UBS provides a detailed analysis of Unitree Robotics: a pioneer in humanoid robot assembly, leading in "cerebellum" and body technologies.

Within two years, Unitree Robotics increased its annual sales of humanoid robots from 5 units to over 5,200 units, capturing 30% of the global market share and becoming one of the few profitable OEMs in this sector.

On August 28, UBS released its inaugural report detailing Unitree Robotics. The report breaks down humanoid robot capabilities into three layers: the "brain" (perception, reasoning, and task planning), the "cerebellum" (motor control and dynamic balance), and the "body" (execution performance and reliability). The conclusion is that Unitree Robotics has a significant lead in the "cerebellum + body" layers, but the "brain"—AI generalization capability—remains a common bottleneck for the entire industry.

UBS analysts say that the current price-to-sales ratio of 54x for 2027 already fully reflects the company's leading position in the hardware sector and the expected revenue CAGR of 53% from 2026 to 2028. Further growth would require clearer progress in the "brain" of the company and sustainable profit conversion.

The company listed on the STAR Market on August 19th, becoming the first A-share listed company specializing in humanoid robots. It raised 6.099 billion yuan, and as of the reporting date, its share price was approximately four times the issue price of 150.8 yuan. From 2022 to 2025, its revenue is projected to grow at a CAGR of 140%, with its gross profit margin climbing from 43.6% to 60.3%, and net profit projected at 278 million yuan in 2025.

In 2025, industrial applications of humanoid robots globally will account for only 3% of demand, while Unitree Robotics derives 76% of its humanoid robot revenue from scientific research and education. The company has already invested approximately 39% of its IPO proceeds in embodied AI research and development, and in August, it entered into a strategic partnership with DeepSeek.

Hardware supremacy: The full-stack barrier of self-developed actuators and motion control

Buyers of humanoid robots today don't test whether a robot "can understand," but rather whether it "can move." Commercial performances require backflips and group dances, research institutions need high-performance hardware platforms, and data acquisition centers need stable execution devices—these scenarios place far greater demands on mobility and hardware reliability than on the generalization capabilities of AI.

Unitree Robotics' competitive advantage lies precisely in this area. The company independently develops actuators, motor drives, control systems, overall structure, and LiDAR. Actuators account for over 40% of the BOM cost of humanoid robots. While most OEMs rely on third parties to supply complete joint modules, Unitree designs its own components, sources externally, and then assembles them in-house. Its self-developed M107 motor boasts a peak torque of 360 N·m and a maximum pulling force of 10,000 N. Technically, it employs a QDD (quasi-direct drive) solution—a high-torque-density motor paired with a low-reduction-ratio planetary reducer and high-bandwidth force control—achieving a balance between dynamic response and engineering feasibility, making it more suitable for high-dynamic scenarios such as running and jumping.

In terms of motion control, the company achieves multimodal perception and motion control closed-loop through deep reinforcement learning and large-scale parallel simulation training. The H1 standard version has a maximum movement speed of 3.3 m/s, with a potential capability exceeding 5 m/s, while the racing modified version reaches a peak speed of 10 m/s. The "Superman" prototype, released in August 2026, claims a top speed of 12.66 m/s, surpassing the human running world record, but this has not yet been verified by an independent third party. Having performed a robot group dance on the Spring Festival Gala for two consecutive years, and winning 11 medals at the first World Robot Games (ranking first in both gold medals and total medals), the company has demonstrated the practical maturity of its full-body coordination control and multi-robot formation capabilities.

This capability translated into shipments. In 2025, global humanoid robot shipments reached approximately 18,315 units, with Unitree Robotics leading the pack with 5,500 units (5,215 units included in revenue). Cumulative sales of quadruped robots exceeded 33,000 units, with 23,037 units shipped in 2025 alone, representing a global market share of approximately 28%. Overseas revenue accounted for over 40%. The momentum continued into the first half of 2026: humanoid robot shipments reached approximately 5,900 units, maintaining a global market share of approximately 31%, with revenue increasing by 48.5% year-on-year to RMB 1.15 billion. During the same period, global demand for humanoid robots reached 19,100 units, representing a year-on-year increase of 272%.

ASP halved, profits rose – cost structure is the real moat.

In 2024 and 2025, the average price of Unitree's humanoid robots decreased by 56% and 36% respectively, while the average price of quadruped robots decreased by 16% and 6% respectively. However, the overall gross profit margin increased from 43.6% in 2023 to 60.3% in 2025. In the hardware industry, it is not common for gross profit margins to rise while prices are reduced and sales volume is increased.

This result is driven by the reduced BOM cost resulting from full-stack self-developed technology—vertical integration from motors to structural components reduces profit sharing with external suppliers. The industry's largest shipment volume also brings purchasing power. In addition, continuous optimization of the product structure—the revenue share of humanoid robots increased from 27% in 2024 to 51% in 2025, while the gross profit margin of humanoids (63%) is higher than that of quadruped robots (57%), and the increase in the proportion directly boosted the overall profit margin.

A comparison with peers further illustrates the point. In 2025, gross profit margins were: Unitree 60%, UBTECH 38%, and YueJian 46%. On the expense side, Unitree's operating expense ratio was 31.8% (including approximately 350 million RMB in equity incentives), and excluding these, its administrative expense ratio was only 2.97%, far lower than the industry average of 47%-75%. In the early stages of the industry, Unitree concentrated its R&D resources on the hardware itself and motion control rather than basic models, resulting in a lower R&D expense ratio than UBTECH and YueJian. However, this hardware competitiveness translated into more orders and stronger profitability.

However, Unitree's profit margin will be under pressure in the first half of 2026. Net profit margin will drop to 24%, partly due to the factory relocation dragging down gross profit margin, coupled with one-off expenses for Spring Festival Gala brand promotion and a 150% year-on-year increase in R&D expenses—the latter being a financial reflection of the strategic shift towards the "brain." UBS estimates indicate that increased R&D investment may drag down net profit margin in 2026-2027, but it is expected to stabilize above 19% from 2028-2030, with a compound annual growth rate of approximately 57% in net profit from 2025-2030.

Completing the "Brain"—DeepSeek Joins Forces with 2 Billion Yuan in R&D Investment

Industrial applications account for only 3% of global demand for humanoid robots, for a straightforward reason—insufficient AI generalization capabilities, making it difficult for robots to handle unstructured tasks. Industry assessments suggest that the current stage of humanoid robots is similar to the era of large language models before the release of ChatGPT—embodied intelligence models continue to evolve, but a breakthrough application inflection point is still some time away.

Unitree is pursuing a two-pronged approach: simultaneously advancing the World Model (WMA) and the Visual-Language-Action Model (VLA). UnifoLM-WMA-0 and UnifoLM-VLA-0 are expected to be open-sourced in September 2025 and January 2026, respectively. Approximately 56% of the IPO proceeds will be invested in R&D, with about 39% of the total funds directly allocated to embodied AI models and core algorithms. R&D investment is projected to exceed 2 billion yuan over the next 2-3 years, higher than the industry average. The company expects net cash of approximately 8 billion yuan in 2026, which should financially support this round of investment.

In August 2026, DeepSeek and Unitree Robotics entered into a strategic and capital cooperation agreement—DeepSeek received 933,400 shares in Unitree's IPO strategic placement, investing approximately RMB 141 million with a 36-month lock-up period. The two companies signed a Memorandum of Understanding covering general artificial intelligence, high-performance robotics, and AI foundational models. The 36-month lock-up period signifies a multi-year industry collaboration rather than a short-term financial investment. DeepSeek's large-scale model capabilities complement Unitree's robot hardware, motion control, and physical world data, potentially shortening the evolutionary cycle from "being able to move" to "being able to understand, make decisions, and execute." However, initial R&D and computing power expenditures may increase, and the integration complexity and commercialization timeline remain uncertain.

On the production capacity side, the new base, once completed and fully operational, will add 75,000 humanoid robots and 115,000 quadruped robots to the production capacity. The company is also building an ecosystem loop encompassing "hardware—toolchain—developers—applications—data." The core logic is to attract developers with a large installed base, rely on developers to enrich applications and data, and continuously iterate products and reduce costs based on real-world data. In China, 64 humanoid robot data collection centers are already in operation across 23 provinces, with another 26 under construction, continuously providing data for AI model training. In an optimistic scenario, global demand for humanoid robots will increase from 18,300 units in 2025 to 325,000 units in 2030, with a compound annual growth rate exceeding 70%, and the market size is expected to reach $16 billion.

Now, Yushu has proven that he can create the best "body".

The only remaining question is whether this body can house a truly intelligent "brain".

This article is from the WeChat official account "Hard AI" . For more cutting-edge AI information, please go here.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.