Ge Long Hui·2025-09-30 03:45Group 1 - The report from Dongfang Securities highlights that process technology upgrades drive chip performance improvements, while advanced packaging serves as another key driver for enhancing chip capabilities [1] - In the context of slowing process technology upgrades, increasing the die area can enhance transistor count and computational power, with Nvidia's H100 die area nearing the reticle limit of approximately 800-900 mm² [1] - Nvidia's B200 adopts advanced packaging with dual die integration, achieving 208 billion transistors in a single package, which is more than double the 80 billion transistors in the H100 [1] - According to Nvidia's roadmap, the Rubin Ultra will integrate four dies in a single package, targeting a computational power of 100PF FP4 per card [1] Group 2 - System assembly is emerging as a new driver for AI server performance enhancement, as wafer manufacturing and advanced packaging may not keep pace with the growing demand for AI computing power [2] - The number of GPUs in AI servers is expected to increase from 8 per server to 72 per cabinet, with projections for the VR Ultra NVL576 cabinet in 2027 to support 144 GPUs, each with four die, totaling 576 die [2] - The increase in GPU count raises cooling requirements and complicates system assembly, exemplified by the production ramp-up challenges faced by GB200 NVL72 due to assembly difficulties [2] - Leading companies in the industry are likely to benefit from the rising technical barriers and improved competitive environment in system assembly [2] Group 3 - In terms of investment targets, companies related to AI server system assembly are maintained, including Industrial Fulian, which has significantly optimized GB200 series product testing and reduced cabinet debugging time [3] - Industrial Fulian has expanded capacity globally and introduced fully automated assembly lines, expecting strong growth in GB200 shipments, primarily driven by large North American cloud service providers [3] - Haiguang Information's merger with Zhongke Shuguang is anticipated to create vertical integration capabilities encompassing CPU, DCU, and system assembly [4] - Lenovo is expected to launch various servers based on Blackwell Ultra starting in the second half of 2025, as indicated by Nvidia [4] - Huaqin Technology, a core ODM supplier for domestic internet firms' AI servers, benefits from the capital expenditure expansion of downstream cloud companies [4]