Author: Peng Zhao (Founder of Zhifang Institute, Co-Founding Partner of Yunhe Partners) Original
With the preliminary inquiry on August 5 and online subscription on August 10, according to the disclosed issuance arrangements, Unitree Robotics (688836) is about to officially list on the STAR Market.
Over the past week, interpretations of this prospectus have been overwhelming: valuation, IPO subscription strategies, founder's shareholding ratio...
Peeling back the noise of the capital market, this is the first fully audited financial sample of the embodied AI industry in the A-share market.
The value of audited financial figures is often two-fold: it not only records a company's growth trajectory but also exposes the industrial ecosystem in which the company operates.
For AIoT practitioners, this document is more like a mirror, worth "reading backwards"—to read the absence of AIoT in the procurement details, to read the rewritten delivery models in accounts receivable, and to read the yet-to-be-named "scheduling rules" in the fundraising projects.
Summarizing the whole text, I would like to distill it into three keywords: internalization, product standardization, and scheduling rights.
Hardware "Internalization": The Folded AIoT Supply Chain and Reshaped Pricing Power
Among the raw material procurement composition disclosed in the prospectus, the proportion of mechanical parts increased from 39.91% in 2023 to 48.45% in 2024, and further to 50.71% in 2025; meanwhile, during the same period, the proportion of electronic components dropped all the way from 31.88% to 22.40%.
In Unitree's 794 million RMB procurement basket in 2025, more than half consisted of machined parts, die-cast parts, and fasteners.
More robots are being manufactured year by year, yet the procurement basket is becoming increasingly "metalized" year by year.
Where did those vanished electronic components go?
The answer lies in the comparison between the supplier list and the in-house R&D list.
The top five suppliers accounted for a combined 22.54%, and their procurement content only included mechanical parts, electronic components, electrical materials, and auxiliary materials—no sensor companies, no module companies, and certainly no AIoT platform enterprises.
In contrast, the in-house R&D and manufacturing list provided in the prospectus includes: high-performance motors, reducers, joint modules, battery modules, LiDAR, high-computing-power modules, dexterous hands, and motion control systems.
Overlaying these two lists, the conclusion is obvious: perception, computing power, energy management... these high-value segments that were once the pride of the AIoT supply chain are being fully absorbed into the robot's "body," leaving the external supply chain with primarily metals and standard parts.
The financial statements provide an audit-level verdict on the cost-effectiveness of this "full-stack in-house R&D" route.
The unit price of humanoid robots dropped from 593,400 RMB/unit in 2023 to 260,400 RMB in 2024, and further to 166,400 RMB in 2025, a decrease of approximately 72% over two years. However, even after the price was slashed by 70%, the gross profit margin of its humanoid robots remained as high as 63.18%, and the gross profit margin of the core business was 60.13%, nearly 16 percentage points higher than the 44.22% in 2023.
The prospectus states the reason very clearly: the vertical integration brought by the in-house R&D and manufacturing of core components "significantly reduced material procurement and manufacturing costs." The rate of decline in unit costs was far faster than the rate of decline in unit prices.
This is worth comparing with the history of smartphones.
The smartphone industry chose the opposite organizational model back then; phone manufacturers outsourced modules such as cameras, acoustics, displays, and connectivity, cultivating a massive component ecosystem through the division of labor and giving rise to a group of listed companies that coexisted with the OEMs. However, the first audited sample of embodied AI shows that, at least among leading enterprises, this scenario has not been repeated. The scale dividends have not spilled over to the supply chain but have been used to deepen "internalization."
On the other hand, the prospectus shows that its robot component business (including externally sold robotic arms, 4D LiDAR, high-computing-power modules, etc.) saw revenue surge from 26.92 million RMB in 2023 to 104 million RMB in 2025, with a gross profit margin of 59.36%. The OEM used the scale of its whole machines to amortize component costs, and then turned around to enter the market as a component supplier.
For AIoT component manufacturers, the situation has become a two-way squeeze: as sellers, they are shut out of the value sharing in embodied AI; as peers, they must now begin to endure the overwhelming asymmetric strike brought by the scale advantage of whole-machine manufacturers.
Restructuring Delivery Models: Crossing the Project-Based Quagmire with 3C Standard Product Logic
By the end of 2025, Unitree's accounts receivable balance was 61.9068 million RMB, accounting for only 3.64% of the operating revenue for the year, and this ratio showed a continuous downward trend over three years (6.95%, 5.60%, 3.64%); during the same period, the books showed 137 million RMB in contract liabilities, meaning customers pay first and the company delivers later.
The top five customers accounted for a combined 12.08% of revenue, with the largest customer accounting for only 3.11%. The sales expense ratio was merely 8.31%.
Placing this set of data in the context of the B2B equipment industry, the anomalies jump off the page.
The default delivery structure for traditional AIoT devices is: system integrators, PoC testing, bidding, on-site implementation, acceptance, and lengthy payment terms... The complexity remains at the customer site, requiring massive manpower for installation, networking, platform docking, and customized development. Consequently, financial statements commonly feature high accounts receivable, lengthy collection cycles, and heavy sales expenses.
However, Unitree's financial statements possess a completely different underlying structure.
The complexity of on-site implementation is packed into the robot body itself; full-stack in-house R&D ensures "out-of-the-box" usability, while SDKs and secondary development interfaces transfer the customization work entirely to buyers or ecosystem partners.
Once "integration" is thoroughly "standardized," "channels" degrade into pure logistics—online direct sales via Tmall, JD.com, Shopify, and Amazon, with offline sales dominated by direct sales and supplemented by distribution. The buyers are universities, research institutions, tech enterprises, and consumers.
A heavy-duty device with sensor and computing power densities far exceeding most industrial gateways achieves scaled delivery through 3C retail shelves—payment before delivery, standard products, returnable and exchangeable.
Unitree has proven with audited data what many AIoT device companies dream of: as long as implementation complexity is standardized, the project-based model can be entirely transformed into a product-based model.
Why can't traditional AIoT device companies achieve this, while embodied AI can? The bottleneck lies in the position of the "value closed-loop."
Traditional AIoT devices are mostly component-level products: a sensor, a gateway. While their own functions are complete, their commercial value can only be realized after the entire system is fully operational. Thus, the deployment process inevitably relies on a whole set of heavy on-site engineering—surveying, wiring, networking, protocol docking, platform integration, and business software customization. All of this happens at the customer site, delivered one by one by integrators on a project basis.
The devices are standard, but the deployment is extremely non-standard; the complexity settles on-site, and the financial statements naturally accumulate high accounts receivable.
Traditional automation equipment takes a different path: AGVs require laying magnetic strips and pasting QR codes, while robotic arms need to be equipped with fixed fixtures and safety fences. Essentially, it uses "environment modification" to exchange for operational "certainty." In short, it is "making the environment accommodate the equipment."
Embodied AI, for the first time, completely reverses this relationship: "making the equipment adapt to the environment." Whole-body motion control allows it to walk freely on unmodified stairs, lawns, and slopes; multimodal perception completely frees it from the constraints of magnetic strips and fixtures. The environment needs to make no compromises or preparations for it; those once time-consuming and labor-intensive "on-site engineering" tasks are already "built-in" when the robot leaves the factory.
Of course, the boundary here must be clarified: what has been successfully absorbed into the body so far is merely the "motion layer."
As for the adaptation of the "task execution layer"—that is, how to enter specific industries and perform tasks correctly at specific workstations—it is entirely transferred to buyers and the developer ecosystem via SDKs and secondary development interfaces. Unitree's current buyers happen to be primarily universities, research institutions, tech enterprises, and individual consumers. For scenarios such as scientific research, teaching, performances, and secondary development, "walking and motion" themselves are the final products, and complicated industry implementation is simply out of the question.
Therefore, the reason this 3C-style standard product delivery can work is half due to the generalization of motion intelligence, and the other half due to the clever bypassing of deep industry adaptation through customer group selection.
When the day comes that robots truly enter factories to work as "full-time employees," the bill for "task adaptation" will inevitably reappear. At that time, will this complexity continue to be "absorbed internally" by more powerful embodied foundation models, or will traditional integrators have to be "invited back on-site"? This is the suspense this prospectus leaves for the entire industry.
But at least within the current product definition and customer groups, the complexity of delivery has been successfully converged from "on-site engineering" to "factory capabilities and buyer self-service." B2B non-standard project contracts have thus been transformed into To C-style standard product orders.
The Battle for Scheduling Rights: The Watershed and Leverage for IoT Platforms
So, exactly how far will the "internalization" and "standardization" of hardware be pushed?
The 1.699 billion RMB revenue and 60.13% gross profit margin in 2025 were entirely earned by the robot's "body" and motion intelligence. The risk factors in the prospectus explicitly state: "Given that embodied foundation model technology globally is still in the R&D and testing phase, the company has not yet scaled the application of its self-developed general-purpose embodied foundation models to robot products during the reporting period." The relevant models are currently undergoing R&D testing and deployment validation in pilot scenarios such as its own factories.
This means that the foundation models at the cognitive layer currently contribute zero revenue.
However, according to this fundraising plan, the largest item, the "Intelligent Robot Model R&D Project," plans to invest 2.022 billion RMB, which is about 71 times the cumulative investment during the reporting period for the "Multimodal General-Purpose Embodied AI Foundation Model Project" (28.3628 million RMB), accounting for 48% of the total 4.202 billion RMB raised.
The audited financial statements describe an extremely profitable hardware company, while the fundraising plan describes a gold-devouring AI behemoth.
The current pricing of the "brain" has been set to "zero" by the company itself. UnifoLM-WMA-0 and UnifoLM-VLA-0 have been open-sourced successively, and the prospectus explicitly states the open-sourcing of "full-process algorithms such as data collection, model training, real-machine deployment, and motion control, to attract global developers to participate in algorithm optimization and general-purpose robot R&D."
The body is sold at a high price with a 63% gross margin, while the brain is opened for free at zero cost. These two seemingly contradictory pricing strategies together form a highly ambitious tactic: using a free brain to exchange for iterations by global developers on its own hardware.
However, the 5,215 robots sold so far currently only constitute revenue, not a network. After the equipment is sold outright, task execution data is not returned by default, manual corrections are not recorded by default, and retrained models have no default channel to return to the field. A disconnect occurs here between shipment volume and learning volume.
The capabilities required to fill this gap include: connection management, data feedback, edge caching, device shadow, OTA, fleet device management... We will find that these are precisely the capability checklist of AIoT platforms.
The prospectus states that this 2.022 billion RMB will be used to "build large-scale real-world datasets, develop cost-effective teleoperation systems, and establish automated annotation mechanisms to form a 'data flywheel'." Embodied AI tells an AI story in the capital market, but hidden within the composition of capital expenditures is a massive AIoT budget.
The reason this budget is so massive is that to make tens of thousands of scattered robots "come alive" and continuously evolve, a massive digital infrastructure must be built in the cloud.
The robot's learning network requires two planes: the data plane is responsible for feedback, and the control plane is responsible for scheduling. The signal given by the prospectus is that the builder for both planes bears the same name: Unitree itself. On the data plane, teleoperation systems and annotation mechanisms are all listed in the self-built checklist; on the control plane, cluster scheduling capability is listed as a core dimension.
The "internalization" that occurred at the hardware component layer is now being replayed at the software platform layer.
Will the scheduling gate for future factories fall on the Industrial IoT platform, or in the cloud of a certain robot manufacturer? This is the largest division of ownership in industrial software over the next five years, and the complexity of existing on-site deployments is the natural leverage in the hands of AIoT platforms.
Conclusion
Looking back on this unstoppable path of evolution:
At the hardware foundation, high-value-added components are ruthlessly "swallowed" into the robot body, leaving the external supply chain with nothing but cold metals;
In the delivery middle ground, heavy on-site integration is reduced to out-of-the-box standard products, completely washing away the project-based quagmire on the financial statements;
At the cloud high ground, the scepter of data feedback and cluster scheduling is being firmly held by the OEMs themselves, attempting to complete a comprehensive closed-loop from edge to cloud.
What the first audited financial statement of embodied AI draws for the AIoT industry is exactly this unstoppable advancing line of "internalization" and "standardization."
In the previous article, "After the Demo: The Life-and-Death Game of the Trillion-Dollar Humanoid Robot is Hidden in the Power Distribution Cabinets of the Workshop," we dissected the electricity bills of robot factories; in this article, we analyzed the capital ledger of robot companies together. Combined, the two articles are two sides of the same coin: The AI dream of embodied AI reaching for the stars will ultimately need the tracks laid by AIoT in the physical world to bear it.
Data Note: All data in this article are cited from Unitree Robotics' "Prospectus" (July 31, 2026). The financial data from 2023 to 2025 are audited figures. The issuance schedule is subject to the final announcement of the exchange.