From hardware bodies to data infrastructure, the logic narrative under the physical AI wave. The expectation that artificial intelligence will move from the digital world into the physical world is being repriced by capital markets. On the news front, in late October, Tesla confirmed that the first batch of Optimus robots with mass-production quality rolled off the line, and NVIDIA founder and CEO Jensen Huang announced at GTC Berlin that "physical AI is the next wave of AI." A series of industrial events and news items landing pushed the A-share robotics sector out of a bottom rebound starting October 16. Unitree Robotics (688836.SH) rose 58% in eight trading days, and segments across the industrial chain such as lead screws and complete machines all posted gains of varying degrees. However, unlike the previous focus on hardware bodies, the core main line of this round of market movement is data. Capital is paying more attention to who can supply high-quality, trainable, and closed-loop embodied data at scale. Data collection equipment and data service providers have thus become the segment with the most concentrated gains.
Why capital is flowing back into robotics. On the October 27 trading session, the robotics sector strengthened in some areas. Xiangyang Bearing (000678.SZ) hit the daily limit, while Shangwei New Materials (688585.SH), Shuanghuan Transmission (002472.SZ) and others followed higher. Suochen Technology (688507.SH), Haitian Ruisheng (688787.SH) and others closed lower, showing clear divergence between robotics hardware and data collection segments. Looking back at the recent trend of the robotics sector, October 16 was the turning point. Before that, the East Money humanoid robot index had seen net outflows for three consecutive trading days, and from October 16 it turned to net inflows. The sectors leading the gains this time were not weighted stocks, but mid-cap names that had been suppressed for a relatively long time. Unitree Robotics is the emotional center of this round. The company listed on the STAR Market on August 19 this year at an issue price of 150.80 yuan. On its first day of trading, it touched an intraday high of 1,100 yuan, then slid all the way down, at one point falling 63.2% from its listing high. From October 16 to 27, over eight trading days, its stock price rose 58%, closing at 639.96 yuan on October 27 with a turnover rate of 23.42%. In addition, industrial chain stocks such as Zhaowei Electromechanical (003021.SZ), Orbbec-W (688322.SH), Suochen Technology, Xiangyang Bearing, and Shangwei New Materials all rose. There are two main factors driving this rebound in robotics. First is the concentrated release of news. On October 19 U.S. Eastern Time, Tesla said in its Form 10-K that the third-generation Optimus is ready for large-scale mass production and that it continues to expand AI training infrastructure. Then in its third-quarter report on October 22, Tesla disclosed that Optimus has started early capacity ramp-up at the Fremont factory and that the first batch of robots meeting mass-production quality standards rolled off the line. The Fremont line is designed with annual capacity of about 1 million units, and a second line in Texas is being prepared, with a long-term designed annual capacity of about 10 million units. On NVIDIA's side, during the GTC Berlin conference it announced an upgrade to the Isaac general-purpose robotics development platform, launching a new humanoid robot reference framework, simulation workflows, and data tools. Second is the positioning of holdings in the robotics sector. The robotics index reached its intra-year high on May 26 and had adjusted for nearly five months by October 27. The Wind robotics index had a maximum drawdown of 29.22% and was still down 13.27% cumulatively for the year. By contrast, AI hardware, also a main line in technology, posted significant gains during the year. The optical module index rose 41.26% for the year, the semiconductor equipment index rose 45.35%, and the memory index rose 18.73%, but the three had retraced 39.20%, 47.48%, and 43.48% respectively from their intra-year highs on June 25, July 1, and July 1. On one side is negative return for the year and a five-month pullback; on the other is considerable gains but ongoing valuation digestion. The relative choice for capital is not hard to explain.
After physical AI, the "data sellers" are getting orders first. The core narrative of this round of physical AI has shifted from "building the body" to "feeding the brain," that is, from the body to data. OpenAI founder Sam Altman previously said the biggest bottleneck for humanoid robots is not hardware, but that AI's understanding of the physical world is still insufficient. Multiple institutional research reports also pointed out that the key bottleneck of physical AI is no longer limited only to computing power, algorithm models, or simulation environments, but more to how to accurately acquire multimodal real-world data covering vision, touch, and more, especially data on dynamic interaction between humans and the physical world, to understand the laws of dynamic physical interaction in complex three-dimensional space, and to seamlessly deploy AI capabilities in embodied intelligence and other practical application scenarios. According to estimates by Zheshang Securities, for embodied foundation models to reach their own "ChatGPT moment," at least tens of millions of hours of real data are needed, while currently the world's available high-quality real data is only at the level of hundreds of thousands to one million hours, leaving a gap of orders of magnitude between supply and demand. Against this backdrop, the market began to focus on "who can provide scalable, high-quality, trainable, and closed-loop data supply capability." The speculative route of market capital also follows the above logic. Most of those at the top of the gainers list are not traditional reducer, lead screw, or motor names, but companies related to "data." Haitian Ruisheng's main business is training dataset products and services; Orbbec provides 3D visual perception sensors and is an important entry point for robot vision data; Suochen Technology engages in CAE simulation software and is benchmarked against "simulation engines plus world models"; Fantu Digital Creative's main business is digital twins and three-dimensional imaging and has been classified by multiple institutions into the direction of "spatial data collection and generation." The above four companies have averaged a gain of 63.8% since October 19, while traditional core component manufacturers' gains in the same period were concentrated in the 12% to 27% range. As the bottleneck of the humanoid robot industry accelerates its shift from hardware bodies to data infrastructure and brain training, data collection equipment, data services, real-scenario entry points, simulation platforms, and real-machine feedback loops—"selling shovels for data"—are expected to become the segment in the industrial chain that first realizes orders and performance. In the A-share market, who can catch these demands is the focus of market attention, and the third-quarter reports currently being disclosed also give investors some informational guidance. Orbbec released its third-quarter report on the 26th, saying that in the first three quarters sales revenue of vision solutions for the global robotics market reached more than twice that of the same period last year, with direct overseas revenue doubling. The company also pointed out that the key bottleneck of physical AI is no longer computing power, algorithms, or simulation environments, but the accurate acquisition of multimodal real-world data covering vision and touch. At its earnings briefing on the 27th, Orbbec said that at the current stage its data collection business is in a critical period of customer introduction and ecosystem building, mainly conducting adaptation, technical support, and prototype verification for specific customer scenarios, with relatively controllable investment intensity and cycle, and with strong upfront and phased characteristics. However, the gap between industrial development expectations and weak performance reality is also clear. Orbbec's data collection business development has not yet translated into performance growth. In the first three quarters, the company achieved revenue of 711 million yuan, down 0.43% year on year, and net profit attributable to shareholders of 73.8615 million yuan, down 31.62% year on year. Unitree Robotics recently announced a reminder that the humanoid robot industry is still in the early stage of commercialization, and there is uncertainty in technology maturity, cultivation of downstream market demand, and supporting capacity of the industrial chain. At the same time, the large-scale implementation of physical AI technology from laboratory demonstrations to real scenarios still faces multiple challenges such as perception accuracy, motion control stability, data accumulation, and software-hardware coordination. Fantu Digital Creative has been speculated by capital since October as a core target for physical AI data collection. Its stock price surged 122% during the month and refreshed a historical high. On investor interaction platforms and institutional research meetings, Fantu Digital Creative said it is building a physical AI data foundation around four major categories of data required for physical AI and embodied training: spatial data, simulation interaction data, bodyless data, and body data. In the first half of the year, Fantu Digital Creative's revenue fell 27.8% year on year, and net profit attributable to shareholders showed a loss of 77.25 million yuan.
Comments