中邮证券：英伟达(NVDA.US)Vera Rubin重构AI存储层级 NAND有望成为通胀品

Core Insights - Nvidia's new AI supercomputing platform, Vera Rubin, has officially entered full production, aiming to address the "memory wall" issue by restructuring the HBM-DRAM-NAND storage architecture [1][2] - The platform features significant upgrades, including the integration of HBM4 memory, which enhances bandwidth and processing capabilities, potentially leading to improved profit margins for manufacturers [1][3] Group 1: Storage Architecture - Vera Rubin redefines the storage pyramid by introducing a new memory architecture that significantly increases the processing capacity of context memory, allowing for a fivefold increase in token processing per second [2] - The integration of BlueField-4 processors within the Vera Rubin platform is designed to manage KVCache effectively, enhancing overall system performance [5] Group 2: HBM and DRAM Upgrades - The Rubin GPU is equipped with HBM4, which has doubled the interface width compared to HBM3e, achieving a bandwidth of 22TB/s, nearly tripling that of the previous generation [3] - Vera CPU has been upgraded to LPDDR5X, providing a high bandwidth of 1.2TB/s while maintaining low power consumption, which supports efficient multi-model execution and reduces data movement costs [4] Group 3: NAND Market Outlook - The introduction of the BlueField-4 driven context memory storage platform is expected to correlate linearly with GPU quantities, indicating a potential inflationary trend in NAND prices due to rising demand from cloud service providers and AI applications [5] - The industry anticipates a double-digit percentage increase in NAND prices throughout 2026, driven by the growing needs of cloud services and AI applications [5] Group 4: Investment Recommendations - The report suggests focusing on key players in the storage supply chain, including major overseas companies like SK Hynix, Samsung, Micron, SanDisk, and Kioxia, as well as domestic firms such as Shannon Semiconductor, Demingli, Zhaoyi Innovation, Purun Co., and Tongyou Technology [6]