ASIC发力，GPU地位松动

Core Insights - The effectiveness and cost structure of AI are determined not just by general-purpose GPUs but by the ability to create application-specific integrated circuits (ASICs) tailored for specific workloads [2][3] - The demand for ASICs is surging as AI model scales and application scenarios expand, making them a critical component in AI infrastructure [2][3] ASIC Reshaping the Computing Landscape - ASICs offer a significantly better performance-to-power ratio and long-term cost advantages compared to general-purpose chips due to their highly customized hardware design [3] - By 2027, global shipments of AI server ASICs are expected to triple compared to 2024, with shipments surpassing 15 million units by 2028, exceeding those of data center GPUs [3][4] Market Dynamics and Competitive Landscape - Google’s TPU v7e has entered mass production, while Microsoft’s Maia series and Meta’s MTIA are being deployed at scale, indicating a significant growth in shipments by 2027 [4] - Taiwanese ASIC companies are transitioning from passive design contractors to core partners in system co-design, which is crucial for the success of complex AI ASIC projects [4] - Broadcom is currently the preferred supplier for cloud AI ASICs but faces increasing competition from emerging players, particularly a strategic alliance between Google and MediaTek [4]