Because it appeared alongside Cai Ming in a skit on the 2026 CCTV Spring Festival Gala, the humanoid robot Xiaobumi quickly became a sensation. Combined with its breakout performance at the 2025 robot marathon, Songyan Power, founded by Jiang Zheyuan, who studied all along on the Tsinghua campus, is hoped by the industry to become a second Unitree Robotics. In March of this year, Songyan Power completed its Series B financing, led by Chendao Capital, an industrial investment platform under the CATL ecosystem. To date, Songyan Power has completed 10 rounds of financing in total. Yet beyond the buzz, the real test lies at the delivery end. Earlier this year, Jiang Zheyuan set a target of at least 10,000 units in Xiaobumi sales for 2026. As of September this year, cumulative deliveries of Xiaobumi had just passed 5,000 units.
How much of Xiaobumi's sales go to B-end (enterprise) versus C-end (consumer) customers? "My personal feeling is roughly half and half," Jiang Zheyuan, founder of Songyan Power, recently candidly told the Daily Economic News reporter at the "2026 BILIBILI NEXT Conference" held on Bilibili. Betting on price is the path that sets Songyan Power apart from its peers. While Unitree Robotics, AgiBot and others compete over "how many units come off the line" and "how many work hours can be delivered," Songyan Power, with a price tag of 9,998 yuan, pulled a bipedal humanoid robot below the ten-thousand-yuan threshold for the first time.
Facing an industry shift toward competition over "whose robot works better," can Songyan Power's ten-thousand-yuan machine support the commercial expectations of large-scale delivery? With the rapid evolution of general large models, as players such as AgiBot, Xiazibianliang, and Galbot have long made embodied brains the core of their narratives, how can the late-starting Songyan Power break through?
Low prices do not mean the market is fully mature
Ten-thousand-yuan humanoid robots are becoming a watershed for the commercialization of embodied intelligence. Compared with Unitree Robotics and AgiBot robots entering performances, household chores, factory applications, shopping mall scenarios and the like, Xiaobumi has avoided these tracks and scenarios since its launch. In early 2026, Jiang Zheyuan said that in the short term he wanted Xiaobumi to first enter home scenarios, especially families with children, to perform companionship, interaction and education functions. He then even called for a target of at least 10,000 units in Xiaobumi sales for 2026. However, a low price does not mean the market is fully mature. Today, Songyan Power is still a considerable distance from that sales target. Data shows that as of September 2026, cumulative deliveries of Xiaobumi exceeded 5,000 units, with B-end orders approaching 2,000 units.
The Daily Economic News reporter noted that as of November 3, on JD.com's "NOETIX JD Self-operated Flagship Store," the 9,998-yuan Xiaobumi Spring Festival Gala same-style home robot made the TOP20 of the November smart robot gold list. Jiang Zheyuan said: "As for the share of Xiaobumi's B-end and C-end sales, I personally feel it is roughly half and half. C-end shipment volume will be larger, but B-end monetary value will be greater." This means that although consumer-grade humanoid robots have a certain user appeal, their commercial value still relies more on institutional scenarios such as scientific research and education and commercial display.
At present, price, mass production and intelligent capability are becoming common keywords across the embodied intelligence track, and competition among ten-thousand-yuan machines is quietly changing: price may no longer be a barrier, just a ticket in. The Daily Economic News reporter noted that in June this year, Unitree Robotics cut the price of its R1 humanoid robot from 39,900 yuan to 29,900 yuan and topped JD.com's sales list; in September this year, AgiBot's Qiyuan Robot put two personal robots, Q1 and T1, on the shelves starting at 19,999 yuan, and has now welcomed the official rollout of its 5,000th personal robot. As more and more leading players aim at cost-effectiveness, competition among ten-thousand-yuan machines is no longer just about "who can sell a robot more cheaply," but also about "who can deliver genuinely worthwhile capability at that price."
Robot companies cannot outcompete companies that focus purely on large models
If ten-thousand-yuan machines solve the question of whether robots can be sold, then what the embodied brain determines is whether robots can move from executing preset actions to autonomously completing tasks. Facing the rapid iteration of today's large models, whether large models will replace robot companies' self-developed brains has also become a focus of market attention. In response, Jiang Zheyuan believes that a generative pretrained transformer will not be an entire brain — the model parameter size is too large to run on the device side, and network latency issues are also difficult to solve. "A general large model can serve as a good planner, or as a front end for System 2 (slow thinking), but the system integration still needs to be handled by the robot company itself. There is actually a lot of room worth exploring here."
"For robot companies, if you compete on computing power or financial resources, you definitely cannot beat those companies that focus purely on large models. The gap is very large." Citing Tesla's intelligent assisted driving system as an example, Jiang Zheyuan further explained that the reason intelligent assisted driving systems are difficult to replace with general large models is that they hold continuously rolling and accumulating driving data, and large models have never seen such data. "Part of the moat for robot companies in the future comes precisely from their own data closed loop in home scenarios and specific operational tasks. If these data can generate a flywheel, they will not be eaten by large models. This is also why we must build a brain," Jiang Zheyuan told the Daily Economic News reporter.
However, a data closed loop is not a slogan; it requires enough robots, sufficiently rich tasks, and real scenarios that can operate continuously. Since the beginning of this year, leading players such as AgiBot, Xiazibianliang and Galbot have long made embodied brains the core of their narratives, but by contrast, Songyan Power did not launch its embodied intelligence technology brand Scalabot until September this year, already far behind leading players in terms of timing. Asked why it first built the body and only later the brain, Jiang Zheyuan told the Daily Economic News reporter: "We are an enterprise, not a research laboratory. The definition of a high-tech enterprise, I believe, is to do technology at a height I can reach by standing slightly on tiptoe or jumping a bit. If, even looking through a telescope into the distance, there is still no possibility of large-scale implementation, then this is not something that should be done at the current stage."
Daily Economic News. All rights reserved.
Comments