Google Unveils AI Breakthrough Cutting Memory Use by 6x, Pressuring Flash Stocks

This article first appeared on GuruFocus.

A sharp two-day selloff in memory-chip stocks is beginning to expose a more nuanced split within the artificial intelligence trade, following Alphabet’s Google (NASDAQ:GOOG) unveiling of its TurboQuant technique. Early market reactions suggest investors are starting to differentiate between segments of the memory ecosystem, rather than treating the space as a uniform beneficiary of AI growth. Shares of Samsung Electronics (SSNLF) and SK Hynix (HXSCL) both key suppliers of high-bandwidth memory used in Nvidia accelerators erased most of their earlier declines in Seoul, while Micron Technology (NASDAQ:MU) traded relatively steady in US pre-market activity. In contrast, losses extended among flash and storage-focused companies, including Kioxia, indicating a more targeted reassessment of demand exposure.

Analysts suggest the distinction may stem from how TurboQuant impacts different layers of the AI stack. The technique is said to improve inference efficiency by reducing memory usage and data movement, with Google indicating it can cut memory requirements for certain large language model processes by at least a factor of six. This has raised concerns that hyperscalers such as Meta Platforms could require less storage capacity over time, potentially weighing on demand for NAND and related components. However, commentary from analysts, including those at Morgan Stanley, indicates that core memory tied to GPU operations particularly high-bandwidth memory and DRAM may remain largely unaffected, as these are still required to store model weights and support compute-intensive workloads.

The divergence follows a period where flash and storage stocks significantly outperformed traditional DRAM leaders during the AI-driven rally, with names like Sandisk and Kioxia posting substantial gains in recent months. As volatility picks up amid broader macro concerns, including inflation pressures linked to geopolitical tensions, investors appear to be rotating more selectively across subsectors on incremental developments. Even so, some market participants continue to frame the reaction as potentially short-term, suggesting that the buildout of AI infrastructure spanning compute and advanced memory could unfold over years and decades, leaving room for longer-term demand to stabilize even as near-term expectations adjust.