Shares of SK Hynix plunged 8.55% to €1,070 on its Düsseldorf listing as investors digested a familiar shock from an unfamiliar source: a Chinese AI startup just showed it can do more with less memory. The question is whether this is a genuine threat to the memory supercycle or another panic that the market will walk back in months.

  • The Trigger: A New AI Model Needs Far Less of SK Hynix's Most Profitable Product

DeepSeek's latest model claims to cut high-bandwidth memory requirements by 75% and solid-state storage needs by 87.5% versus its predecessor. HBM — stacked memory chips placed right next to processors to move data at extreme speed — helped turn SK Hynix, Samsung, and Micron into some of the biggest semiconductor trades of 2026. Any model that shrinks the need for HBM strikes directly at the revenue stream driving SK Hynix's record margins. In Q2 2026, the company posted an operating margin of 76% on revenues of 79.3 trillion won.

  • The Selloff Echoes January 2025 — When Panic Faded and Stocks Recovered

When DeepSeek's earlier model rattled markets in January 2025, SK Hynix fell as much as 12% in a day. Investors initially feared more efficient models would weaken hardware demand — instead, AI spending kept climbing, and the argument that lower costs stimulate greater usage won the year that followed. That playbook is exactly what bulls are citing again. One fund manager called this "sentiment-driven pressure rather than a lasting sector selloff," adding that "cheaper AI could drive greater usage, offsetting efficiency gains."

  • SK Hynix's Order Book Provides a Short-Term Buffer

Some analysts argue this "does not change the HBM or DRAM shortage," noting 2026 HBM output at all three major suppliers is described as sold out.

SK Hynix recently raised its shareholder return policy above 50% of cumulative free cash flow and announced a 40 trillion won buyback program. JPMorgan estimates a cumulative total shareholder return yield of 41.8% from 2026 through 2028.

  • The Real Risk Is 2027 and Beyond, Not Today's Contracts

The current order book is locked. The danger is what happens when contracts reset. DeepSeek has published a chart showing per-token memory requirements falling more than 400-fold since January 2024 — and says the same architecture is intended for larger models. If that efficiency trajectory holds across the industry, AI data centers will eventually need less memory per unit of work, pressuring both volumes and the premium pricing that funds SK Hynix's extraordinary profitability. The tension between a multi-year memory supercycle and AI models that need less memory per task will likely define trading in Korean chip stocks for the near term.