Nomura China Robotics: Data Emerges as the Critical Part for Robotics

Nomura highlights that data acquisition is becoming the critical bottleneck and value driver for humanoid robots in 2026. Based on an estimated industry shipment of 100,000 units, annual data demand is projected to reach approximately 10 million hours. The market exhibits distinct volume and price stratification across four data types. Non-embodied data accounts for 40-50% of total hours but commands only RMB 100-300 per hour, translating to a TAM of RMB 1-1.5 billion. Conversely, real machine teleoperation represents roughly 30% of hours but generates the highest unit price at RMB 500-1,000 per hour, making it the most valuable sub-market at RMB 2.2-2.5 billion.

The firm argues that a closed-loop hardware-software model is a structurally defensive strategy for data providers, mitigating vertical integration risks from downstream humanoid OEMs that standalone Data-as-a-Service (DaaS) models face. Simulation and synthetic data remain the cheapest at roughly RMB 50 per 10,000 frames. While companies like NVIDIA and Physical Intelligence have demonstrated significant performance gains using synthetic data, precision assembly and contact-rich manipulation still fundamentally require real-world machine data for fine-tuning.

This dynamic directly relates to physical hardware constraints. The trade-off between dexterous hand size and sensor payload capacity heavily limits tactile fidelity, ensuring simulation cannot independently cover household scenarios within the next three to five years. Nomura anticipates that industrial use cases, such as machine tending and sorting, may achieve qualitative breakthroughs by 2027-2028. However, broader household deployment will likely be delayed until post-2030.