After two consecutive days of sharp declines, the storage sector begins to "diversify": Samsung and SK Hynix stabilize, while "flash memory" continues to decline.

A breakthrough algorithm from Google that boosts AI inference efficiency is triggering a structural split in the memory chip industry: flash-memory companies’ stock prices are under sustained pressure, while related high-bandwidth memory (HBM) plays have largely stabilized.

The U.S. stock storage-chip sector has been under pressure for two consecutive trading days. On Thursday, SanDisk (Sandisk) plunged 11% at the U.S. market close. On Friday, selling sentiment briefly spread into the Asia-Pacific markets, with Korean memory stocks continuing to slide. SK Hynix fell by more than 5% during the trading day, and Samsung Electronics dropped by more than 4%.

However, thanks to HBM’s core position in the AI training arena—enabling Samsung Electronics and SK Hynix to quickly stabilize after supplying high-bandwidth memory to Nvidia AI accelerators—Samsung has essentially recovered all of its losses, and SK Hynix’s decline has narrowed to 1%, showing resilience that is better than that of flash-memory vendors. By contrast, stocks such as Kioxia, which had gained more than 600% cumulatively over the prior several months, continued to fall.

In a research note, Morgan Stanley analyst Tiffany Yeh said that Google’s “TurboQuant” technology significantly improves AI inference efficiency by compressing memory usage and reducing data movement, but “does not weaken demand for core memory chips such as HBM.” The market is gradually realizing that the threat this technology poses to flash-memory companies is far greater than what it poses to the HBM space.

The flash-memory segment is hit first, with earlier gains already largely unwound

Spurred by expectations of widespread AI adoption, flash-memory and storage-product manufacturers have attracted a large influx of investors in recent months. Since late August, SanDisk’s share price has at one point risen by more than 1,000% cumulatively, and Kioxia is up more than 600%, far outperforming traditional storage giants such as Samsung Electronics, SK Hynix, and Micron Technology.

However, market sentiment flipped this week. After investors realized the far-reaching impact of Google’s technology breakthrough, they first sold the shares of the above companies. Google said its TurboQuant algorithm can compress by at least one-sixth the amount of memory required for specific stages when running large language models, thereby significantly lowering overall operating costs for AI. The market is worried that this move will reduce storage procurement demand from hyperscale data center operators such as Meta, which in turn will weigh on storage chip prices in the smartphone and consumer electronics sectors.

In a research note, Bloomberg Intelligence analyst Jake Silverman pointed out: “Since model weights need to be stored in GPU memory, HBM demand and the DRAM produced by Micron are very likely not affected. By comparison, NAND flash demand will face a more profound long-term impact.”

HBM demand logic remains unchanged; Samsung and SK Hynix stabilize

Against the backdrop of the flash-memory sector’s continued slide, HBM-related stocks have shown stronger resilience in this round of selling. Samsung Electronics has already recovered all of its losses on Friday, and SK Hynix’s share price has also basically returned to pre-selloff levels.

Analysts believe there is a clear rationale behind this split. During the AI large language model training stage, GPU demand for HBM is highly concentrated, and TurboQuant optimizes memory efficiency for the inference stage—it does not touch the core HBM demand on the training side. Samsung and SK Hynix became market darlings earlier in the AI investment boom precisely because of their HBM products, and this advantage has not been shaken by the current algorithm breakthrough.

Analyst: Near-term volatility may not drown out the industry’s long-term narrative

It is also worth noting that this selloff in storage chips is occurring amid a macro backdrop in which overall technology-stock valuations are under review. Inflation concerns stemming from the Middle East situation have made the market more cautious about high-valuation stocks. Investors are highly sensitive to new information, and profit-taking could trigger at any moment.

SGMC Capital Chief Investment Officer Ed Gomes said that the hardware demand driving the deployment and application of AI technology is a long-term structural proposition—“it will keep evolving over years, even decades, rather than over a few days or weeks.” He believes the selloff around TurboQuant is “short-term noise that provides a good buying opportunity for quality names.”

However, analysts also noted that, amid ongoing iterations of AI efficiency algorithms, the internal divergence in the storage-chip industry may further intensify. Whether flash-memory companies can return to the earlier high-growth expectations remains to be seen.