全球半导体：英伟达采用推理上下文内存存储，NAND 供应短缺预计进一步加剧-Global Semiconductors NAND Supply Shortage Expected to Deepen Further as Nvidia Adopts Inference Context Memory Storage

Summary of Key Points from the Conference Call Industry Overview - Industry: Global Semiconductors - Focus: NAND Supply and AI Storage Solutions Core Insights 1. Nvidia's Adoption of ICMS: Nvidia announced the adoption of Inference Context Memory Storage (ICMS) for its Vera Rubin platform to address memory bottlenecks in large-scale inference computations. This new architecture will utilize 16TB TLC SSDs and offload Key-Value (KV) Cache to more scalable storage solutions to enhance AI capabilities [1][2] 2. Projected NAND Demand Increase: The implementation of ICMS is expected to require an additional 1,162TB SSD NAND per Vera Rubin server. The projected demand for NAND in 2026 and 2027 is estimated to reach 34.6 million TB and 115.2 million TB respectively, which represents 2.8% and 9.3% of global NAND demand for those years [1][5] 3. NAND Supply Shortage: The demand increase due to ICMS adoption is anticipated to exacerbate the existing NAND supply shortage, indicating a significant upward pressure on NAND prices and availability in the market [1][6] 4. Performance Improvements: By offloading KV Cache, Nvidia aims to achieve up to 5x higher tokens per second, 5x better power efficiency, and reduced latency, which are critical for enhancing AI performance [2] 5. New Architectural Layer: ICMS introduces a new layer between local SSDs and shared enterprise storage, facilitating faster data processing and improved coordination with High Bandwidth Memory (HBM) to boost overall AI performance [4] Beneficiaries - Key Beneficiaries: The companies expected to benefit from the increased NAND demand include Samsung Electronics, SK Hynix, SanDisk, Kioxia, and Micron, as they are major suppliers in the NAND market [6] Additional Insights - KV Cache Explanation: KV Cache is a memory optimization mechanism in transformer models that stores previously computed key-value pairs to avoid redundant calculations, which is essential for efficient AI processing [3] - Market Implications: The announcement of Nvidia's ICMS adoption is viewed as a positive catalyst for NAND suppliers, indicating a robust growth opportunity in the semiconductor sector driven by AI advancements [6] This summary encapsulates the critical points discussed in the conference call regarding the semiconductor industry, particularly focusing on the implications of Nvidia's new storage architecture and its impact on NAND supply and demand dynamics.