Embodied Intelligence Companies Struggle with Spending Amid Capital Influx

A single investment nearly brought down a company.

An investor discovered that a well-known company in the embodied AI industry, without notifying its board of directors or obtaining shareholder approval, paid a total of 100 million yuan in cash and equity to a single evening event—receiving only a few seconds of exposure in return.

And the results weren’t great—he said, “It didn’t showcase the product’s value or advantages; viewers of the event felt nothing, and some even mocked it.”

More seriously, this massive marketing expenditure directly reduced the company’s R&D ratio, triggering the red line under Hong Kong Stock Exchange listing rules—according to the exchange’s strict R&D intensity requirements, the company may have to reapply and rejoin the queue to re-enter the capital markets.

In fact, after the party, several investors gathered in the entrepreneur’s office and interrogated him; one became emotional and even slammed the table.

Sometimes, money sitting idle on the books can also be dangerous. Last autumn, a robotics company with hundreds of millions of yuan on its books had its founder insist on cutting costs and refused to allow the hardware team to travel to oversee production. As a result, over a hundred mass-produced robots were all inspected remotely via video. Once potential investors learned of this, the funding deal fell through.

This is not an isolated mishap by a few companies—it’s the first sign of financial backlash following the frenzied fundraising in the embodied AI sector.

According to incomplete statistics, domestic embodied AI startups raised approximately RMB 43.8 billion in the first half of this year. Data from IT Juzi shows that 288 funding events occurred in the first half of this year, involving 226 companies, with disclosed funding exceeding RMB 46 billion. If the statistical period is extended to July 2025 through June 2026, the number rises to 503 funding events and over RMB 96 billion.

On average, more than one funding round occurs per day. This sector, regarded as the "ultimate vehicle for AI," is absorbing capital at an unprecedented pace. In the first half of 2026, over 25 transactions exceeding RMB 1 billion each have been recorded. Among them, Yishi Zhihang secured $455 million in a single round in April, setting a new record for the largest funding round in China’s embodied AI space.

Such intense capital inflow has given rise to a phenomenon never before seen in China’s internet industry—the embodied intelligence sector, which may well be the highest "per capita disposable amount"赛道 in the history of China’s internet. A company with only a few hundred employees could have hundreds of millions of yuan in cash on its balance sheet, something nearly unimaginable in previous internet startup cycles.

But the rapid influx of money has brought new challenges. The industry has attracted a large group of scientists and engineers, most of whom excel at reconstructing world models in three-dimensional space, yet few have experience in capital operations or spending large sums. They find themselves in an equally complex but entirely unfamiliar environment: as laboratory thinking meets the flood of capital markets, generating real value from money has become a more difficult challenge than raising funds.

Similarly lacking a reference asset are the investment institutions that were previously the most calculating.

Spending money has become the most unfamiliar lesson in the embodied intelligence sector.

Behind the glamour of financing lie different ways embodied companies survive.

Earlier, embodied AI entrepreneur Liu Yuan attended a dinner party. As the evening progressed and the alcohol flowed, the conversation naturally shifted from technology and products to the same topic—funding and who had more money left. He couldn’t help but reveal that his company had enough cash on hand to last 50 months.

“I thought I was being very frugal and had plenty of money on paper,” but Liu Yuan didn’t expect that someone at the table immediately told him his company still had enough cash to last at least 100 months.

In other industries, this would be a number to envy, but in the field of embodied intelligence, just how much security these figures can truly buy—even entrepreneurs themselves aren’t sure.

Soon, Liu Yuan heard that more than one company in the industry, after securing funding, relied on returns from financial products to develop their products and sustain their teams. As long as the team size didn’t expand rapidly, the principal on paper could remain virtually untouched for the long term.

A group of lightweight companies emphasizing "brain algorithms and concept implementation" have become the biggest beneficiaries of the capital frenzy and the most frugal players in the industry.

A frequently cited example is Xinghaitu, a leading embodied AI company in the industry, which has raised over RMB 4 billion in total funding within two years, setting a record for early-stage financing in the sector. According to insiders, Xinghaitu’s total capital expenditure over the two-year period was less than RMB 1 billion, indicating extremely low capital utilization and substantial cash reserves on hand.

However, according to insiders, Star Map enforces extremely strict budget controls, slashing marketing, hiring, and operational expenses to the bare minimum, and has now nearly achieved break-even—so much so that even the use of office printing paper is tightly restricted. This approach has been jokingly dubbed the “outlast them” strategy within the industry.

A former employee of an embodied AI company revealed that the near-disaster for their former company stemmed from saving tens of thousands of dollars annually on redundant cloud services and offsite backup costs. Management retained only a single local server to store scenario programs and localization maps for thousands of robots. When an unexpected power outage occurred on the campus, the hard drive failed, resulting in the complete loss of all data. Crucially, customized programs developed for government and enterprise clients were not archived. “The company ultimately paid a heavy price to barely pull through.”

Another well-known example is Unitree. According to Unitree’s prospectus, in 2025, its gross profit margin exceeded 60%, while many other embodied AI companies were still burdened with debt and continuously burning cash, Unitree had already achieved profitability at such a high gross margin.

Even so, Unitree remains very frugal with its spending. It does almost no PR and has spent only around 90 million on R&D—an amount that doesn’t even cover the annual computing costs of some AI startups.

This is entirely different from the financing model during the internet boom, where companies raised funds and quickly spent them on large-scale marketing and competition. A similar business approach can be traced back to the “Hundred Group Battles” during the rise of group buying, when Meituan ultimately outlasted its competitors by operating at extremely low costs.

More than one entrepreneur has told us that embodied AI companies are using their cash reserves as proof of security, making this a new metric for assessing a company’s survival capability. This appears to be an extremely prudent approach—given that entrepreneurs who have weathered multiple capital cycles would find it hard to call cash conservation a mistake. But the issue is, in an industry where the technological path is still uncertain and product capabilities require ongoing validation, it’s difficult to determine from financial figures alone whether holding cash for long periods reflects disciplined restraint or a lack of viable investment opportunities.

The most immediate consequence is that the embodied AI industry spends nothing, easily sparking speculation: How much R&D has this company actually done?

This is also one of the hardest questions to answer in the embodied AI industry. Money saved or spent is easy to see, but missed product opportunities rarely appear on financial statements in a timely manner. Many consequences only become apparent one or two years later.

The management of a prominent startup once set "extending the cash runway" as its top operational priority. To reduce expenses, the company froze several R&D positions, reduced the volume of component validation batches, and postponed its planned data collection initiative. From a financial perspective, these measures quickly took effect—the company’s monthly expenditures dropped significantly, and its cash reserves were extended by nearly a year.

But months later, the company found that many engineers had been poached by competitors, and a product slated for small-scale delivery remained stuck in development due to compatibility issues with its core components.

In the startup world, extreme measures taken due to insufficient funding and cash flow constraints are somewhat understandable. However, the embodied AI industry is flush with substantial capital. Across multiple embodied AI projects, a similar pattern repeatedly emerges: to control expenses, companies first reduce testing cycles, then delay supply chain validation, and finally freeze R&D hiring. Individually, each decision may seem justifiable—but together, they cause products to remain stuck at the prototype stage.

A hardware lead once referred to this state as "rich stagnation."

But while some save, others freely spend. Liu Yuan heard of a case where, two years ago, a humanoid robotics company, after securing funding, did not first refine viable use cases but instead launched multiple different product lines simultaneously and spent heavily on business and government relations. More seriously, the company built a “logistics scenario” in its office using shelves and cardboard boxes, drew a few detection boxes with open-source models, and claimed to have multimodal perception capabilities. What was presented as automated sorting was in fact manually controlled by people hidden beneath desks, with control wires concealed out of sight.

“This is basically taking a packaged demo and going out to raise funds and tell a story,” said Liu Yuan.

The former U.S. celebrity embodied robotics company, Vicarious Surgical, was once cited as a cautionary tale of reckless spending. It raised over $300 million from investors including Bill Gates, and its market valuation once exceeded $1.2 billion. However, the company poured its funds into persistently developing an overly complex proprietary “bio-inspired robotic arm,” rejecting established commercial solutions. This decision led to indefinitely extended development cycles, runaway iteration costs, and an inability to finalize or mass-produce its product.

In the end, Vicarious Surgical lost over $100 million during its two most critical years of development and was forced into quiet bankruptcy liquidation.

To understand people’s mindset toward “spending,” you first need to grasp the true cost of this business. But calculating the finances of embodied AI is precisely a lesson some entrepreneurs haven’t yet learned.

An individual with long-term exposure to embodied AI companies attempts to estimate a company’s true expenses using the simplest possible method.

Typically, start by visiting the company’s careers page, tallying the number of roles in algorithms, hardware, and engineering, then calculate labor costs based on office locations, team sizes, and market salaries. Next, examine the company’s model releases, robot shipment volumes, and supply chain status to infer investments in computing power, data, and infrastructure.

After adding up several numbers, he found a significant gap between the cash burn rates predicted by many companies and the actual costs that might occur.

Labor is the first unavoidable cost. Embodied intelligence is a highly interdisciplinary field that demands an enormous number of algorithmic talents. In major cities, the monthly salary for a typical embodied intelligence algorithm engineer has reached around 50,000 RMB. Some recruitment data indicate an even higher average, approaching 63,000 RMB. When factoring in year-end bonuses, social insurance, housing fund, and other employment costs, the annual expenditure for a seasoned algorithm engineer easily approaches one million RMB. Senior professionals in areas such as reinforcement learning, world models, motion control, and core hardware commonly earn annual salaries of 2 to 3 million RMB.

More notably, in April 2026, UBTECH announced a global recruitment drive for a Chief Scientist in Embodied Intelligence, offering a starting annual salary of 15 million yuan, with a maximum of up to 124 million yuan. ByteDance’s Volcano Engine is also recruiting Senior Experts in Operational Algorithms (Embodied Intelligence), with monthly salaries ranging from 95,000 to 120,000 yuan. Even recent PhD graduates from top institutions such as Tsinghua, Peking University, Fudan, Shanghai Jiao Tong University, Zhejiang University, and Harbin Institute of Technology are commonly offered starting salaries of 600,000 to over 700,000 yuan. Yu Hongxiang, Chief Technical Officer for Industrial Applications at the Zhejiang Humanoid Robotics Innovation Center, revealed, “Some recent graduates are exceptionally talented and may receive offers in the range of 2 to 3 million yuan.”

For a embodied AI company with 200 employees, R&D staff typically make up about half the team, meaning personnel salaries alone would require at least 100 million yuan—this is just an average figure. As the team grows, the costs rise accordingly; according to someone familiar with industry compensation structures, a 300-person embodied AI company would spend as much as 300 million yuan annually on labor costs alone.

“Previously, there was a claim circulating in the industry that top brain companies spent only 100 million yuan over two years—this might reflect inaccurate accounting or entrepreneurs holding back on the full truth,” said the investor. “With a team of 300 people, even if salaries were halved, annual labor costs would still reach 150 million yuan.” The implication is that, even if a company is extremely conservative in hiring, as long as it maintains a substantial R&D team, labor costs are unlikely to be reduced to a sufficiently low level.

Moreover, demand for such talent far exceeds the current supply in the job market. According to data from Zhipin Recruitment, the number of job postings in the robotics industry increased by 38% in the first quarter of this year, with growth rates of 38%, 37%, and 60% for industrial robotics engineers, robotics algorithm engineers, and robotics calibration engineers, respectively.

The shortage of supply relative to demand has also driven up labor costs for embodied AI companies. One algorithm engineer who transitioned from a major internet company to an embodied AI unicorn revealed that he received job offers from three headhunters, with some startups offering salary increases of up to 150%.

But labor costs are just the beginning. What truly obscures the financial picture are the bottomless pits of computing power and data.

Developing a few VLA large models can cost at least $30 to $50 million per year; if you aim to build a world model, annual compute costs rise to hundreds of millions. The procurement and operational expenses for 100,000 A100 chips easily exceed $1 billion. “Even the top companies that truly develop their own large models and hardware find their funding far from sufficient,” confessed the founder of a company committed to in-house large model development.

So, should we just skip the embodied brain? From the current state of industry development, barriers in embodied hardware—such as dexterous hands, flexible materials, and joints—are nearing their limits. The industry now views the embodied brain as the critical factor determining whether embodied intelligence can achieve true commercialization. If companies ignore the embodied brain, they will struggle to present compelling narratives to investors; furthermore, as the embodied brain matures, they risk being rapidly outpaced and rendered obsolete.

Of course, some industry players, such as Unitree Technologies, plan to wait until embodied brain technology matures and then acquire companies to “pick the low-hanging fruit.” However, not all embodied intelligence companies can afford to wait as Unitree does; most have chosen to develop their own “brains.”

Data is another massive money sink. Industry consensus holds that achieving a general-purpose humanoid robot brain requires at least one million hours of high-quality data. A single data collection robot costs approximately 200,000 yuan, has a lifespan of about 1,000 hours, and incurs labor costs of roughly 120 yuan per hour, with high-quality real-robot data accounting for only around 20%. Roughly estimating, the cost per hour of effective data is about 1,600 yuan. One million hours of effective data translates to an investment of approximately 1.6 billion yuan.

But the reality is far harsher than the numbers on paper suggest.

The founder of a leading company once inadvertently revealed during a private exchange that acquiring and collecting data costs approximately 100 million to 200 million RMB for one million hours, while the cost of training on this data is roughly ten times higher. Even more concerning is data quality: according to an industry insider, Xu Qing (pseudonym), out of ten thousand hours of real-world data collected, only dozens to a few hundred hours are typically usable for model training.

If it’s a very complex task, it might take only a few dozen hours; if the scenario is slightly simpler, it could take two to three hundred hours.” He believes, “With millions spent to acquire a hundred thousand hours of data, the model’s capability increased by only five percent.”

This means that over ninety-nine percent of the mining costs are likely sunk costs.

Zheng Sipeng, partner at Zhi Zai Wu Jie, calculated a more detailed figure in a public setting: the cost of collecting real-device data for 30 seconds ranges from 10 to 15 yuan, resulting in an approximate cost of 1,000 yuan for one hour of real-device data; to complete pre-training with a million hours of real-device data, the investment would reach a scale of one billion yuan.

Even more startling is the severe misalignment in value distribution along the data collection chain. It is reported that most frontline data collectors earn only tens of yuan per hour, yet the data they gather is sold to embodied enterprises at rates of 300 to 500 yuan per hour. The most labor-intensive collection stage receives the least compensation, while the least value-adding intermediary环节 receives the most.

Beyond the data, there is ontology.

The material cost for a large humanoid robot is commonly estimated by the industry to be between 150,000 and 200,000 yuan. If a company produces 500 prototype units or a small batch, the hardware material costs alone could approach 100 million yuan. This estimate does not include expenses for mold development, testing, rework, warehousing, after-sales service, or products that fail quality inspection.

Once money enters these channels, it becomes difficult to fully account for it. When all these figures are added together, they likely point to a dilemma with no standard solution.

An investor recalled that in other industries, one can at least cross-verify a company’s operational status through revenue, inventory, customers, and bank statements to determine when to increase investment and when to scale back. The return on expenditures could generally be anticipated. However, the core assets of embodied AI companies often remain stuck in the R&D phase.

But embodied intelligence is entirely different; a company’s core assets often remain stuck in the R&D phase. Past proven approaches now no longer work.

This also explains why some highly funded smart individuals collectively lose their judgment when it comes to spending. To date, the industry remains divided on the actual final cost of embodied intelligence—spanning human resources, computing power, and data—and whether significant expenditures will continue in the future.

It’s not just some entrepreneurs in the embodied AI industry who are careless about spending money—so are the people investing in them.

Entrepreneurs have noticed that during fundraising in 2026, an increasing number of investors are asking how much the company spent over the past six months and where exactly those funds were allocated.

Obtaining a standard answer is relatively easy. “The industry generally has a consistent line on this,” said one investor, noting that in early founders’ statements, each figure had a clear proportion. But the problem is that, just as costs are unclear, these numbers are also difficult to verify further, making it hard for external shareholders to make precise judgments.

“A major shareholder of a leading company basically has no idea where the money went—can you believe that? They have absolutely no idea,” revealed someone familiar with the situation. “Some investors invest and then completely disengage, especially smaller shareholders who have no access to the company’s actual financial information.”

For example, a small company was on the verge of closing a funding round of tens of millions of yuan, with the investment process nearly complete. The finance manager accidentally sent an internal set of books to the investor instead of the correct documents. After reviewing the materials again, the investor halted the funding.

It cannot be confirmed externally whether this erroneous dispatch was a work mistake or intentionally done by the finance staff. What truly alerts investors is that if the email had not been sent in error, they might never have noticed the discrepancy between the two sets of numbers.

Why is a company developing embodied large models reluctant to disclose how much it has spent? The answer isn't complicated—if the remaining cash on hand appears too substantial, it could hurt its chances in the next funding round; but if it admits to burning hundreds of millions in a year, it signals that the funds will only last a few months, requiring constant fundraising, which would severely undermine investor confidence.

In the embodied intelligence industry, everyone must tell a story about "how I can survive for a long time."

However, similar incidents have prompted some investment firms to change their post-investment management approaches. Since the beginning of this year, rumors have occasionally surfaced within the industry—powerful investment institutions have begun sending personnel to reside at portfolio companies. In the past, firms typically appointed directors or observers to attend key meetings and periodically review operational reports, rarely intervening directly in day-to-day financial matters. Now, some investors have placed finance, audit, and even anti-fraud staff inside embodied AI companies to conduct ongoing reviews of procurement, expense reimbursements, and related-party transactions.

Among these, there are even more extreme cases. A logistics-focused embodied AI company revealed to us that a large portion of its finance team consists of personnel seconded by investors to work on-site, and nearly every expenditure requires approval from the investors—“even the purchase of office supplies must be reviewed and approved by the investors’ finance staff.”

An investor stated that their greatest concern is not merely entrepreneurs spending recklessly, but rather the emergence of new opportunities for personal gain within the company after rapid expansion of funding. Robotics development involves chips, sensors, servers, components, data services, and external testing, with a large number of suppliers and no standardized pricing. The same service can vary in price by several times across different companies.

When there is no public market price for technology procurement, investors struggle to distinguish between reasonable premiums and improper transfers of value. As a result, some investment institutions are beginning to combine financial audits with technical reviews.

This concern is not unfounded. Previous media reports revealed that a leading embodied AI company, while preparing for its IPO, brought in investment banks and accounting firms for financial guidance—only to have half of its revenue eliminated during auditing, due to low-quality revenue stemming from related-party transactions and fragmented income. Even more concerning is the emergence of “hand-to-hand” data trading within the industry: companies sell robots to data collection centers, collect payment, and then repurchase data from the same centers.

And so, the money ultimately circulates uselessly within the industrial loop.

This shift has not made the relationship between investors and entrepreneurs any easier. The core of their debate has gradually shifted from “whether to spend” to “who has the right to decide how to spend.”

During the most vibrant period of internet entrepreneurship, capital and entrepreneurs developed a relatively mature collaborative model: after securing funding, companies would expand their teams, purchase traffic, subsidize users, and then use growth metrics to secure the next round of financing. Even if the company never became profitable, investors could still assess whether their capital had been effective by evaluating metrics such as new user acquisition, retention rates, transaction volume, and market share.

Embodied intelligence lacks a similar reference framework. Autonomous driving once went through a similar capital-intensive phase, but vehicle development, road testing, and mass production milestones were relatively clear. Large model companies also require massive computing power, yet their software products can reach users more quickly. Embodied intelligence must simultaneously bear the pressure of hardware, algorithms, data, and scenario deployment—each of which could continuously consume cash.

We understand that some organizations are adopting phased goals instead of simple annual quotas. Companies can only proceed with the next round of computing power investment after completing a technical validation, and can only move to large-scale production once their product meets stability benchmarks. In addition to hours, data collection must also evaluate efficiency and model improvements.

This approach can reduce some inefficiencies, but it still cannot solve all problems. Research in embodied intelligence is highly uncertain; a failed training session is not necessarily without value. If investors only reward successful outcomes, entrepreneurs may be inclined to choose projects that are easier to demonstrate and carry lower risk.

A more realistic issue is that not all investors have sufficient motivation to delve deeply into the company.

One respondent said that some minority shareholders hold only 2% to 3% of the equity and find it difficult to obtain comprehensive operational information. Additionally, some investors, after completing their investment, are more focused on the next funding round and secondary share transfers rather than continuously monitoring the company’s financials and R&D. If industry valuations continue to rise, early investors may exit through a subsequent round of financing, even if the company faces operational challenges.

For example, in May 2026, the listed company Hangzhou Kelin announced its intention to acquire up to 41.57% of Kepler Robotics for no more than RMB 300 million. However, the market soon discovered that Kepler’s CEO, Hu Debo, had officially stepped down in February 2026 and registered a new company, “Sota Infinite,” in April to focus on developing “brains” for embodied intelligence. More intriguingly, Kepler revealed that Hu had ceased serving as CEO as early as June 2025, retaining only responsibilities for sales and marketing, while the company also revoked his equity incentives. With a co-founder departing just before the acquisition, the company’s valuation plummeted from RMB 1.06 billion six months prior to RMB 720 million—this “sale” saga is rife with unspoken博弈 between the founding team and investors.

This has created a dangerous misalignment in the embodied AI industry: as long as funding continues to flow, many issues can be postponed, but once the pace of financing slows, the gap between products, revenue, and cash burn will be exposed simultaneously.

A seasoned investor believes the true turning point may begin in the second half of 2026. Leading companies will still secure substantial funding, while small and medium-sized enterprises that have not completed their Series B financing or secured stable orders will see their survival space rapidly shrink.

But there are no signs that capital wants to leave for now.

An investor and industry peer estimated the capital scale for the second half of the year. He specifically emphasized that this was merely a rough estimate, not a rigorous statistical analysis—given the influx of more long-term capital into the domestic investment market, new funds in the second half could reach 500 billion yuan.

Embodied intelligence will remain a key focus for investment institutions, as they believe China’s embodied intelligence industry possesses a complete supply chain, engineering talent pool, and manufacturing capabilities—making it one of the few technology sectors with the potential to reach global leadership.

This means the industry may not immediately cool down due to a few high-profile failures. More capital will likely continue to flow in, and companies that have already raised funds may still push for larger next-round valuations.

Today, the embodied intelligence industry has spent the past few years addressing the question of whether it has funding. Next, it must confront a more difficult question: when funds far exceeding the company’s current operational capacity enter the account, who decides how quickly and in what direction they should be spent?

As more money flows into the industry, the lesson of spending wisely becomes even more urgent.

(Text by Leo Zhang’s B2B Musings, Author: Zhang Shenyu, Editor: Yang Lin)