NVIDIA's (NVDA.US) newly released solution adopts a hierarchical concept at the architectural level, inspired by Intel SCM. The combination of BF4+DRAM+memory pooling represents NVIDIA's software-defined SCM, whereas existing potential solutions face various challenges related to cost and the stability of large-scale training. Logically, AI is accelerating the importance of all storage. Previously, the market lacked a consensus on the growth trajectory for DRAM and SSDs driven by AI demand. However, following this announcement, the firm believes AI is progressively reshaping the storage stack and quantifying the demand and capacity for SSDs. The primary views of Guolian Minsheng Securities are as follows:
At CES 2026, Jensen Huang's latest keynote introduced the new-generation AI chip architecture, NVIDIA Vera Rubin. Faced with the exponential growth in AI computing demands, NVIDIA officially launched this new architecture. The Vera Rubin platform is not merely an iteration of a single chip but a comprehensive rethinking of the entire stack, from transistors to system architecture. Huang showcased six new chips, including the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 Smart NIC, BlueField-4 DPU, and Spectrum-X 102.4T CPO.
NVIDIA contends that in an era of slowing Moore's Law, performance gains solely from process node advancements are limited. True breakthroughs originate from a systemic redesign of how components work together. In traditional computing architectures, data transfer often becomes a bottleneck. In Vera Rubin, by redesigning high-speed interconnects between chips, the total system communication bandwidth reaches "twice the total bandwidth of the global internet."
NVIDIA confirms that AI is reshaping the storage stack. Previously, the volume of KV Cache, driven by growing context memory, longer conversations, and the need to store more temporary knowledge in the wave of large models, could no longer be accommodated by traditional GPU high-bandwidth memory. Consequently, Vera Rubin's solution employs four BlueField-4 DPUs to directly manage a massive, shared, persistent, high-speed context memory pool of up to 150TB. This memory pool connects directly to all GPUs via an ultra-low latency intra-rack network, enabling dynamic allocation of up to 16TB of dedicated, high-speed context space per GPU.
NVIDIA is positioning itself as the definer and provider of full-stack computing infrastructure for the AI era. Currently, training a next-generation frontier model with 10 trillion parameters using Vera Rubin requires only one-quarter the cluster size needed for a Blackwell system, significantly accelerating the iteration cycle from research to product. Under equivalent power and space constraints, the AI computational throughput of a Vera Rubin data center is projected to be approximately 100 times that of a data center based on the Hopper architecture. The token generation cost for large-scale AI services is expected to drop to about one-tenth of current levels.
In the AI era, the importance of storage is being elevated. If each GPU adds 16TB of NAND, and each cabinet adds 1,152TB of NAND, then based on 140,000 cabinets, the NAND demand increases by 161 exabytes. This represents approximately 16% of the global total NAND demand forecast for 2025 and 54% of the enterprise SSD demand, potentially exerting a sustained influence on the NAND supply-demand balance.
Risk warnings: AI development may fall short of expectations; the commercial logic of AI may not meet expectations; storage demand may fall short of expectations.
Comments