NextFin News -- As embodied intelligence becomes the next battleground for the AI industry, global tech giants have been ramping up their bets: Tesla’s Optimus and Boston Dynamics’ Atlas have continued to iterate, while domestic players such as UBTECH and Dobot have been accelerating to catch up.Amid this double game of technology and capital, Unitree Robotics, a benchmark company in China’s high-performance general-purpose robotics space, sparked heated discussion across the industry with a set of figures disclosed in its IPO prospectus: in 2024, the company’s R&D spending totaled just RMB 90.4663 million. What does RMB 90 million actually mean? Online rumors have claimed that Tencent once paid RMB 100 million in annual salary to poach Yao Shunyu from OpenAI—meaning Yao’s annual pay alone would have been enough to cover Unitree’s entire R&D budget for the year.By comparison, from 2022 to 2024, the average R&D expense ratio of comparable companies in the same sector (such as UBTECH and Dobot) was higher than Unitree’s; in the first three quarters of 2025, Unitree’s R&D expense ratio fell to less than half the industry average. The company attributed this to “economies of scale,” but during a breakout period of rapid technological iteration, such a sharp decline in the R&D expense ratio may be read by the market as a shift from “technology-driven” to “market-driven”—or as a warning sign about its ability to sustain high-intensity investment over the long term.A Pragmatic Roadmap or Only Good at "Jumping Around"?The world's leading chip maker TSMC Chairman Che-Chia Wei has pointed out that "the robots in mainland China are just jumping around, which is "useless and merely for show." His criticism may sound harsh but it reveals some truth. Unitree Robotics, a typical player in China's robotics industry, has submitted its initial public offering plan to the Shanghai Stock Exchange, aiming to raise 4.2 billion yuan in the IPO. In a tech-industry narrative where companies casually claim annual R&D outlays of tens of billions—or even hundreds of billions—Unitree’s prospectus figures look decidedly “out of step.” From January to September 2025, the company’s R&D spending was RMB 90.2094 million. Even with a simple linear extrapolation, its full-year R&D investment would be expected to come in at only around RMB 120 million. That number is not only far below the cross-industry investments of tech giants; it is even lower than that of many AI startups that have yet to turn a profit. Yet behind this RMB 90 million R&D budget, Unitree claimed in its prospectus that its 2025 humanoid-robot shipments had already exceeded 5,500 units (humanoids only, excluding wheeled dual-arm robots) and, citing third-party data, asserted that it ranked No. 1 globally by shipment volume. That raises a pointed question: if its R&D spending is nowhere near astronomical, what exactly has Unitree relied on to build what appears to be a solid competitive moat?The crux of the answer lies, first and foremost, in its highly focused and pragmatic R&D strategy—“full-stack in-house development” and “cerebellum-first.” This is not a simple slogan, but a survival philosophy that runs through its technology roadmap, cost structure, and product iteration.Full-stack in-house development: spend every dollar on building a closed loop. Unlike many robotics companies that take an “integration” or “assembly” approach, Unitree Robotics emphasizes independent R&D spanning the robot body, core intelligence algorithms, and critical components and subsystems. Its prospectus lists 12 core technologies in detail, covering key hardware and software links such as integrated joint modules, highly compact body design, self-developed LiDAR, and high-dynamic motion-control algorithms.In particular, the technology it calls “in-house development of core components and high-performance actuators” has enabled it to develop and manufacture its own joint modules (integrating motors, reducers, encoders, etc.). This means Unitree Robotics’ R&D spending has largely been converted into deep control over the supply chain and proprietary engineering know-how, rather than “black-box” modules purchased from outside. This model requires heavy upfront investment, but once the loop is closed, it creates pronounced dual barriers in both technology and cost: iteration is no longer constrained by others, and vertical integration substantially lowers the bill of materials. The prospectus disclosed that the starting price of its Go2 Air quadruped robot has already fallen below RMB 10,000, the base version of the G1 humanoid robot starts at RMB 85,000, and the R1 Air has dropped as low as RMB 29,900—direct reflections of this cost-control capability.“Cerebellum-first”: concentrate firepower on nailing the motion fundamentals. In the sweeping narrative of embodied intelligence, the “brain” (cognition and decision-making) and the “cerebellum” (motion and control) matter equally. Unitree Robotics’ choice was that, against the backdrop of “brains” (embodied foundation models) still being in the R&D and testing stage worldwide, it heavily tilted its limited early R&D resources toward building a world-leading “cerebellum” capability.The prospectus repeatedly emphasizes that its general-purpose robot products have already integrated an in-house developed embodied intelligence model, giving them “world-leading embodied-intelligence motion capabilities.” Through technologies such as deep reinforcement learning, the company has enabled its robots to perform eye-catching, highly dynamic movements—backflips, high-speed running (the H1 once set a world record at 3.3 m/s), complex dances, and impact resistance against knocks and falls. These capabilities are not mere showmanship; they are the foundation for robots to operate reliably in complex physical environments, and the “ticket to entry” for commercialization. Unitree Robotics has created a virtuous cycle of rapid algorithm iteration by open-sourcing its reinforcement learning platform and drawing in collaborations with universities and research institutions. This strategy is pragmatic: by leading with outstanding motion performance that is visible, demonstrable, and mass-producible, it has built brand awareness and market confidence early, buying valuable time and room for product sales and ecosystem building.Efficient technology reuse and organizational structure. The prospectus notes that the company’s humanoid and quadruped robots can achieve “technology sharing and reuse” across core modules such as joint actuation, mechanical structure, battery management, and software algorithms. This not only spreads R&D costs across products, but also dramatically accelerates the journey from prototype to mass production. From Laikago, its first quadruped robot in 2017, to the launch of the humanoid robot H1 in 2023, and then G1 in 2024, followed by R1 and H2 in 2025, the pace of iteration has been striking. Supporting this speed is the “highly flat organizational structure” mentioned in the prospectus, which reduces internal coordination friction and allows R&D to respond quickly to technological shifts and market demand.Beyond that, there is also a non-technical factor—educating the user mindset. As the first embodied-intelligence company to appear on the CCTV Spring Festival Gala, and to return for two consecutive years, Unitree Robotics has effectively become synonymous with “robots” in the minds of mainstream consumers. At last year’s WAIC, many non-industry attendees even mistook competitors’ products for Unitree’s. This kind of early move to capture user awareness has also given Unitree Robotics an advantageous position on the consumer side.“The brain” is the “shortest plank in the barrel”Despite its very strong profitability, the “moat built with 90 million yuan in R&D spending” revealed in the prospectus will need to be re-examined and reinforced in the face of ample capital after listing. The company’s primary battleground will undergo a profound shift, and the sustainability of its moat will face three major tests.The first and biggest test is the pressure to achieve a breakthrough in the “cerebrum.” Unitree’s core strength lies in motion control (i.e., the “cerebellum”): its robots’ running, backflips, and coordinated group performances have won broad recognition. However, both industry consensus and the company itself acknowledge that it is relatively weak in cognition, decision-making, and environmental understanding (i.e., the “cerebrum”). The robots the company delivers today are more like high-performance “hardware platforms” or “development platforms” that require secondary development by developers, rather than agents that can independently understand, decide, and execute. The prospectus explicitly warns that global embodied foundation-model technology is still in the R&D and testing stage, and that large-scale application remains uncertain—directly constraining the large-scale rollout of robots in non-standardized industrial settings and home scenarios.Wang Xingxing also said publicly in recent days that embodied intelligence’s true “GPT moment” is still not quite there yet. “Some people in the industry are relatively optimistic and estimate it can be achieved in 18 months. I’m probably a bit more pessimistic—I think it will take at least two to three years—but the pace will definitely still be very fast,” he said.The prospectus specifically emphasizes that embodied foundation-model (i.e., “brain”) technologies worldwide are still in the R&D and testing stage, and that during the reporting period the company had not yet deployed its self-developed general-purpose embodied foundation model at scale in robot products. The company also candidly notes that if “brain” technology fails to make major progress, it “will introduce uncertainty into the timeline for large-scale deployment of general-purpose robots.” At present, Unitree’s lead is to a large extent built on an outstanding “cerebellum” (motion control).However, for robots to ultimately reach broad application scenarios (such as home services and complex industrial operations), they must rely on a “brain” that can understand ambiguous instructions, plan tasks, and cope with unknown environments. This is the central battleground in today’s global race in embodied intelligence. The prospectus shows that the company has already laid out efforts along two major technical tracks—WMA (World Model–Action) and VLA (Vision–Language–Action)—and it open-sourced related models in September 2025 and January 2026, respectively. That also means that after listing, the capital market and competitors will closely watch for substantive progress in foundation-model R&D and the speed of commercialization and real-world deployment. The “intelligent robot model R&D project” in the fund-raising investment plan is precisely a response to this.Therefore, in Unitree Robotics’ IPO plan, the most important use of funds is R&D for intelligent robot models. According to the prospectus, Unitree Robotics plans to publicly issue no fewer than 40.4464 million new shares in this IPO and raise RMB 4.202 billion. The planned total investment for the intelligent robot model R&D project alone reaches RMB 2.022 billion.Going forward, R&D spending will have to tilt heavily toward this “money sink” and “deep-tech waters.” A 90-million-yuan R&D playbook may need to be upgraded to a new logic measured in the hundreds of millions—or even billions.Second is the intensifying, cutthroat competition in the industry. The prospectus lists “the risk of intensified industry competition and disorderly or improper competition” as a special risk. It specifically names Tesla, arguing that with its “capability for large-scale mass production and artificial intelligence technology resources,” Tesla may be able to rapidly reduce the cost of Optimus, thereby “directly exacerbating price competition in the industry.”In addition, tech giants at home and abroad, automakers, and startups are all piling in. The battleground has also expanded—from pure hardware specs to an all-around contest spanning full-stack “perception–decision–execution” technologies, supply-chain efficiency, capital strength, and application ecosystems. The scale and market reputation Unitree Robotics has built on its first-mover advantage serve as a valuable buffer. But after listing, it will need to use the funds raised (RMB 4.2 billion planned in this offering) not only for technological R&D, but also to speed up construction of an “intelligent robot manufacturing base” to expand its capacity advantage. At the same time, it may have to devote more resources to global market expansion and deeper industry solutions to meet competition on all fronts.Lastly is the deeper expansion of commercialization scenarios. The prospectus’ warning about the “risk that downstream large-scale commercial applications fall short of expectations” is a shared challenge across the industry. At present, the rollout of quadruped robots in scenarios such as inspection and firefighting is relatively clear, while “killer apps” for humanoid robots are still being explored. After listing, Unitree Robotics will need to prove that its products can not only sell well, but also truly integrate into a wide range of industries and create irreplaceable value. That requires evolving from a “provider of high-performance general-purpose robot platforms” into an “ecosystem builder” that understands the pain points of vertical industries more deeply and delivers end-to-end solutions. The IPO-funded “new intelligent robot product development project” is intended to broaden the product portfolio—precisely to cover more scenarios.RMB 90 million in R&D spending is a necessary foundation for Unitree Robotics to sustain its current product lead and market momentum, but it is far from enough to help the company build the decisive industry moat it will need for the next stage of global competition—one centered on high-level AI and general intelligence.Even though this prospectus is an outstanding example of “low-cost innovation,” going public means the game is entering its second half. Will the RMB 4.2 billion in funds raised help it break through the “brain” technology bottleneck? Can it translate its strengths in the “cerebellum” and mass production into all-around competitiveness against tech giants? Can it truly open the floodgates for large-scale commercial deployment of humanoid robots? The answers to these questions will determine whether Unitree Robotics can upgrade the “value-for-money moat” showcased in its prospectus into an “ecosystem-level moat” that can endure in both the capital markets and the deeper waters of the industry.(Note: 1 U.S. dollar equals 6.9 Chinese yuan.)更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App