CITIC SEC released a research report stating that the embodied AI industry is experiencing rapid growth, driven by the synergy of capital, policy, enterprises, and talent, propelling the industry forward at an unprecedented pace. Training data is a critical factor in transitioning from "semi-commercialization" to full commercialization. The report recommends paying extra attention to companies that hold data dominance, particularly those with successful deployments across three key elements: data, models, and hardware.

From a commercialization perspective, compared to embodied model companies, data corpus providers—acting as "shovel sellers" in the AI era—are progressing faster in monetization and deserve close attention. Key focus areas include: 1) OEM manufacturers; 2) data enablers; 3) motion capture devices; 4) capital investors; and 5) computing power and cloud services. CITIC SEC's main views are as follows:

**Why the Focus? Indispensable Yet Scarce Real-World Data** The development of embodied AI faces challenges due to the lack of large-scale training data. While large language models benefit from vast internet knowledge and autonomous driving leverages continuous real-world data collection, embodied models struggle with insufficient directly usable data assets, often relying on controlled lab environments. The industry categorizes embodied model training data into a "data pyramid." Although synthetic and internet data scalability is widely recognized, real-world data at the pyramid's apex remains irreplaceable due to its physical authenticity and annotation value, making it the linchpin in addressing embodied data challenges.

**Market Potential? Data Factories Foster New Business Models** Since 2025, data collection factories have proliferated, with companies like Zhiyuan leading the charge. Public-private partnerships between local governments and OEMs to establish such factories are increasing. Common collaboration models include: 1) governments procuring robot hardware and providing operational sites, and 2) companies supplying data-collection hardware and ongoing technical support. A factory with hundreds of units can generate tens of thousands of hours of annual data, potentially yielding millions in sales. Currently, large model firms without mass-production capabilities are the primary demand source for embodied data. Companies like Qunkor Tech and Paxini Perception have listed data products on major domestic exchanges. Cost-wise, China’s lower labor expenses for data collection compared to North America provide a competitive edge. Though immediate ROI from data sales is limited, these factories may serve as strategic hubs for OEMs’ nationwide distribution networks, talent acquisition, and branding. Projections indicate over 700,000 hours of real-world data output in China by 2025, with nearly 8,000 data-collection units deployed by 2028.

**Competitive Edge? Setting Standards is as Crucial as Participation** The industry grapples with data silos due to: 1) inconsistent dataset standards; 2) fragmented data from divergent technical approaches; and 3) closed ecosystems. Standardization and benchmark cases are urgently needed. Some OEMs have gained first-mover advantages by fostering open ecosystems through national-regional innovation centers. In May 2025, the National-Local Humanoid Robot Innovation Center partnered with OEMs to launch the Embodied AI Data Alliance. By September, China’s first humanoid robot dataset standard was released, with Zhiyuan Robotics and Coupas earning CR certification as data quality benchmarks. Coupas, a professional data corpus firm backed by Shanghai’s government, aims to reduce industry data costs. The rise of data corpus players warrants long-term attention.

**Risks:** Slower-than-expected robotics advancements, limited application scenarios, technical route shifts, inadequate policy support, heightened competition, regulatory risks, and geopolitical uncertainties.