A single marketing spend nearly pushed one embodied AI company to the brink.
An investor said a well-known company in the sector paid RMB 100 million in cash and equity to a gala event for only a few minutes of exposure, without notifying the board or shareholders or obtaining their consent. In the investor’s telling, the result was poor: 「It did not show the value or advantages of the product. People who watched the gala felt nothing, and some even mocked it.」
The damage went beyond optics. The spending reduced the company’s R&D ratio enough to trigger a hard threshold under Hong Kong listing rules. Under the Hong Kong Stock Exchange’s requirements on R&D intensity, the company may need to re-queue before making another run at the capital market.
After the event, several investors gathered in the founder’s office and pressed for answers. One of them was upset enough to slam the table, according to the report.
At the other end of the spectrum, not spending can be just as costly. Last autumn, one embodied AI company had more than RMB 100 million on its balance sheet, but the founder insisted on tight cost control and refused to let the hardware team travel to supervise production lines. More than 100 production robots were accepted remotely by video. After a prospective investor learned that, the financing collapsed.
The report presents these incidents not as isolated mistakes, but as signs of a broader backlash now surfacing in a sector that has raised money at breakneck speed.
Capital poured in faster than operating discipline
By incomplete estimates cited in the report, China’s embodied AI sector raised about RMB 43.8 billion in the first half of this year. Data from IT Juzi showed 288 financing events involving 226 companies in the same period, with disclosed funding of more than RMB 46 billion. Over a longer window from July 2025 to June 2026, that expanded to 503 deals and more than RMB 96 billion.
That pace works out to more than one financing event a day on average. The sector, often described as the ultimate carrier for AI, has been absorbing capital at unusual speed. In the first half of 2026 alone, there were more than 25 deals worth RMB 1 billion or more. Among them, Tashi Zhihang raised $455 million in a single round in April, setting a record for embodied AI financing in China.
The result is a funding profile that was almost unimaginable in earlier internet startup cycles in China: companies with only a few hundred employees can be sitting on several hundred million yuan in cash.
That has changed the core management question. For many founders, raising money is no longer the hardest part. Turning that money into actual progress has become more difficult.
The report says the sector is packed with scientists and engineers who know how to rebuild world models in three-dimensional space, but have limited experience in capital operations and even less experience deciding how to spend very large sums. Investors, despite their reputation for financial discipline, also lack mature benchmarks in a category this new.
Spending too little can be as dangerous as spending too much
Embodied AI founder Liu Yuan described a dinner conversation that shifted from technology and products to financing and cash reserves. He said his company had enough money on the books to last 50 months. Someone at the same table responded that their own company had enough cash for at least 100 months.
In another sector, those numbers would signal comfort. In embodied AI, they do not necessarily translate into safety.
Liu later heard of more than one company that used returns from wealth management products to fund R&D and keep teams running after closing financing. As long as headcount stopped expanding quickly, the principal could remain almost untouched for a long time.
That dynamic has particularly benefited lightweight companies focused on brain algorithms and conceptual deployment.
One often-cited example in the report is Xinghaitu, described as a leading embodied AI brain company. People familiar with the matter said the company raised more than RMB 4 billion in two years, setting an early financing record for the sector, yet spent less than RMB 100 million over the same period. That left it with abundant reserves and, in the report’s framing, very low capital utilization.
Industry insiders quoted in the piece said Xinghaitu runs extremely tight budget controls, compressing marketing, hiring and operating expenses to the limit and having now basically reached breakeven. One person described the approach as so strict that even office printing paper use was heavily restricted. Inside the industry, that style has been jokingly called a strategy of outlasting everyone else.
Still, over-saving can create a different failure mode. A former employee at another embodied AI company said the issue that nearly crippled his former employer was not a major investment mistake, but tens of thousands of yuan saved each year by cutting cloud redundancy services and off-site backup. Management kept only a single local server to store scenario programs and localization maps for more than 1,000 robots. A power outage in the company’s industrial park damaged the hard drive and wiped out all the data. Customized programs built for government and enterprise customers were not archived either. The company eventually paid heavily and absorbed a large operational cost just to get through the fallout.
Unitree is another example the report highlights. According to Unitree’s prospectus, its gross margin exceeded 60% in 2025. While many embodied AI companies were still heavily cash-burning, Unitree had already shown profitability and a high gross margin.
Even so, the company remained frugal. The report says it does almost no PR, and its R&D spending was only a little above RMB 90 million. That number, the piece notes, would not even cover a year of compute costs for some AI startups.
This is a sharp contrast with the logic that defined China’s internet boom years, when fundraising was often followed by rapid spending on marketing and competition for scale. Multiple founders told the publication that embodied AI companies are now using cash reserves as proof of safety, turning cash burn discipline into a new survival metric.
The complication is obvious. In a sector where technical paths are still unsettled and products still require repeated validation, money staying on the balance sheet can mean either discipline or an inability to identify where it should be deployed. Financial statements alone do not answer that question.
One immediate consequence is suspicion. If an embodied AI company is not spending, outsiders quickly ask how much R&D it is really doing.
The bigger cost often appears later. A management team at one high-profile startup reportedly made extending cash runway its top operating target. It froze multiple R&D roles, cut the number of component validation batches and postponed planned data collection. The financial effect was immediate: monthly spending fell and the company gained almost an extra year of runway on paper. But months later, several engineers had been poached by rivals, and a product that should have entered small-batch delivery was still unable to settle into mass production because of compatibility problems in core components.
The report says that in many embodied AI projects, the same chain keeps appearing. Companies first reduce testing rounds, then delay supply-chain validation, then freeze R&D hiring. Each decision can be defended on its own. Taken together, they leave products stuck at the prototype stage. One hardware executive called it 「a shutdown with money in the bank.」
There are also companies that spend freely in the wrong places. Liu cited one humanoid robot company from two years ago that, after raising money, did not first polish deployable scenarios. Instead, it launched multiple product lines at once and spent heavily on business development and government-enterprise reception. More troubling, the company built a so-called logistics scenario in the office using shelves and cardboard boxes, drew a few detection boxes with open-source models, and claimed multimodal perception capability. Its “automatic sorting” was actually controlled manually with wires hidden beneath the desk.
In Liu’s words, it was basically a packaged demo used to raise money and tell a story.
The report also points to an overseas cautionary tale: Vicarious Surgical. The company reportedly raised more than $300 million from investors including Bill Gates, and at one point had a market value above $1.2 billion. But it spent heavily on a highly complex in-house bionic robotic arm and abandoned mature commercial options. That decision stretched out the R&D cycle, sent iteration costs spiraling, and left the product unable to reach stable mass production. In its two most critical years, the company lost more than $100 million and ultimately entered liquidation.
The cost base is real, but the accounting still looks fuzzy
To understand why the sector is so conflicted about spending, the report argues, it helps to look at the underlying economics. That is also where many founders still appear least prepared.
One person with long-term exposure to embodied AI companies described a simple approach: open the company’s hiring page, count algorithm, hardware and engineering roles, estimate labor costs based on location, team size and market pay, then look at model releases, robot shipments and supply-chain activity to back out likely spending on compute, data and hardware.
When he put the numbers together, he found a visible gap between many companies’ projected cash burn and the costs they were likely to face in reality.
Labor is the first unavoidable line item. Embodied AI is deeply interdisciplinary and especially hungry for algorithm talent. In some core cities, monthly pay for ordinary embodied AI algorithm engineers has already reached about RMB 50,000. Some hiring data put the average closer to RMB 63,000. Once bonuses, social insurance, housing fund contributions and other employment costs are included, the annual cost of one mature algorithm engineer can approach RMB 1 million. Senior specialists in reinforcement learning, world models, motion control and core hardware commonly command RMB 2 million to RMB 3 million a year.
In April 2026, UBTech announced a global search for a chief scientist for embodied intelligence with annual pay starting at RMB 15 million and reaching as high as RMB 124 million. ByteDance’s Volcano Engine was also hiring a senior operations algorithm expert in embodied intelligence at RMB 95,000 to RMB 120,000 a month. Even newly graduated top PhDs from labs at Tsinghua, Peking University, Fudan, Shanghai Jiao Tong University, Zhejiang University and Harbin Institute of Technology were said to be starting at RMB 600,000 or RMB 700,000 and above. Yu Hongxiang, industrial application technology chief engineer at the Zhejiang Humanoid Robot Innovation Center, said some newly graduated junior colleagues could receive offers in the RMB 2 million to RMB 3 million range.
For a 200-person embodied AI company, with R&D staff making up about half the team, payroll alone would usually require at least RMB 100 million. If headcount rises, the total goes up sharply. One person familiar with industry compensation estimated that a 300-person embodied AI company would spend RMB 300 million a year on labor alone.
That is why, according to one investor quoted in the story, old industry claims that a top embodied AI brain company spent only RMB 100 million over two years may simply reflect incomplete accounting or incomplete disclosure. At 300 employees, he said, even cutting average annual pay in half still leaves labor costs at RMB 150 million a year.
Demand for talent is also far ahead of supply. Data from Zhilian Recruitment showed job postings in the robotics industry rose 38% in the first quarter. Growth rates for industrial robot engineers, robot algorithm engineers and robot commissioning engineers were 38%, 37% and 60%, respectively. Tight supply has pushed salaries even higher. One algorithm engineer who moved from a major internet company to an embodied AI unicorn said he received three separate recruiter approaches, and some startups were offering pay increases of as much as 150%.
Labor, however, is only the start. The lines that make the books hardest to read are compute and data.
Building a few VLA large-model demos can cost RMB 30 million to RMB 50 million a year, according to the report. Training world models can push annual compute costs above RMB 100 million. Buying and operating 100,000 A100 chips can easily exceed RMB 1 billion. One founder at a company committed to developing its own large models privately summed it up this way: for top companies trying to build both large models and hardware in-house, there is nowhere near enough money.
Skipping the embodied brain is not an easy way out either. The report says industry barriers in embodied hardware such as dexterous hands, flexible materials and joints are gradually nearing a ceiling, while the brain has become the next key question for real commercialization. A company that does not build on the brain side may struggle to tell investors a compelling story and later face a serious risk of elimination when embodied brain technology matures.
Some companies, including Unitree, are said to be taking the view that they can wait and acquire mature brain technology later. But the report says most embodied AI companies do not have that luxury and are still choosing to build the brain themselves.
Data is the other major cash sink. One widely circulated figure in the industry holds that a humanoid robot brain with general capability needs at least 1 million hours of high-quality data. A single data-collection robot costs about RMB 200,000, lasts roughly 1,000 hours, and requires labor costing around RMB 120 per hour. High-quality real-world robot data accounts for only about 20% of the total. That puts the effective cost of one usable hour at roughly RMB 1,600, implying about RMB 1.6 billion for 1 million hours.
Actual efficiency may be worse than the headline numbers suggest. In one non-public discussion, a founder at a leading company reportedly said buying and collecting 1 million hours of data would cost about RMB 100 million to RMB 200 million, while training those data would cost roughly ten times more.
Xu Qing, a pseudonymous industry insider cited in the article, said that out of 10,000 hours of real data collection, only tens to hundreds of hours often end up usable for model training. 「For a very complex task, it may be only a few dozen hours. For a somewhat simpler scenario, it may be 200 to 300 hours,」 he said. His conclusion was stark: tens of millions of yuan could be spent collecting 100,000 hours of data, only for model capability to improve by 5%.
That suggests more than 99% of collection costs can become sunk cost.
Zheng Sipeng, partner at Zhi Zai Wu Jie, gave another public estimate. He put the cost of 30 seconds of real-machine data collection at RMB 10 to RMB 15, or about RMB 1,000 per hour. On that basis, pretraining on 1 million hours of real-machine data would require investment on the order of RMB 1 billion.
The report adds another distortion: value distribution in the data collection chain is heavily inverted. Frontline data collectors are often paid hourly wages in the tens of yuan, while the data they generate may be sold to embodied AI companies at RMB 300 to RMB 500 per hour. The hardest part of the chain earns the least, while middle layers take the largest spread.
Then there is the robot body itself. A commonly cited estimate in the industry puts the bill of materials for a large humanoid robot at RMB 150,000 to RMB 200,000. If a company builds 500 prototypes or small-batch units, hardware materials alone can approach RMB 100 million. That figure excludes tooling, testing, rework, warehousing, after-sales support and units that fail acceptance.
Once money enters all those categories, it becomes difficult to fully explain where it went. One investor recalled that in other industries, revenue, inventory, customers and bank flows can at least be cross-checked to judge business conditions and decide when to expand or retreat. In embodied AI, the company’s core assets often remain stuck in the R&D process itself. Old methods stop working.
Investors are now asking harder questions about where the money went
The uncertainty is not limited to founders. It extends to the people funding them.
Some founders said that during financing in 2026, more investors began asking exactly how much the company had spent over the prior six months, and on what.
One investor said there is now a fairly standard way of answering those questions in the industry. Founders usually present clean percentages across categories. The problem is that, just like the broader cost structure, those numbers are hard to verify at a deeper level, leaving outside shareholders unable to judge with much precision.
One person familiar with the matter said investors in a top-tier company basically did not know where the money had gone at all. In his account, some investors write the check and stop there, especially smaller shareholders who have little access to the company’s real financial picture.
The report gives an example. A small company was close to completing a financing round of RMB 20 million to RMB 30 million, with due diligence nearly finished. Its finance lead accidentally sent the investor a different internal set of accounts. After rechecking the numbers, the investor immediately halted the financing.
No one outside the company can say whether the mistake was accidental or deliberate. What alarmed the investor was the possibility that, if the wrong email had never been sent, the gap between the two sets of numbers might never have been visible.
Why would a company building embodied large models be reluctant to fully disclose its spending? The report’s answer is straightforward. If too much cash appears to remain on the balance sheet, the next financing round may be harder. But if the company admits it has burned several hundred million yuan in a year, investors may conclude that the runway is only a matter of months and confidence may weaken just as quickly. In embodied AI, everyone needs a story about being able to survive for a long time.
That pressure is changing post-investment oversight. Since the start of this year, the report says, there have been repeated signs that strong investment institutions are placing staff inside portfolio companies. In the past, firms typically sent directors or observers to major meetings and listened to regular operating updates. They rarely intervened in day-to-day finance. Now, some investors are placing finance, audit and even anti-fraud personnel directly inside embodied AI companies to monitor procurement, reimbursements and related-party transactions on an ongoing basis.
In one more extreme case, a logistics-focused embodied AI company reportedly had a finance team in which a large share of personnel were dispatched by investors to work on-site. Nearly every expense required investor review and approval, even some office-supplies purchases, according to the company executive cited in the article.
One investor said the main concern is not only that founders may overspend, but that rapidly expanding financing rounds create new room for internal benefit transfers. Robot R&D touches chips, sensors, servers, components, data services and external testing. Supplier counts are high, pricing is inconsistent, and quotations for the same service can differ by multiples across companies. When there is no clear market price for technical procurement, investors struggle to distinguish a fair premium from benefit transfer. Some institutions have therefore started linking financial review with technical review.
The report says those fears are not hypothetical. It cites prior media coverage about a leading embodied AI company that brought in an investment bank and accounting firm for IPO financial coaching, only to see half of its reported revenue cut during audit because nearly half the revenue was low quality, consisting of related-party transactions and fragmented income streams.
Another warning sign is the appearance of circular data transactions. Some companies sell robots to data-collection centers, receive payment, and then buy data back from those same centers. The money circulates inside the industrial loop instead of creating clear new value.
The argument is shifting from whether to spend to who gets to decide
These changes have not made the founder-investor relationship any easier. The core argument has gradually shifted from whether money should be spent to who has the authority to decide how it should be spent.
During the most active years of internet entrepreneurship, capital and founders had a relatively mature playbook. Companies raised money, expanded teams, bought traffic, subsidized users and then used growth metrics to raise again. Even if they never turned profitable, investors could still judge results through new-user counts, retention, transaction volume and market share.
Embodied AI has no such reference system. Autonomous driving also went through a capital-intensive phase, but vehicle testing, road testing and mass-production milestones were comparatively clear. Large-model software companies also consume huge amounts of compute, but software products can reach users much faster. Embodied AI must simultaneously bear pressure from hardware, algorithms, data and scenario delivery. Any one of those can keep burning cash.
The report says some institutions are trying to replace simple annual budgets with stage-based goals. A company may be allowed to launch the next round of compute spending only after finishing one round of technical validation. A product enters larger-scale production only after hitting stability targets. Data-collection reviews now look not just at total hours, but also at usable-rate and model improvement.
That method can trim some ineffective spending. It still leaves major gaps. R&D in embodied AI is highly uncertain. One failed training run is not necessarily worthless. If investors reward only visible success, founders may drift toward projects that are easy to show and carry less risk.
Another hard reality is that not every investor has enough incentive to go deep. One interviewee said some small shareholders own only 2% to 3% and struggle to obtain complete operating information. Others focus more on the next financing round or secondary share sales than on sustained tracking of finance and R&D. If valuations keep rising, even investors in problematic companies may still exit through later transactions.
The report points to a deal involving Kepler Robotics. In May 2026, listed company Hangzhou Kelin announced a proposed acquisition of a 41.57% stake in Kepler Robotics for no more than RMB 300 million. The market later found that Kepler CEO Hu Debo had already formally stepped down in February 2026 and registered a new company, Sota Wujie, in April to pursue embodied AI brain R&D. Kepler also disclosed that Hu had in fact stopped serving as CEO as early as June 2025, remaining responsible only for sales and marketing, and that the company had canceled his equity incentive. A co-founder left on the eve of the sale, while the company’s valuation fell from RMB 1.06 billion six months earlier to RMB 720 million.
The report frames that episode as a concentrated example of the difficult bargaining between founding teams and capital.
That leaves the sector with a dangerous mismatch. As long as money keeps flowing in, many problems can be deferred. Once the pace of financing slows, the gap between products, revenue and cash burn may show up all at once.
One veteran investor said the real dividing line may start to appear in the second half of 2026. Leading companies are still likely to secure large rounds. Smaller firms that have not completed Series B financing and lack stable orders could see their survival space shrink quickly.
For now, capital does not seem ready to leave. One investor said he and peers made a rough estimate for second-half funding and stressed that it was only a back-of-the-envelope calculation, not a rigorous statistical model. With more long-term money entering China’s investment market, they think new capital in the second half could reach RMB 500 billion.
Embodied AI is still expected to remain a priority target for investors, who see China’s complete supply chain, engineering base and manufacturing capability as a rare combination that could put the country at the front of the field globally.
That means the sector may not cool immediately because of a few failed cases. More money may still enter, and companies that already raised may continue pushing the size of their next round higher.
The report’s central point is blunt: over the past few years, embodied AI solved the question of whether there would be money. The next question is harder. Once funding far beyond a company’s current operating capacity lands in its account, who decides the speed and direction in which that money should be spent?
The article was credited to Leo Zhang ToB Notes, written by Zhang Shenyu and edited by Yang Lin.