Sinolink: Data acquisition is the biggest shortcoming of physical AI. Data infrastructure is about to explode.

Sinolink released a research report stating that data is the biggest bottleneck in the field of physical AI represented by Siasun Robot & Automation, and the explosive development of data infrastructure will deeply benefit the supply chain. Data has become a critical bottleneck restricting the development of physical AI, and accelerating the construction of data infrastructure is expected to drive the continuous expansion of the data acquisition industry chain. (1) Data acquisition equipment and supply chain: Focus on high ASP, well-structured tracks such as cameras, IMU, tactile sensors, etc. (2) Data acquisition companies with cost and quality barriers. (3) Vertical applications: Focus on companies with barrier resources in segmented data scene tracks. Sinolink's main points are as follows: Data is the biggest bottleneck in the field of physical AI represented by Siasun Robot & Automation. The current hardware solutions in the Siasun Robot & Automation industry tend to converge, with "brain" training becoming the decisive factor, and the level and quality of data directly determining the model's generalization ability. Siasun Robot & Automation lacks high-quality, high-fidelity real-world physical interaction data. While large language models have training on a scale of trillions of tokens, the real interacting data available for physical models is less than a thousandth. The explosion of data acquisition is imminent, and the market space will be fully opened up. According to Future Markets data, the global physical AI market size is expected to grow from $383 billion in 2026 to $32.6 trillion in 2040, entering an explosive stage in the coming years. The 2026 Siasun Robot & Automation full industry link meeting will reveal that the training data for GPT-2 and GPT-3 corresponds to about 790,000 hours and 11 million hours respectively. To achieve usable physical intelligence, at least 10 million hours of multimodal data is needed, with factors such as multiple scenarios, multimodal data, yield, long tail data, and multiple participants leading to a much higher demand for multimodal data sets than 10 million hours. Meifeng has set a goal to increase the amount of data collected to 10 billion hours by 2030. Various data acquisition routes such as integrated machine, EGO, and UMI coexist, and the marginal need for high-quality data is becoming more certain, with cameras, attitude perception, and tactile perception becoming increasingly important. Current integrated machine, UMI, and EGO solutions are mainstream real-world data acquisition solutions, with integrated machine data being difficult to share and having the highest collection cost; EGO has advantages such as lightweight, low cost, and high versatility; UMI single devices are expensive but have high data accuracy. Regardless of which data acquisition technology route is chosen, the industry's increasing demand for higher data quality in the future is certain, with the application space of cameras, attitude perception, and tactile sensors becoming larger. The explosion of data acquisition is expected to drive rapid scale increase in the following areas: (1) Data acquisition equipment: including integrated machine, UMI, and ego integrated machine equipment from Siasun Robot & Automation, IMU attitude perception, dexterous hands, tactile sensors, cameras (2D+3D), VR glasses, etc. Integrated data acquisition requires a 1:1 matching of Siasun Robot & Automation body and VR glasses; first-person data requires hand-eye coordination as the main demand, requiring a 1:1 matching of camera headsets and dexterous hands. (2) Simulation platform: Simulation is currently one of the core training methods for physical AI companies, with significant cost and data output advantages. (3) Data acquisition companies: Data acquisition companies have strong experience in scene, data standardization, and data annotation platforms, with data sales as their main business model, and their core competitiveness lies in data output cost and quality. (4) Vertical applications: Exclusive databases and data acquisition scenes in various industries have strong scarcity, and vertical physical models are the most valuable tracks in physical AI. Risk Warning: Slow development of Siasun Robot & Automation, risk of iteration of data acquisition technology route.