申万宏源研究晨会报告-20250929

Core Insights - The report highlights the significant advantages of ASIC over GPU in terms of cost-effectiveness and energy efficiency, marking a turning point for ASIC development [2][12] - The increasing penetration of AI is driving a surge in inference demand, expanding the market space for ASICs [3][12] - Domestic cloud providers are making strides in self-developed ASICs, indicating a strong demand in the Chinese AI cloud market [12][13] Summary by Sections ASIC vs. GPU - ASICs are specialized chips tightly coupled with downstream applications, focusing on specific needs like text and video inference, while GPUs are general-purpose chips covering a broader range of applications [2][12] - The energy efficiency of Google's TPU v5 is 1.46 times that of NVIDIA's H200, and Amazon's Trainium2 reduces training costs by 40% compared to GPU solutions [2][12] Market Growth and Demand - The global AI ASIC market is projected to reach $125 billion by 2028, with significant contributions from major clients [3][12] - The demand for inference computing is directly linked to throughput, with ChatGPT's weekly active users reaching 700 million by July 2025, driving the need for increased computational power [3][12] ASIC Design Services - ASIC design requires a high level of specialization, with major service providers like Broadcom and Marvell leading the market [3][12] - Broadcom's collaboration with Google on TPU has been pivotal, leveraging a comprehensive IP system and advanced packaging technologies [3][12] Domestic Developments - Leading Chinese cloud providers are achieving results in self-developed ASICs, with significant orders and advancements in technology [12][13] - The trends of PD separation and super nodes are emerging as key developments in the domestic ASIC landscape [12][13] Industry Outlook - The report anticipates a robust growth trajectory for the ASIC market, driven by increasing AI applications and domestic innovation [12][13]