Zhi Tong Cai Jing·2026-01-13 08:37Group 1 - The core viewpoint of the report is that NVIDIA's new AI chip architecture, NVIDIA Vera Rubin, represents a comprehensive rethinking of system architecture, addressing the exponential growth in AI computing demands and redefining storage stack requirements [1][2][3] - NVIDIA's Vera Rubin platform includes six new chips designed to enhance system communication bandwidth, achieving a bandwidth that is twice that of the global internet total bandwidth, thus overcoming traditional data transfer bottlenecks [1][3] - The introduction of BlueField-4DPU allows for the management of a shared, persistent, high-speed context memory pool of up to 150TB, directly connected to GPUs, enabling dynamic allocation of up to 16TB of dedicated context space per GPU [2][3] Group 2 - The demand for NAND storage is projected to increase significantly, with an estimated 161EB needed by 2025, which would account for approximately 16% of the global NAND demand and 54% of enterprise SSD demand [4] - The Vera Rubin architecture allows for training a next-generation model with 100 trillion parameters using only a quarter of the cluster size required by the previous Blackwell system, greatly accelerating the research-to-product iteration cycle [3][4] - The AI computing throughput of a Vera Rubin data center is expected to be about 100 times that of a data center based on the Hopper architecture, significantly reducing the token generation cost for large-scale AI services to about one-tenth of the current level [3]