Orient: NVIDIA Corporation (NVDA.US) launches an inference context memory storage platform as AI storage demands continue to expand.

Orient issued a research report stating that NVIDIA Corporation (NVDA.US) launched the Inference Context Memory Storage Platform at the CES 2026 conference. This platform is a POD-level AI native storage infrastructure, with the core goal of creating a new memory layer optimized for inference between GPU memory and traditional storage to support long-term AI operations. In the process of large-scale AI model inference, frequent access to data is required to achieve high-quality content generation, leading to significant changes in storage structures and an increased demand for storage chips. The current storage supply is inadequate, and the expansion progress of overseas storage giants in general storage may be limited, providing a historic opportunity for domestic storage manufacturers to expand and increase their market share. Key points from Orient are as follows: Event: NVIDIA Corporation founder and CEO Huang Renxun delivered a speech at CES 2026, announcing the launch of NVIDIA Vera Rubin POD AI supercomputer, NVIDIA Spectrum-X Ethernet encapsulated optical devices, NVIDIA Inference Context Memory Storage Platform, and the NVIDIA DGX SuperPOD based on DGX Vera Rubin NVL72. NVIDIA Corporation introduces the Inference Context Memory Storage Platform to create an AI native storage infrastructure The Inference Context Memory Storage Platform released by NVIDIA Corporation this time is a POD-level AI native storage infrastructure. The core goal is to create a new memory layer optimized for inference between GPU memory and traditional storage, to support long-term AI operations. Technically, this platform is the result of a collaborative design, including: (1) BlueField-4 is responsible for accelerating the management and access of context data at the hardware level to reduce data movement and system overhead. (2) Spectrum-X Ethernet provides high-performance networking, supporting fast data sharing based on RDMA. (3) Software components such as DOCA, NIXL, and Dynamo are responsible for optimizing scheduling, reducing latency, and improving overall throughput at the system level. Through collaborative design, this platform can extend the context data originally stored in GPU memory to an independent, high-speed, and shareable "memory layer", enabling quick sharing of context information among multiple nodes and AI agents while releasing GPU pressure. In terms of actual performance, NVIDIA Corporation stated that using this method can increase the number of tokens processed per second by up to 5 times, while achieving equivalent energy efficiency optimization. The bottleneck of AI inference is shifting from computation to context storage, and the demand for storage chips is expected to continue to grow rapidly Huang Renxun emphasized in his speech that the bottleneck of AI inference is shifting from computation to context storage. As model sizes increase and user usage grows, AI processing for complex tasks that require multiple rounds of dialogue and multi-step inference will generate large amounts of context data. Traditional network storage is inefficient for short-term contexts, and AI storage architecture needs to be restructured. Some investors may still underestimate the impact of AI on the demand for storage chips. Orient previously emphasized that in the process of large-scale AI model inference, frequent access to data is required to achieve high-quality content generation, leading to significant changes in storage structures and an increased demand for storage chips. Looking ahead, AI is expected to evolve from "chatting Siasun Robot&Automation" in a one-time dialogue to intelligent cooperation for understanding the real world, continuous inference, and task completion using tools, which requires continuously expanding context capacity and accelerating cross-node sharing, thus driving the rapid growth of storage chip demand. Storage supply remains inadequate, with a focus on the opportunity for domestication of the storage industry chain The storage supply continues to be inadequate, and the progress of overseas storage giants in general storage expansion may be limited, providing a historic opportunity for domestic storage manufacturers to expand and increase their market share. In terms of technology, in the DRAM field, Longxin Technology launched DDR5 products in November 2025, reaching international first-line levels in key technical parameters such as peak rate; in the NAND field, Changjiang Storage's independently developed Xtacking architecture achieved a breakthrough in 3D NAND technology. In terms of IPO progress, Longxin Technology's IPO has been accepted, and Changjiang Storage's parent company Changfeng Group completed the shareholding system reform in September 2025. Orient believes that both companies are expected to achieve significant expansion volumes after financing in the future, benefiting deeply from the entire industry chain. Related targets Chinese semiconductor equipment companies such as Advanced Micro-Fabrication Equipment Inc. China (688012.SH), Shenzhen SEICHI Technologies (688627.SH), Beijing Jingyi Automation Equipment (688652.SH), Jiangsu Leadmicro Nano Technology (688147.SH), Piotech Inc. (688072.SH), NAURA Technology Group (002371.SZ), etc. Domestic semiconductor testing and packaging companies such as Shenzhen Kaifa Technology (000021.SZ), Union Semiconductor (688403.SH), TongFu Microelectronics (002156.SZ), etc. Supporting logic chip manufacturers like Nexchip Semiconductor Corporation (688249.SH); GigaDevice Semiconductor Inc. (603986.SH) and Ingenic Semiconductor (300223.SZ) focusing on edge AI storage solutions. Beneficiaries of storage technology iterations like Montage Technology (688008.SH), Maxio Technology (688449.SH), etc. Domestic storage solution manufacturers such as Shenzhen Longsys Electronics (301308.SZ), Shenzhen Techwinsemi Technology (001909.SZ), Biwin Storage Technology (688525.SH), LENOVO GROUP (00992), etc. Risk Warning AI adoption falls short of expectations, technological iteration speeds are lower than expected, and domestication progress is slower than expected.