存储“涨声”再起：一季度NAND闪存涨幅预期超40%

Core Viewpoint - The price increase of NAND flash memory is driven by the rising demand from AI applications, particularly in large-scale inference processes, leading to significant upward revisions in price forecasts by market research firms [1][2][3]. Group 1: Price Forecasts and Market Trends - Samsung Electronics raised NAND flash contract prices by over 100% in January, prompting multiple market research firms to revise their price forecasts upward [1]. - TrendForce increased its first-quarter NAND flash price growth forecast from 33-38% to 55-60%, indicating potential for further upward adjustments [1]. - Counterpoint predicts NAND flash prices will rise by over 40% in the current quarter [1]. Group 2: AI Demand and Storage Architecture - The surge in NAND flash demand is primarily attributed to AI applications, particularly in retrieval-augmented generation (RAG) which enhances the accuracy of large language models [1][2]. - The transition from training to large-scale inference in generative AI has led to increased demand for NAND flash, as systems require high-speed access to vast amounts of data [2][3]. - The need for high-frequency access to context data during inference has resulted in a shift towards a storage architecture that includes HBM, DRAM, and NAND [3]. Group 3: Supply Constraints and Future Outlook - The global NAND flash production capacity is concentrated among a few major players, including Samsung, SK Hynix, and Micron, with investments in NAND lagging behind HBM and advanced DRAM [4][5]. - Morgan Stanley forecasts a 40% year-over-year increase in average NAND sales prices by 2026, with only a slight decline expected in 2027 [5]. - The introduction of High Bandwidth Flash (HBF) aims to address the limitations of traditional NAND SSDs, providing higher bandwidth and capacity suitable for AI inference applications [5][6]. Group 4: Technological Advancements - HBF combines 3D NAND flash with high-bandwidth interface technology, offering 8 to 16 times the capacity of traditional HBM, making it a competitive solution for AI applications [5][6]. - The industry is moving towards a multi-layer architecture of "DRAM cache + HBF acceleration + NAND mass storage," which is expected to alleviate supply-demand imbalances and drive growth [6].