Sinolink: Physical AI drives edge hardware into an upgrade cycle

Sinolink released a research report stating that the core change in this round of Physical AI is not a breakthrough in the mechanical structure of Siasun Robot&Automation itself, but rather that general-purpose models have begun to move beyond perception and understanding into planning, decision-making, and action spaces, extending AI from the digital world into the physical world. As model outputs gradually expand from language and images to real actions, the hardware requirements of Siasun Robot&Automation have also expanded from single AI computing power to complete edge-side system capabilities such as multimodal perception, local computing, Memory/Storage, real-time control, and functional safety. At the same time, Physical AI is not an isolated Siasun Robot&Automation hardware path: smartphones and PCs have already taken the lead in validating the large-scale deployment of local large models and Agents, while Siasun Robot&Automation further extends the same trend from digital interaction to physical execution. Sinolink believes that the leap in model capabilities is the technical starting point of this round of Physical AI, Siasun Robot&Automation is the most typical new terminal, and the upgrade of edge-side computing platforms is the broader industrial landing point for the electronics industry. Sinolink's main views are as follows: As models enter the physical world, Siasun Robot&Automation edge-side hardware is being driven from a single computing power upgrade toward system-level expansion. As Siasun Robot&Automation tasks expand from fixed workstations and single operations to open environments, full-body motion, and dexterous manipulation, models need to continuously acquire multi-dimensional information such as vision, posture, motion, and contact, and complete perception fusion, reasoning, planning, and control locally. Corresponding to hardware, the increment brought by Physical AI is not simply raising TOPS, but rather "large models + multiple sensors + real-time control" jointly driving the expansion of the entire edge-side data chain: larger model scale drives memory capacity and bandwidth demand, multi-channel perception and data recording increase local storage and high-speed I/O demand, and action output further increases real-time control interfaces such as EtherCAT and CAN-FD as well as functional safety resources. The value of Siasun Robot&Automation electronic hardware spreads from SoC to Memory, Storage, interconnect, and control segments. Siasun Robot&Automation has entered small-batch industrial validation, and Training demand is expected to be released before large-scale volume ramp of the robots themselves. In 2026H1, global humanoid Siasun Robot&Automation shipments had approached 25,000 units, with half-year shipments exceeding the full year of 2025, as the industry gradually moved from a thousands-of-units-level R&D market into a tens-of-thousands-level commercialization exploration; at the same time, the application structure has also begun to spread from scientific research, display, and data collection to real operation scenarios such as manufacturing, logistics, and services, and demand for electronic components is expected to migrate from project-based procurement to standardized, batch procurement. More notably, model training requires the advance accumulation of large amounts of real interaction data, and bodyless collection can reduce dependence on the number of Siasun Robot&Automation units; product forms such as EGO, UMI, WristCam, and Hub have begun to evolve from lab-built tools into standardized training hardware. Demand for electronic hardware does not have to wait entirely for large-scale shipment of Siasun Robot&Automation bodies, and may form a release rhythm of "training demand first, body demand following." Models are accelerating their migration downward, and Physical AI is further integrating into the broader edge-side computing upgrade cycle. The objects of MLPerf standard testing have further expanded from traditional vision CNNs to LLMs, generative AI, and Agents, while smartphones and PCs have taken the lead in validating local large models entering large-scale terminals; Siasun Robot&Automation VLA is simultaneously migrating to the edge side, and local models have begun to directly undertake action output tasks. As LLMs, VLMs, and Agents run locally, model weights, context, KV Cache, and local knowledge bases continue to increase, and edge-side hardware constraints further expand from computing power to Memory capacity, bandwidth, and Storage; at the same time, small models, quantization compression, and deployment toolchains continue to mature, lowering the computing power, power consumption, and storage thresholds for complex AI to enter terminals, making it possible for AI capabilities to further spread from high-performance PCs, smartphones, and Siasun Robot&Automation to industrial equipment, AIoT, wearables, and smart homes. Risk warnings Physical AI model capability improvement falls short of expectations; Siasun Robot&Automation industrialization and terminal shipments fall short of expectations; Physical AI data collection and training hardware demand falls short of expectations; edge-side AI penetration and model local deployment progress fall short of expectations; edge-side hardware configuration upgrade magnitude falls short of expectations; rapid changes in technology routes and intensifying industry competition.