A single marketing push almost brought down a company.
An investor discovered that a well-known embodied AI company they had invested in paid a full 100 million RMB—in cash plus equity—to a TV gala show without reporting to or seeking approval from the board and shareholders. In return, they got mere minutes of exposure.
And the effect was less than ideal. "It didn't showcase the product's value and advantages. Viewers felt nothing, even ridiculed it," the investor said.
More seriously, this massive marketing expense directly lowered the company's R&D ratio, triggering a red line under Hong Kong Stock Exchange's strict requirements on R&D intensity. The company might have to go back to the end of the queue to knock on the capital market's door again.
In fact, after the gala, several investors gathered in the founder's office for a confrontation. Emotions ran high, and someone even slammed the table.
Sometimes, money sitting idle in the account is also a danger. Last fall, an embodied AI company, despite having hundreds of millions on its books, insisted on cutting costs. The founder refused to let the hardware team travel to oversee the production line. Hundreds of mass-produced robots were accepted remotely via video. When potential investors learned of this, the financing round fell through.
This is not just a few companies making occasional mistakes. After a frenzy of financing in the embodied AI sector, the first signs of backlash regarding money are emerging.
According to incomplete statistics, the domestic embodied AI sector raised about 43.8 billion RMB in the first half of this year. IT桔子 data shows that in the first half of the year, there were 288 financing events involving 226 companies, with disclosed funding exceeding 46 billion RMB. Extending the period from July 2025 to June 2026, this number swells to 503 financing events, totaling over 96 billion RMB.
On average, that's more than one financing event per day. This sector, seen as the "ultimate carrier of AI," is devouring capital at an unprecedented rate. Single transactions of 1 billion RMB or more exceeded 25 in the first half of 2026 alone. Among them, Tashi Zhihang secured $455 million in a single round in April, setting a new record for embodied AI financing in China.
Such intensive capital injection has spawned a phenomenon previously unseen in China's internet history—the embodied AI industry is likely the sector with the highest "disposable amount per capita" ever. A company with just over a hundred employees might have billions of RMB in cash on hand, something almost unimaginable in past internet startup cycles.
But the money came too fast, bringing new problems. The industry is crowded with scientists and engineers, most adept at reconstructing world models in three-dimensional space but rarely seasoned capital operators, let alone experienced in spending large sums. They find themselves in an equally complex but entirely unfamiliar environment: when lab thinking meets the floodwaters of the capital market, how to make money generate real value is becoming a tougher test than raising it.
And similarly lacking reference points are the investment institutions that were previously the best at accounting.
Spending money has become the most unfamiliar subject in the embodied AI track.
To Spend or Not to Spend
Behind the glory of financing lies the different survival strategies of embodied AI companies.
Embodied AI entrepreneur Liu Yuan once attended a dinner where, after a few drinks, the conversation drifted from technology and products to the same topic—financing, and who had more money left. He couldn't help revealing that his company had enough cash on its books to last 50 months.
"I thought I was very frugal and had a lot of money on hand," Liu Yuan said. But to his surprise, someone at the table immediately told him their company had at least enough cash for 100 months.
In other sectors, that would be an enviable number. But in the embodied AI industry, how much security these numbers actually buy is something even the entrepreneurs probably can't say.
Soon, Liu Yuan heard that more than one company in the industry, after securing funding, was using returns from financial products to fund R&D and sustain the team. As long as the team size doesn't expand rapidly, the principal on the books can remain largely untouched.
A group of lightweight companies focusing on "brain algorithms and conceptual landing" became the biggest beneficiaries of the capital boom and the most frugal players in the industry.
A frequently cited example is the industry-leading brain-focused embodied AI company Star Chart (Xinghai Tu). It raised over 4 billion RMB cumulatively in two years since its founding, setting an early-stage financing record for the sector. An insider claimed Star Chart spent less than 100 million RMB in total over two years, with extremely low capital utilization and very ample cash reserves.
But according to industry insiders, Star Chart implements extremely strict budget controls internally. Marketing, recruitment, and operational expenses are compressed to the extreme. It has reportedly achieved basic breakeven, "to the point where even office printing paper usage might be strictly limited." This model is jokingly called the "last-man-standing" strategy within the industry.
A former employee of an embodied AI company revealed that the "scourge" that nearly killed his previous company was saving a few hundred thousand RMB per year on redundant cloud service fees and off-site backup costs. Management kept only a single local server storing scene programs and positioning maps for thousands of robots. An unexpected campus power outage damaged the hard drive, and all data was lost. More critically, customized programs for government and enterprise clients weren't archived either. "The company ultimately spent a huge sum and paid a high price to barely pass the crisis."
A more well-known case is Unitree. Unitree's IPO prospectus shows that in 2025, its gross profit margin reached over 60%. While many embodied AI companies were "deep in debt" and burning money, Unitree had already achieved profitability with such a high margin.
Yet, even so, Unitree remains very frugal in spending. It hardly does any PR, and its R&D expenditure is only just over 90 million RMB—not even enough to cover a year's compute power costs for some AI startups.
This is completely different from the understanding of financing during the internet boom—taking money and spending it fast, pursuing scale through marketing and competition. A similar operating style can be traced back to the "Groupon Wars" period, where Meituan ultimately dragged down competitors with extremely low costs.
More than one entrepreneur told us that embodied AI companies are using cash reserves to prove their safety, which has become a new metric for measuring a company's viability. This appears to be an extremely prudent way of operating—for entrepreneurs who have gone through multiple capital cycles, controlling cash burn can hardly be called a mistake. But the question is, in an industry where the technical path is not yet determined and product capabilities still need continuous validation, does money sitting idle for a long time mean the company is sufficiently disciplined, or does it mean it hasn't found a direction worth investing in? It's hard to get a direct answer from financial numbers.
The most direct consequence is that not spending money in the embodied AI industry easily sparks speculation: How much R&D is this company actually doing?
This is also one of the hardest questions to answer in the industry. Money saved or spent is easily seen, but missed product windows rarely appear on financial statements in time. Many consequences only become apparent one or two years later.
The management of a star startup once set "extending the cash runway" as its most important operating goal. To reduce expenses, the company froze multiple R&D positions, reduced the number of component validation batches, and postponed planned data collection. From a financial perspective, these measures quickly showed results—monthly expenses dropped significantly, and the cash runway extended by nearly a year.
But months later, the company found many engineers had been poached by competitors. And a product that should have entered small-batch delivery was delayed from mass production qualification due to compatibility issues with core components.
Looking at the broader startup field, if cash flow tightness is due to insufficient financing, such relatively extreme measures are at least understandable. But in the embodied AI industry, companies are sitting on large piles of cash. In multiple embodied AI projects, a similar chain has repeatedly appeared: to control spending, companies first reduce testing rounds, then postpone supply chain validation, and then freeze R&D hiring. Each decision can be justified individually, but together, they can keep products stuck at the prototype stage.
A hardware head once called this state "affluent stagnation."
But where some save, others spend freely. Liu Yuan heard of a case two years ago where a humanoid robot company, after securing funding, didn't first polish a viable application scenario. Instead, it simultaneously launched multiple different product lines and spent a lot on business development and government/enterprise receptions. More seriously, the company built a "logistics scene" in the office using shelves and cardboard boxes, claimed multimodal perception capability by using open-source models to draw a few detection boxes, and its so-called automated sorting was actually manually controlled by someone hidden from view.
"This is basically using a packaged demo to raise funds and tell stories," Liu Yuan said.
Overseas, the once-starry embodied AI company Vicarious Surgical was also rumored to be a cautionary tale of reckless spending. It reportedly raised over $300 million from investors including Bill Gates, and its market value once exceeded $1.2 billion. But the company spent its money on persisting with self-developed, overly complex "bionic robotic arms," abandoning mature commercial solutions. This decision led to infinitely extended R&D cycles, runaway iteration costs, and products that never reached mass-production qualification.
In the end, Vicarious Surgical lost over $100 million in the most critical two years of its development and ultimately had to file for bankruptcy liquidation.
The Unclear Accounts
To understand the industry's mindset towards "spending," one must first grasp the real costs of this business. But how to account for embodied AI costs is precisely a lesson some entrepreneurs haven't yet learned.
A person long involved with embodied AI companies tried to estimate a company's real expenses in the simplest way for us.
Typically, you open the company's recruitment page, count the number of algorithm, hardware, and engineering positions, then calculate labor costs based on office location, team size, and market salaries. Next, you look at the company's model releases, robot shipments, and supply chain situation to back-calculate compute power, data, and hardware investment.
Adding these numbers together, he found a significant gap between many companies' projected cash burn and the actual costs likely to occur.
Labor is the first unavoidable expense. Embodied AI is a highly interdisciplinary field, with intense demand for algorithm talent. In some core cities, the monthly salary for an ordinary embodied AI algorithm engineer has reached around 50,000 RMB. Some recruitment data suggests an even higher average, close to 63,000 RMB. Including year-end bonuses, social security, housing fund, and other employment costs, the annual expenditure for a mature algorithm engineer easily approaches 1 million RMB. For senior talent in reinforcement learning, world models, motion control, and core hardware, annual salaries of 2 to 3 million RMB are not uncommon.
More tellingly, in April 2026, UBTECH announced a global recruitment for a "Chief Scientist of Embodied AI," offering a starting annual salary of 15 million RMB, potentially up to 124 million RMB. ByteDance's Volcano Engine is also recruiting a "Senior Operation Algorithm Expert (Embodied AI)" with a monthly salary of 95,000 to 120,000 RMB. Even top fresh PhD graduates from renowned labs at Tsinghua, Peking, Fudan, Jiaotong, Zhejiang University, Harbin Institute of Technology, etc., commonly command starting salaries of 600,000 or even 700,000 RMB. Yu Hongxiang, Chief Industrial Application Technology Engineer at the Zhejiang Humanoid Robot Innovation Center, revealed, "Some very outstanding recent graduates might get offers in the 2 to 3 million RMB range."
For an embodied AI company with around 200 people, R&D personnel typically make up half. Labor costs alone would require at least 100 million RMB—and that's just an average. If the team scales up, the number climbs accordingly. A person familiar with industry salary structures calculated for us that a 300-person embodied AI company would spend around 300 million RMB per year on labor alone.
"Previously, there was a rumor in the industry that a top brain company spent only 100 million RMB in two years. This might be unclear accounting or the entrepreneur's selective storytelling," the investor said. "For a 300-person scale, even halving the average annual salary, annual labor costs would still be 150 million RMB." The implication is that even with extreme hiring restraint, maintaining a sizable R&D team makes it hard to compress labor costs to a very low level.
Moreover, this talent demand far outstrips current supply. Zhaopin data shows that in Q1 this year, job postings in the robotics industry grew 38%, with positions for industrial robotics engineers, robotics algorithm engineers, and robotics debugging engineers growing 38%, 37%, and 60% respectively.
This supply-demand imbalance also drives up labor costs. An algorithm engineer who jumped from an internet giant to an embodied AI unicorn revealed he received three job offers from headhunters, with some startups offering salary increases up to 150%.
But labor costs are just the beginning. What truly blurs the accounts are the bottomless pits of compute power and data.
Making a few VLA (Vision-Language-Action) large model demos costs at least 30 to 50 million RMB per year. If you're building world models, annual compute power consumption hits the 100-million-RMB mark. Procuring and operating 100,000 A100 chips easily exceeds 1 billion RMB. "For a truly self-developed large model and hardware company, the money is far from enough," the founder of a company committed to self-developed large models privately lamented to us.
So, is not building an embodied brain the answer? From the current industry development, barriers in the embodied hardware—like dexterous hands, flexible materials, joints—are gradually peaking. The industry already views the embodied brain as key to future commercialization. Not developing the brain presents two problems: it becomes hard to tell a compelling story to investors, and companies face high risk of obsolescence when brain technology matures.
Of course, there are players like Unitree who plan to "pick the fruit" via M&A once embodied brain tech matures. But not all embodied AI companies can afford to wait like Unitree. Most choose to build their own "brain."
And data is another money-guzzling beast. An industry-common estimate is that achieving a general-purpose humanoid robot brain requires at least one million hours of high-quality data. A data collection robot costs about 200,000 RMB with a lifespan of about 1,000 hours. Labor costs about 120 RMB per hour, with high-quality real-world data accounting for only about 20%. A rough estimate puts the cost per effective data hour at about 1,600 RMB. One million hours of effective data implies an investment of about 1.6 billion RMB.
But reality is far harsher than numbers on paper.
A founder of a leading company once "slipped up" in a non-public exchange—buying and collecting data for one million hours probably costs 100 to 200 million RMB, and training on this data costs about ten times that. More anxiety-inducing is data quality. An industry insider, Xu Qing (pseudonym), revealed that out of 10,000 hours of real-world data collected, often only dozens to a few hundred hours are usable for model training.
"If it's a very complex task, maybe just dozens of hours. For slightly simpler scenarios, maybe 200 to 300 hours," he said. "Spending tens of millions to collect 100,000 hours of data might only improve model capability by five percent."
This means over 99% of collection costs could be sunk costs.
Zheng Sipeng, partner at Wisdom Without Borders (Zhizai Wujie), calculated a more detailed public estimate: 30 seconds of real-world data collection costs 10-15 RMB, accumulating to about 1,000 RMB per hour. Completing pre-training on one million hours of real-world data would require investment on the order of 1 billion RMB.
More astonishing is the severe value inversion in the data collection chain. Reportedly, most frontline data collectors earn only tens of RMB per hour, but the data they collect is sold to embodied AI companies for 300-500 RMB per hour. The hardest-working collection环节 gets the least, while the least value-adding middle环节 gets the most.
Beyond data, there's the hardware body.
The material cost for a large-size humanoid robot is commonly estimated at 150,000 to 200,000 RMB. Producing 500 prototypes or small-batch products could cost nearly 100 million RMB in hardware materials alone. This doesn't include mold opening, testing, rework, warehousing, after-sales, or products that fail acceptance.
Once money enters these环节, it's hard to fully account for. Adding all these numbers together likely pieces together a dilemma with no standard answer.
An investor recalled that in other industries, you could at least cross-check a company's operations through revenue, inventory, clients, and bank statements to decide when to ramp up spending or tighten the belt. What you got for the money spent was roughly predictable. For embodied AI companies, core assets often remain in the R&D process. Previously effective paths are now失效.
This also explains why some brilliant minds who secured huge funding collectively lose their judgment on spending. It seems the true cost of embodied AI (labor, compute, data) and whether there will be more major expenditures in the future still lack consensus in the industry.
Where Did the Money Actually Go?
When it comes to spending, it's not just some embodied AI entrepreneurs who are confused, but also the people who gave them the money.
Entrepreneurs found that during 2026 financing rounds, more and more investors began asking: How much exactly did the company spend in the past six months, and on what?
Getting a standard answer is relatively easy. "The industry basically has a fixed script for this," an investor said. In the entrepreneur's narrative, each item has a clear proportion. The problem is, just as costs are hard to clarify, these numbers are difficult to verify further down. External shareholders struggle to make precise judgments.
"Can you believe that investors in a certain leading company basically have no idea where the money went? Completely no idea," a person familiar with the situation revealed. "Some investors just put money in and don't follow up, especially some minority shareholders who can't access the company's real financials."
An example: a small company was about to close a 20-30 million RMB financing round, with the process nearing completion. The CFO mistakenly sent the potential investor another set of internal accounts. After re-checking, the investor stopped the financing.
It's unclear if this was a genuine mistake or intentional by the finance staff. What alarmed the investor was that if that email hadn't been sent wrong, they might never have seen the discrepancy between the two sets of numbers.
Why would a company developing embodied large models dare not admit how much it spent? The answer isn't complicated—if remaining cash is too ample, it might affect the next funding round. But admitting to burning through hundreds of millions a year means the money only lasts a few months, requiring constant fundraising, which would severely shake investor confidence.
In the embodied AI industry, everyone must tell a "I can survive a long time" story.
But such incidents prompted some investment firms to change their post-investment management approach. Starting this year, news kept coming from the industry—aggressive institutions began stationing personnel at portfolio companies. Previously, they typically sent directors or observers to important meetings, regularly listening to operational reports, rarely directly介入 daily finances. Now, some investors have sent finance, audit, and even anti-fraud personnel into embodied AI companies for long-term checks on procurement,报销, and related-party transactions.
There are even more extreme individual cases. For a logistics vertical embodied AI company, a large proportion of their finance team are personnel stationed by the investors to work on-site, and almost every expense needs investor confirmation and review, "even some office supply purchases need review by the investor's finance personnel before proceeding," the company's负责人 once told us.
An investor said their biggest worry isn't just entrepreneurs spending recklessly, but new利益 spaces emerging internally after rapid funding scale-up. Robotics R&D involves chips, sensors, servers, components, data services, and external testing, with many suppliers and缺乏统一标准的报价. The same service can vary in price severalfold between companies.
When technical procurement lacks公开市场 prices, it's hard for investors to distinguish reasonable premiums from利益输送. Hence, some began combining financial checks with technical reviews.
This concern isn't unfounded. Previously, media reported that during IPO preparations, a leading embodied AI company brought in investment banks and accounting firms for financial辅导. The audit slashed revenue by half, as nearly half was low-quality, comprising related-party transactions and碎片化收入. More alarmingly, "left-hand-to-right-hand" data transactions have emerged: companies sell robots to data collection centers, get paid, then turn around and buy data from those same centers.
Thus, money ultimately空转s within the industry闭环.
This change hasn't made relations between investors and entrepreneurs easier. The core of their debates gradually shifted from "whether to spend" to "who has the right to decide how to spend."
During the most active internet创业 period, capital and entrepreneurs developed a relatively mature合作 model. Companies raised money, expanded teams, bought traffic, subsidized users, then used growth metrics to secure the next round. Even without最终盈利, investors could judge if the money worked through new users, retention rates, transaction volume, and market share.
Embodied AI lacks such a reference system. Autonomous driving experienced a similar capital-intensive phase, but vehicles, road testing, and production milestones were relatively clear. Large model companies also needed巨额算力, but software products reached users faster. Embodied AI simultaneously bears pressure from hardware, algorithms, data, and场景交付, any of which can持续消耗 cash.
We learned some institutions are trying to replace simple annual budgets with stage goals. After completing a round of technical verification, a company can启动 the next batch of compute power investment. Only after products reach stability metrics do they enter larger-scale production. Additionally, beyond data collection hours, they also check effective rates and model improvement.
This method can reduce some无效开支 but still doesn't solve all problems. Embodied AI R&D has high uncertainty. A failed training run未必毫无价值. If investors only reward successful outcomes, entrepreneurs might偏向选择 easy-to-showcase, low-risk projects.
A more现实 problem: not all investors have enough motivation to dig deep.
An interviewee said some minority shareholders hold only 2-3%股份, struggling to get complete operational information. Others, after investing, care more about the next funding round and secondary share sales than持续追踪 company finances and R&D. If industry valuations keep rising, even with operational issues, early investors might exit via the next round.
For instance, in May 2026, listed company Hangzhou Kelin announced plans to acquire up to 41.57% of Kepler Robotics for不超过 300 million RMB. The market soon discovered that Kepler's CEO Hu Debo had officially left in February 2026 and registered a new company, "Suota Wujie," in April, diving into embodied AI "brain" R&D. More intriguingly, Kepler revealed Hu had stopped serving as CEO as early as June 2025,仅负责 sales and marketing, and the company canceled his equity incentives. A co-founder出走 right before acquisition, company valuation shrinking from 1.06 billion RMB six months prior to 720 million RMB—this "sell-off" drama is filled with unspoken博弈 between the founding team and capital.
This creates a dangerous misalignment in the embodied AI industry: as long as capital keeps flowing in, many problems can be postponed. Once fundraising slows, gaps between product, revenue, and cash burn will be exposed simultaneously.
A资深 investor判断 the real watershed might emerge in the second half of 2026. Leading companies can still secure large funds, while SMEs that haven't completed Series B and lack stable orders will see生存空间 shrink rapidly.
But capital shows no signs of wanting to leave yet.
An investor estimated with peers the funding scale for the second half. He特意强调 this was just a "back-of-the-envelope" calculation, not rigorous statistics—with more long-term capital entering China's investment market,新增资金 in H2 could reach 500 billion RMB.
Embodied AI will remain a key battleground for investors. In their view, China's embodied AI industry has a complete supply chain, engineering talent pool, and manufacturing capability, making it one of the few tech sectors with a chance to lead globally.
This means the industry likely won't cool immediately due to a few failures. More money will come in, and already-funded companies might raise even larger next rounds.
It seems the embodied AI industry spent the past few years solving the problem of having money. Next, it faces a harder question: when funds far exceeding a company's current operational capacity enter its account, who decides the speed and direction in which it should be spent.
As more money pours into the industry, this lesson on spending money will only become more urgent.
(By Leo Zhang from ToB Zatan, Author: Zhang Shenyu, Editor: Yang Lin)
