记忆晶片股出现为期两天的抛售，使AI相关定位出现新的分化。三星电子股份有限公司(Samsung Electronics Co.)与SK Hynix Co.（其高频宽记忆体用于Nvidia Corp.的AI加速器）以及Micron Technology Inc.在首尔及美国盘前交易中大多已回补，而由Kioxia Holdings Corp.领先的快闪记忆体（flash）同业继续下跌。市场如今在区分不同记忆体类别，而非把这个产业视为同质的AI受益者。

分析师表示，Google的TurboQuant透过降低记忆体使用与资料移动来提升AI推论效率，对flash/储存需求威胁最大。在上一次AI热潮中，投资者更偏好flash个股：Sandisk Corp.股价自上月8月底以来仍上涨超过1,000%(即10倍)，Kioxia上涨超过600%，超过了Samsung、SK Hynix和Micron。抛售自本周Google公布突破后开始，现已集中在该组别，而以HBM为主的标的则相对撑住。

Google表示TurboQuant可将LLM推论某一环节所需记忆体至少降低6倍（原始表述为factor of six），这可能压低Meta Platforms Inc.等超大规模资料中心运营商的成本，并进一步压低手机与消费电子中的组件价格。Bloomberg Intelligence分析师Jake Silverman认为，由于模型权重必须存放于GPU记忆体，HBM和Micron的需求大致不受影响；但NAND需求可能承受更深的长期影响。受伊朗战争带动的通膨忧虑影响，AI科技估值敏感，投资者在获利了结与逢低买入之间轮转，SGMC Capital的Ed Gomes也认为这次下跌只是短期噪音，而AI硬体议题是跨越多年、十年的长期主题。

A two-day sell-off in memory-chip shares is exposing a new split in AI-related positioning. Samsung Electronics Co. and SK Hynix Co., whose high-bandwidth memory is used in Nvidia Corp. AI accelerators, and Micron Technology Inc. mostly recovered in Seoul and in U.S. premarket trading, while flash-memory peers led by Kioxia Holdings Corp. continued to fall. The market is now distinguishing between memory categories rather than treating the sector as one uniform AI beneficiary.

Analysts say Google’s TurboQuant, which improves AI inference by reducing memory use and data movement, is most threatening to flash and storage demand. In the latest AI rush, investors had favored flash players: Sandisk Corp. shares were still up more than 1,000% (equivalent to 10x) since late August, and Kioxia had gained over 600%, outpacing Samsung, SK Hynix, and Micron. The selloff began this week after Google’s announcement and is now concentrated in that group, while HBM-focused names held up.

Google said TurboQuant can reduce the memory required for a specific part of LLM inference by at least a factor of six, which could lower costs for hyperscalers such as Meta Platforms Inc. and pressure component prices for smartphones and consumer electronics. Bloomberg Intelligence analyst Jake Silverman argued HBM demand should remain supported because model weights must stay in GPU memory, but NAND demand may face deeper long-term effects. With AI-tech valuations still vulnerable to inflation fears linked to the Iran war, investors are rotating between profit-taking and dip-buying, and SGMC Capital chief Ed Gomes said this decline is short-term noise in an AI-hardware story expected to play out over years and decades. (Key numbers: 6)