanshan's "Robot Valley," about 80% of components can be sourced locally. Eight AgiBot Genie G2 units ran on the line at Longcheer's Nanchang factory for six days, 10 hours a day, with a full live stream, completing 64,828 operations, with an officially disclosed success rate of 99.99%. Body manufacturing can still generate profit, but the way of making profit is increasingly close to manufacturing: cost, capacity, yield, and delivery can all be calculated item by item.

What is truly expensive and scarce is still the "brain" that can work across scenarios. It does not have an equally mature supply chain, nor a unified, verifiable basis for pricing.

In the price of the brain, expectation still outweighs fact.

But this distance is longer than many people originally imagined.

From the influx of capital to robots replacing labor at scale, there are years in between. These few years are the winter the industry must cross.

This is exactly the real meaning of the current scramble for money and contraction: it is not that the industry is finished, but that the industry has begun to squeeze out the water.

The money has not left robotics either; it has only changed its evaluation method.

Chaoyong AI reported in "Embodied Money Begins to Go 'Up to the Sky and Into the Sea'" that after Zhu Xiaohu exited humanoid robot companies in batches, he turned to investing in underwater robots; Xianyin Weilai, which was founded just over a year ago and completed three rounds of financing in four months, is in the business of repairing satellites in space and refueling satellites with propellant.

Capital preferences are changing: from long-term narratives whose returns are hard to calculate to scenarios with clearer customers, costs, and collections.

Of course, not everyone can endure it; a large number of companies will die, perhaps even most of them.

Wang Huadong has warned of two historical patterns: after a representative company goes public, within at most one year, the industry begins to have problems; he also used the EV industry as an analogy: the total market value of all EV startups after listing is lower than the money raised by the entire industry, and industry investors as a whole are losing money.

Whether these two patterns will repeat in embodied intelligence is the most important question to watch in the coming year.

The Ministry of Industry and Information Technology's humanoid robot full-lifecycle platform has assigned codes to more than 28,000 robots, covering R&D, production, market access, sales, use, maintenance, scrapping, and recycling. It means full-lifecycle management has begun to be implemented, but it still cannot be concluded from this that an "exit mechanism" for enterprises has been established.

Meanwhile, Unitree's first financial report after listing has also been published: net profit attributable to shareholders in the first three quarters was 387 million yuan, a surge of 270.54% year on year, which became the main promotional theme, but the market cannot ignore that the company's Q3 net profit was 113 million yuan, while Q2 net profit was 224 million yuan, a quarter-on-quarter decline of 49% in net profit; revenue maintained high growth, but short-term profit fell sharply. Regardless of whether there is an issue of a heavier R&D ratio, it also shows that market uncertainty is still playing out fiercely, and its stock fell more than 6% at the open the next day.

Unitree's financial performance needs to be watched continuously, and there are four more signals worth monitoring next:

Monthly financing data starting in November, to test the depth of the cooling;

When the IPO window returns to normal, to test the regulator's attitude;

When the second-tier players below Unitree, valued at over 50 billion and 60 billion yuan, will open their prospectuses, and what real numbers those prospectuses contain;

The prices at which leading companies acquire mid-tier assets, to test the real price tag of an embodied company in this winter.

More important than all four signals is still one specific scenario: robots truly entering large-scale commercial scenarios and continuously working, rather than only boxing and dancing on stage. That day will be the starting point of the next wave.

What Li Feifei repeatedly emphasizes is still data, and the commercial scenarios that can continuously generate data.

For embodied intelligence companies, valuation must ultimately return to repeat purchases, delivery, and production-line data; companies that stay at the table must also answer the real account of "what is the gross margin of a single scenario."

Whoever can first make robots create sustainable value in real scenarios will be the first to get a ticket to the next spring.