SK Hynix: A November Deadline, a Talent Drain to China, and the Case for 3D DRAM

SK Hynix fell 17% in seven sessions as foreign investors sold heavily, while CXMT hiring and US export deadlines cloud the memory chip outlook.

Seoul — For much of the past two years, the bull case for memory chips rested on a single comfortable assumption: demand for high-performance storage would outrun supply indefinitely, and the incumbents' technological lead was untouchable. That conviction is now being stress-tested in real time, and SK Hynix sits at the center of the examination.

The stock traded at 1,426,000.00 KRW on Wednesday, down 1.2% on the day and 17% over the past seven sessions. The retreat follows a wave of foreign selling that has hit South Korea's leading memory makers with particular force. According to Dow Jones, overseas investors dumped a net ?2.3613 trillion worth of SK Hynix shares on Tuesday alone. Market watchers cited by Reuters pointed to a mix of nerves over the durability of AI capital spending and plain profit-taking — the shares are still up 118% year-to-date, a gain that invites trimming.

Two Explanations, One Selloff

Whether the pullback marks the start of a cyclical cool-down or an entry point into the next leg of the semiconductor cycle is the question dividing the market. Dan Ives, an analyst who described the recent rout in memory names as overdone, argued that South Korean producers remain at least three years ahead of their Chinese rivals on technology. On that view, persistent shortages of advanced memory parts could set up a W-shaped recovery once foreign outflows ease.

The more uncomfortable explanation lies in the competitive landscape. An investigation by a South Korean lawmaker's office found that 27 former employees of Samsung Electronics and SK Hynix have moved to Chinese memory maker CXMT, with the hires spanning stages of chip development through to mass production. The report concerns a rival's recruitment drive rather than any personnel decision by the Korean company itself, but it underscores an awkward reality: expertise that once looked like an unbreachable moat is migrating faster than expected, and with it the long-term pricing power of the incumbents.

CXMT has said it will introduce a 4F² DRAM architecture before the end of this year. Its manufacturing maturity and yields may still trail the global leaders, yet the announcement has stoked fears of medium-term price pressure in standard DRAM.

Management Looks Past the Cycle

SK Hynix's leadership, meanwhile, is focused on a horizon measured in years rather than quarters. Speaking at the SEDEX industry fair in Seoul, Ahn Hyun predicted that performance demands on global AI infrastructure will multiply twentyfold by 2030. Between 2026 and 2030, the company expects an average annual growth rate of 109%, driven chiefly by complex inference workloads and rapidly expanding context windows.

That shift is redrawing hardware requirements. Context lengths have grown from roughly 128,000 tokens two years ago to about one million today, with 100 million tokens in sight. Raw processor throughput is no longer the sole determinant of speed; the movement of data between SRAM, HBM, DRAM and SSDs has become the critical variable. High Bandwidth Memory has dominated the conversation around AI accelerators so far, but the rise of autonomous agents is swelling the volume of data that must be processed.

To break the bottleneck, SK Hynix is developing 3D-stacked DRAM, an additional intermediate layer designed to connect processors more directly and multiply bandwidth. Should the architecture reach industrial maturity on schedule, it would hand the company a technological edge capable of sustaining high margins in the lucrative high-end segment. The company is also pursuing High-Bandwidth Flash, a NAND technology built on silicon wafer through-vias, intended as an intermediate tier offering roughly ten times the capacity of HBM at about half the bandwidth.

Solidigm Repositions as Washington Looms

The geopolitical map is being redrawn in parallel. SK Hynix operates significant production in China, including a DRAM fab in Wuxi and NAND manufacturing in Dalian. US authorities have suspended a key review rule until November 9, 2026; if the relief is not extended beyond that date, importing advanced manufacturing equipment could become considerably harder. The current US-China tariff agreement also expires in early January 2027, a prospect that could further dampen the investment appetite of large technology groups.

The company's US subsidiary, Solidigm, is hedging accordingly. It is expanding in Taiwan to build out production of data-center solid-state drives for the AI server supply chain. On Monday, Solidigm named Andres Castrillon as its first vice president of government affairs, establishing its own representation before policymakers independent of the parent company. SK Hynix said no decisions have yet been made on Solidigm's capital plans or possible external financing options, adding that it is weighing investment needs for AI data-center memory solutions and prioritizing long-term corporate and shareholder value.

What to Watch

Two markers will shape the next phase. The first arrives on November 9, 2026, when the US exemption expires and the outlook for future export restrictions becomes clearer. The second follows in January 2027, with the planned finalization of HBF specifications in cooperation with Google and SanDisk — a test of the company's claim to technological leadership.

Until those waypoints pass, investors should brace for continued volatility. The narrative of an effortless self-propelling cycle has given way to the realities of a capital-intensive displacement contest, defined by talent poaching and geopolitical tug-of-war. Whether the stock can defend its recent intermediate low — and whether foreign capital flows show signs of bottoming — will determine if the recovery thesis holds or whether deeper support levels come back into play.