According to a research report released by Orient Securities, NVIDIA (NVDA.US) unveiled its Inference Context Memory Storage Platform at the CES 2026 conference. This platform is a POD-level AI-native storage infrastructure, with the core objective of creating a new memory layer optimized for inference between GPU memory and traditional storage, thereby supporting the long-term operation of AI. During the inference process of large AI models, high-frequency data access is required to achieve high-quality content generation, which will significantly alter storage architectures and increase the demand for storage chips. The current supply shortage in the storage market persists, while overseas storage giants may have limited progress in expanding production capacity for general-purpose storage, presenting a historic opportunity for domestic storage manufacturers to expand production and increase their market share. The main viewpoints of Orient Securities are as follows.
At the event, NVIDIA founder and CEO Jensen Huang delivered a speech at CES 2026, introducing the NVIDIA Vera Rubin POD AI supercomputer, NVIDIA Spectrum-X Ethernet co-packaged optics, the NVIDIA Inference Context Memory Storage Platform, and the NVIDIA DGX SuperPOD based on DGX Vera Rubin NVL72, among other products.
NVIDIA's launch of the Inference Context Memory Storage Platform aims to build an AI-native storage infrastructure. The newly released platform is a POD-level AI-native storage infrastructure, with its core goal being to establish a new memory layer specifically optimized for inference, situated between GPU memory and traditional storage, to support sustained AI operations. From a technical perspective, the platform is the result of co-design, comprising several key elements: (1) BlueField-4, which is responsible for accelerating the management and access of contextual data at the hardware level, thereby reducing data movement and system overhead. (2) Spectrum-X Ethernet, which provides a high-performance network supporting high-speed data sharing based on RDMA. (3) Software components such as DOCA, NIXL, and Dynamo, which optimize scheduling at the system level, reduce latency, and enhance overall throughput. Through this co-design, the platform can extend context data originally held in GPU memory to an independent, high-speed, shareable "memory layer." This approach alleviates pressure on the GPU while enabling rapid sharing of contextual information across multiple nodes and multiple AI agents.
In terms of practical performance, NVIDIA states that this method can increase the number of tokens processed per second by up to 5 times and achieve equivalent levels of energy efficiency optimization.
The bottleneck in AI inference is shifting from computation to contextual storage, suggesting that demand for storage chips is poised for sustained high-speed growth. During his speech, Jensen Huang emphasized that the bottleneck for AI inference is transitioning from computation to contextual storage. As model scales increase and user usage grows, AI processing of complex tasks requiring multi-turn conversations and multi-step reasoning generates vast amounts of contextual data. Traditional network storage is inefficient for handling short-term context, necessitating a restructuring of AI storage architectures. Some investors still underestimate the extent to which AI will drive demand for storage chips. Orient Securities has previously stressed that the high-frequency data access required during the inference process of large AI models for high-quality content generation will lead to significant changes in storage structures and boost demand for storage chips. Looking ahead, AI is expected to evolve from a "single-session chatbot" into an intelligent collaborator that understands the real world, performs continuous reasoning, and utilizes tools to complete tasks. This evolution will require continuously expanding context capacity and accelerating cross-node sharing, thereby driving high-speed growth in demand for storage chips.
The supply shortage in the storage market persists, highlighting the importance of opportunities for the localization of the storage industry chain. The ongoing supply-demand imbalance in the storage market, coupled with potentially limited capacity expansion for general-purpose storage by overseas storage giants, presents a historic opportunity for domestic storage manufacturers to expand production and increase their market share. On the technological front, in the DRAM sector, ChangXin Memory Technologies (CXMT) launched its DDR5 product in November 2025, achieving mainstream technical parameters such as peak data rates comparable to international leading standards. In the NAND sector, Yangtze Memory Technologies Co., Ltd. (YMTC) independently developed the Xtacking architecture, enabling leapfrog development in 3D NAND technology. Regarding IPO progress, CXMT's IPO application has been accepted, and its parent company, ChangCun Group, completed its corporatization reform in September 2025. Orient Securities believes that after advancing financing, these two major memory players are expected to achieve significant production capacity expansion, and the entire upstream and downstream industry chain is poised to benefit deeply.
Related investment targets include domestic semiconductor equipment companies such as AMEC (688012.SH), Jingzhida (688627.SH), Jingyi Equipment (688652.SH), Leadmicro (688147.SH), Piotech (688072.SH), and NAURA (002371.SZ). Domestic packaging and testing companies include Shenzhen Kaifa Technology (000021.SZ), Huicheng (688403.SH), and Tongfu Microelectronics (002156.SZ). Supporting logic chip manufacturers include Nexchip (688249.SH). Companies focusing on end-side AI storage solutions include GigaDevice (603986.SH) and Ingenic Semiconductor (300223.SZ). Beneficiaries of storage technology iteration include Montage Technology (688008.SH) and Union Memory (688449.SH). Domestic storage solution providers include Longsys (301308.SZ), DMEGC (001909.SZ), BIWIN (688525.SH), and Lenovo Group (00992).
Risk warnings include potential shortfalls in AI implementation, slower-than-expected technological iteration speeds, and delays in the progress of localization efforts.
