2025中国AI算力大会：系统级集成崛起，AI基础设施进入软硬协同新阶段

Investment Rating - The report does not explicitly provide an investment rating for the industry Core Insights - The 2025 China AI Computing Power Conference marked a significant milestone in the transformation of domestic AI infrastructure, indicating a shift towards system-level integration and the prevalence of large-scale GPU cluster deployments [1][15] - The era of domestic AI clusters has begun, with a transition from card-level optimization to cluster-level scheduling, emphasizing the need for energy efficiency and infrastructure collaboration [2][16] - The core logic of the AI industry is shifting from model-centric to infrastructure-centric paradigms, highlighting the importance of system-level capabilities and the emergence of AI-native companies [3][20] Summary by Sections AI Infrastructure Development - The conference showcased breakthroughs in domestic AI computing, emphasizing the importance of system-level integration and the transition to large-scale deployments [1][15] - Companies like Zhonghao Xinying are accelerating the localization of TPU chips, focusing on high-performance architecture and inter-chip connectivity, with their first-generation TPU nearing mainstream global performance standards [2][17] Technological Advancements - Standardization of low-precision floating-point computation (FP8) has been achieved, significantly improving training efficiency and reducing power consumption, becoming essential for deploying 10,000-card training clusters [5][19] - AI training is evolving towards heterogeneous mixed-architecture systems (GPU + TPU + NPU + CPU), driving traditional HPC stacks to fully cloudify [5][19] Competitive Landscape - System-level capabilities will define future competitive advantages, with players excelling in scheduling systems and energy-efficient designs building stronger economic moats [3][20] - The dual opportunity of "Domestic + System" is emerging, where domestic substitution encompasses complete system-level solutions, fostering holistic ecosystems [3][20] AI-native Workloads - The rise of AI-native workloads, led by AI Agents, is transforming enterprise services and process automation, necessitating compute platforms to support multi-model parallelism and task scheduling [5][19]