DeepSeek's Memory-Efficient AI Model Pressures Samsung, SK Hynix

Memory Stocks Tumble on Efficiency Claims

Samsung Electronics and SK Hynix shares dropped more than 3% in Seoul trading Friday after Chinese AI startup DeepSeek announced its latest model requires significantly less high-bandwidth memory to operate. The decline contrasts with U.S. memory stocks Micron Technology and SanDisk, which remained in positive territory during overnight trading.

DeepSeek disclosed Thursday that its V4.1 Flash model employs a more efficient architecture that reduces its KV cache, cutting the amount of HBM and solid-state storage needed to run the system. The startup unveiled the model at substantially lower prices and plans to retire its V4-Pro version on September 14, automatically routing inference tasks to the new Flash architecture.

The V4.1 Flash reportedly activates only a small portion of its 552-billion-parameter model at any given time, enabling the reduced memory footprint.

Why It Matters

High-bandwidth memory has become a critical bottleneck in AI infrastructure, with costs rising sharply amid supply constraints. The technology powers the GPUs and accelerators used to train and deploy large language models, driving a prolonged shortage that has boosted earnings expectations for Samsung, SK Hynix, and Micron. If AI models can achieve comparable performance with substantially less memory—and if DeepSeek's architecture proves replicable—a key demand driver for the memory industry could weaken even as AI adoption accelerates. That scenario would force investors to recalibrate growth assumptions for companies that have positioned HBM as a multi-year tailwind.

Market Reaction Extends Beyond Korea

The announcement pressured other AI-related stocks in Asia. Shares of MiniMax Group and Z.AI fell more than 8% in Hong Kong trading Friday.

For Samsung and SK Hynix, the selloff adds to a difficult stretch. Both companies remain more than 25% below their record highs following a steep July decline. Retail investors have sold over $10 billion of the two stocks this month, according to reports cited by the source.

Divided Investor Sentiment

Retail traders showed mixed reactions to the news. Some view the decline as a buying opportunity, arguing that hyperscale cloud providers possess substantial cash reserves and continue to require maximum memory capacity regardless of efficiency gains. Others flagged concerns about reduced AI flash demand as models become more resource-efficient.

The divergence reflects broader uncertainty about how quickly architectural improvements might offset volume growth in AI infrastructure spending. Memory manufacturers have benefited from AI workloads requiring ever-larger memory pools, but efficiency breakthroughs could alter that trajectory.

These details were first reported by AI Watch.