Investment Rating - The report does not explicitly state an investment rating for the industry or specific companies involved Core Insights - The industry is witnessing a shift towards "CPU-first" on-device AI, emphasizing efficiency and practicality, with Arm's SME2 architecture reallocating tasks from GPU/NPU to CPU, significantly reducing latency for real-time applications [2][13] - Arm is transitioning from traditional IP licensing to a platform-based delivery model, exemplified by the Lumex Mobile Computing Subsystem, which integrates hardware and software solutions, reducing design cycles and costs for OEMs [3][14] - The mobile experience is improving, with notable performance gains in processors, but memory bandwidth and cache hit rates remain bottlenecks for large model inference, necessitating a collaborative architecture involving CPU, GPU, and NPU [4][15] - The rapid development of China's on-device AI ecosystem is driven by diverse application scenarios and strong industry collaboration, with companies like vivo and Alibaba Cloud leading the way [5][16] Summary by Sections Event Overview - On September 10, 2025, Arm held the "Arm Unlocked" event, showcasing the Armv9.3 architecture and SME2 in mobile platforms, and introduced the Arm Lumex Mobile Computing Subsystem [1][12] Commentary on AI Architecture - The "CPU-first" approach for on-device AI reflects a trend towards efficiency, allowing for significant reductions in latency for applications like camera and voice processing, with zero code modification required for existing applications [2][13] Structural Changes in Delivery Model - Arm's shift to platform-based delivery reduces design cycles and system integration costs, enhancing product development timelines and market responsiveness for device manufacturers [3][14] Mobile Experience Insights - Processor performance improvements are practical, with significant gains in C1-Ultra and G1-Ultra, but memory bandwidth remains a critical issue for large models, necessitating a collaborative architecture among CPU, GPU, and NPU [4][15] Ecosystem Development in China - The growth of China's on-device AI ecosystem is supported by a complete closed loop from chip architecture to application, with a strong user base facilitating model iteration and testing [5][16]