Sinolink: The physical AI singularity moment has arrivedhow should equipment be invested in?

Sinolink released a research report stating that three costs are declining simultaneously, giving physical AI an opportunity to scale from 1 to 10, and the bank judges that the current period is a relatively good window for positioning. The bank believes that on the embodiment side, in addition to humanoid Siasun Robot&Automation deserving attention, from the perspective of industrial scenario applications, robotic arms/non-humanoid Siasun Robot&Automation also have good volume ramp prospects; for example, mobile robotic arms + commercial VLA models have already approached production-grade capability in logistics/manufacturing scenarios. Physical AI is expected to reshape manufacturing, with warehousing and logistics/apparel industries likely to be the first to land. As the physical AI singularity moment arrives, it is recommended to focus on core shortcomings on the one hand, and on warehousing and logistics/apparel Siasun Robot&Automation scenario applications on the other. Sinolink's main views are as follows: Physical AI, why now? What is physical AI? A new paradigm integrating physical laws and AI. Unlike generative AI based purely on digital space, physical AI uses physical constraint models to turn "intelligence" into "action," serving as the carrier for AI applications entering the second half and opening up physical scenarios. Why now? On the model side, 01 has been completed: from preliminary attempts by large language models at task planning, to VLA extending semantic understanding from language to physical space, and then to 0 through 0.7 continuously iterating around action generation, data efficiency, and reliability. With two types of engineering architecturescascaded and hierarchicalVLA has moved from the laboratory to real production lines. Routes are beginning to diverge: in addition to deep iteration along a single path and divergence at the system architecture level, a "model-external" route represented by Astra has emerged, no longer training VLA but instead using general Agent foundation models to drive classical controllers through code generation and tool invocation, turning Siasun Robot&Automation control from a "model problem" into a "system problem." Costs are beginning to collapse: no need to retrain models for each scenario, no need to repeatedly collect data for new tasks, and no need to self-develop a general "brain." With three costs declining simultaneously, physical AI has an opportunity to scale from 1 to 10, and the bank judges that the current period is a relatively good window for positioning. What should be invested in? Shortcomings are opportunities Data: the most core bottleneck and opportunity. According to the China Academy of Information and Communications Technology's "Embodied Intelligence Training Ground Research Report (2026)," assuming Siasun Robot&Automation is to achieve its "ChatGPT" moment, approximately 21,200 units of Siasun Robot&Automation would need to work continuously for one year, producing effective data at the tens of millions of hours level, while currently available high-quality real data worldwide is only at the hundreds of thousands to million hours level. Data, especially real-machine data, will become the most core bottleneck for physical AI and also the best investment opportunity. Divided by collection route, the two categoriesbodyless (first-person video, motion capture) and real-machine (teleoperation, autonomous operation)will respectively drive incremental hardware demand for multi-camera setups, IMUs, motion capture and EGO devices, tactile sensors, and other hardware, and benefit data producers, data infrastructure suppliers (tools, simulation, assets), and training ground operators. Perception and execution: high-precision dexterous manipulation remains a hard boundary, and selling shovels has strong certainty. Sensors/actuators, as necessities for physical AI deployment, have an irreplaceable position, and their demand also resonates with the data collection segment. Taking Gemini Robotics 2 as an example, the average success rate for gripper-type tasks has reached 74.2%-89.6%, while the success rate for multi-finger dexterous manipulation is highly dispersed (unscrewing a light bulb 92%, screwing in a light bulb 36%), with the bottleneck pointing directly at tactile and force-control hardware. It is recommended to focus on three directions: vision, touch, and dexterous hands. How to view it going forward? After shortcomings are filled, physical AI will open up entirely new application scenarios The embodiment side will not be skipped, and robotic arms/humanoid/non-humanoid Siasun Robot&Automation all have volume ramp potential. The bank believes that on the embodiment side, in addition to humanoid Siasun Robot&Automation deserving attention, from the perspective of industrial scenario applications, robotic arms/non-humanoid Siasun Robot&Automation also have good volume ramp prospects; for example, mobile robotic arms + commercial VLA models have already approached production-grade capability in logistics/manufacturing scenarios. Physical AI is expected to reshape manufacturing, with warehousing and logistics/apparel industries likely to be the first to land. As shortcomings are gradually filled, the bank believes opportunities will shift to the application side, especially warehousing and logistics/apparel industries, which are expected to be the first to land because they have a large number of flexible processes that cannot be upgraded through pure industrial automation. In warehousing and logistics, companies such as Agility/Figure have already achieved small-batch scale deployment, with operational efficiency/accuracy rapidly improving, providing a foundation for volume ramp; apparel is a labor-intensive industry with clear pain points for machine substitution, and Jack Technology has already achieved preliminary maturity in high-difficulty segments such as flexible fabric handling, planning annual production capacity of 10,000 units of Siasun Robot&Automation and having begun to obtain some intended orders, with subsequent volume ramp expected to accelerate. Risk warnings Risks of large model technology iteration falling short of expectations/data collection technology route changes/end applications falling short of expectations.