Reading listed companies' financial reports to calculate the operating economics of embodied intelligence

+1

Robots have already begun generating revenue, but companies have not necessarily made money. The prospectus disclosed by Unitree Robotics in July 2026 shows that the company's 2025 revenue was 1.699 billion yuan, with net profit attributable to shareholders of 278 million yuan. Meanwhile, another robotics company, UBTech, had operating revenue of approximately 1.269 billion yuan in the first half of 2026, but still recorded a loss of approximately 339 million yuan during the period. The data from the two companies cannot be directly compared, but this is enough to show that embodied intelligence companies cannot be lumped together into the category of "not yet profitable." What is more worth asking is: where do the profits that have already been realized come from, and can they be sustained? For companies still making losses, is investment approaching returns? There is also an easily overlooked account: after manufacturers make profits from selling equipment, can the customers who buy the equipment also make the numbers work?

This article focuses on the new generation of embodied intelligence applications represented by humanoid robots. The profit experience of mature industrial robots can be borrowed, but it cannot directly prove the economic viability of these new products. To judge where the industry has reached, one must look at both the suppliers' financial reports and the customers' costs of using robots.

Some are already profitable, while others are still bearing the costs of commercialization

Humanoid robots have become an important source of revenue and gross profit for Unitree. In 2025, revenue from this business was approximately 868 million yuan, contributing approximately 548 million yuan in gross profit, accounting for 54.40% of the gross profit of the main business. Humanoid robots have already contributed considerable gross profit. However, gross profit has not yet deducted all period expenses and cannot be treated as the net profit of an independent business.

Unitree also has two easily confused figures: the 2025 net profit attributable to shareholders was 278 million yuan, while the net profit attributable to shareholders after deducting non-recurring gains and losses was 591 million yuan. The difference mainly comes from the accounting treatment of equity incentives: the company included approximately 349 million yuan in share-based payments as expenses for the period, depressing book profit; when calculating net profit after deducting non-recurring items, this expense was excluded again. Therefore, 591 million yuan cannot be directly written as the amount the company "netted" that year.

UBTech shows another state. In the first half of 2026, the company's gross profit was approximately 567 million yuan, but selling, administrative, and R&D expenses were approximately 238 million yuan, 149 million yuan, and 303 million yuan, respectively. The three totaled approximately 690 million yuan, already exceeding gross profit, after which there were still impairments, financial expenses, and taxes that affected profit or loss. Therefore, the loss cannot simply be attributed to R&D, nor can it be asserted on that basis that every robot sold is losing money.

The company disclosed that the improvement in gross margin is related to the increased revenue share of high-gross-margin full-size embodied intelligence humanoid robots, but data that can directly determine the independent net profit of humanoid robots is still lacking. What can be confirmed is that the product mix is improving, but overall expenses have not yet been fully covered by revenue.

Companies that have already entered production sites may also still be at a stage of very small revenue and very large investment. According to Agility Robotics' materials, its 2025 net sales revenue was only 1.782 million US dollars, and it disclosed continuing operating losses and negative operating cash flow. As early as 2024, it signed a multi-year robot service agreement with logistics company GXO to use robots to move totes in warehouses. Commercial deployment has already occurred, but the revenue scale to support the entire company still needs to be established.

These differences are closely related to what companies sell and to whom. Universities and laboratories purchase robots for research and model training; commercial events purchase display and interactive experiences; factories purchasing equipment, in turn, look at the ability to stably complete tasks. All of them may form real business, but their growth drivers, delivery difficulty, and market boundaries differ.

Scientific research equipment manufacturers can be profitable before large-scale production applications mature. Display rental also has its value, but it must face fading novelty, idle equipment, and fluctuations in events. Treating all such revenue as evidence that robots replace productive labor would overestimate the maturity of end applications; conversely, denying their commercial value because they have not yet entered assembly lines would also be unfair.

After selling robots, the cost of doing work must also be calculated clearly

Manufacturers' financial reports can explain a company's operating conditions, but they cannot answer for customers whether the equipment is cost-effective. According to disclosures by BMW, in the 2025 pilot at its Spartanburg plant in the United States, the Figure 02 robot accumulated approximately 1,250 hours of operation, moved more than 90,000 parts, and handled the picking and placing of sheet metal parts required for welding processes. These records prove that the robot performed real tasks, but calculating returns also requires knowing equipment and integration costs, the frequency of manual intervention, maintenance costs, and the efficiency of the original solution.

A common method is to divide the robot's selling price by a worker's annual salary to estimate the payback period in years. The problem is that after the equipment is purchased, it still needs to be connected to systems, workstations need to be adjusted, and during use it needs charging and repairs, and sometimes personnel must take over. When the robot stops, does it slow down adjacent processes? None of these costs can disappear from the ledger.

A worker's job is also difficult to compress into a single repetitive action. A person moving boxes may simultaneously check for damage, identify anomalies, and adjust the order of materials. After the robot takes over the moving task, the other work still needs to be allocated. Therefore, how much labor a piece of equipment saves must be measured according to the actual process and cannot be directly equated with a person's entire salary.

A more reliable method is to add up equipment depreciation or rent, deployment and modification, operation and maintenance, manual assistance, and failure losses, and then divide by the actual number of qualified tasks completed. Only when the cost per task declines does it mean economic viability is improving. If dangerous exposure can also be reduced and night-shift gaps filled, the specific sources of these benefits should also be explained.

Before factories purchase humanoid robots, they will also compare other automation solutions: conveyor lines or mobile robots can be used to move materials, and robotic arms can handle grasping and assembly at fixed workstations. Humanoid robots need to prove that their flexibility can bring benefits sufficient to offset the extra cost. When tasks remain fixed over the long term, specialized equipment may be cheaper and more reliable; only when products change models frequently and workstations are inconvenient to modify may the flexibility of embodied intelligence show its advantages. Whether a humanoid is needed, and whether legs are needed, should both be judged by returning to the specific task.

Manufacturers also have to bear the cost of making customers willing to try. Agility disclosed that its customer acceleration program charges a project fee of about $500,000, but requires months of engineering, deployment, and hardware resources, and participants may ultimately not convert to commercial deployment. Paid pilots can share expenses, but may not cover the full investment, and certainly cannot guarantee subsequent orders.

This also explains why "robots as a service" is worth looking forward to, yet cannot automatically solve the profitability problem. Charging monthly or per task reduces customers' initial investment; the risks of equipment purchase, idleness, maintenance, and upgrades remain more on the supplier's side. Manufacturing and deployment cost money first, and service revenue arrives in installments afterward. Manufacturers need to manage utilization, delivery costs, and collections at the same time; otherwise, business expansion may increase financial pressure.

To judge revenue quality, it is therefore necessary to continue tracking whether orders are delivered, whether customers accept them, and whether payments are collected. A project may generate considerable revenue, but if it relies long-term on engineers stationed on-site to handle exceptions, profits may still be swallowed by service costs.

Temporary losses may be allowed, but the results of investment must also be tested

Companies that are already profitable also face the trade-off between current profit and long-term investment. Unitree's prospectus discloses that revenue in the first quarter of 2026 was about 423 million yuan, up 68.49% year on year, while net profit attributable to the parent company after deducting non-recurring gains and losses fell 52.55% year on year to about 40.25 million yuan. The company explained that growth in R&D and sales expenses affected profit. This set of data shows that revenue growth does not necessarily bring profit growth. After companies increase spending on R&D and marketing, profits may also decline. Whether these investments can bring subsequent returns still needs to be observed.

But "investing in the future" needs specific content. Can R&D shorten the adaptation time for new tasks? After customers increase, can similar projects be delivered faster? Can one engineer support more equipment? Only when these changes gradually occur do investments have a chance to translate into replicable operating capability.

This is also where the large market potential of embodied intelligence lies. If the same set of software and hardware, after limited adjustments, can take on more tasks, then R&D and manufacturing investment can be shared by more customers. Only work that has too small a batch size, too much variation, and was not worth automating in the past may be able to enter the market. If every change of scenario requires retraining and reintegration, and delivery personnel increase proportionally with the number of projects, companies will easily remain stuck for a long time in high-cost customized business.

Therefore, losses should be broken down and examined separately. Mature products can already cover manufacturing, delivery, and after-sales costs, and an overall loss caused only by developing new products is a different operating situation from service expenses that cannot be recovered and expand with every additional project. The former requires evaluating R&D direction and funding horizon, while the latter requires re-examining products, pricing, and target customers.

For companies, more mature businesses can be used to maintain cash flow, and budgets and validation milestones can be set for developing new capabilities. For investors, patience can be reflected in allowing time for technical validation while continuously checking repeat purchases, delivery efficiency, and collections. When unlisted companies lack complete financial disclosure, judgment should be reserved, and financing amounts, framework orders, or planned capacity must not be used as substitutes for profit.

A factory buying a robot for the first time may bring technical reserves and the expectation of trying it out. After using it for a period of time, it has already seen equipment fail and has paid for maintenance and manual assistance, yet is still willing to allocate budget to add a few more workstations. Such a decision carries more weight.

In short, profitable companies have already emerged in embodied intelligence, and humanoid robot sales can also contribute considerable gross profit. What remains to be verified is whether this profitability can continue to expand, and whether end customers can obtain stable returns from robot labor. Only when manufacturers can continue to be profitable and customers can also benefit from use will the growth of embodied intelligence have a solid commercial foundation.

The Paper Weibo; The Paper WeChat official account; The Paper Douyin account; Paisheng Wanwu Open Platform; IP SHANGHAI; SIXTH TONE

Tip hotline: 021-962866; Tip email: news@thepaper.cn