SK Hynix's Record Quarter and $100 Billion US Expansion Sit Uneasily Beside the Market's AI Doubts

Seoul — Two very different pictures of SK Hynix are circulating at the same time. One is of a company that just posted record numbers, has effectively sold out its memory output through 2029 and is preparing to hand more than 90 trillion KRW back to shareholders. The other is of a stock caught in a sector-wide retreat, down 5.0% in a single session as investors once again question whether the enormous sums being poured into AI infrastructure will ever earn an adequate return.

The Korean chipmaker's shares fell 5.0% yesterday, swept along by weakness across technology and semiconductor names. Reports linked the decline at Korean memory makers to the broader selloff in US tech stocks, with the economics of AI infrastructure spending at the centre of the debate. For SK Hynix, the pressure was part of a sector-wide move rather than a verdict on its own results.

That distinction matters, because the company's recent disclosures describe a business firing on all cylinders. On 28 October, SK Hynix reported record figures for the third quarter of 2026: revenue of 104.6155 trillion KRW and operating profit of 83.3338 trillion KRW. Management said DRAM and HBM volumes were practically sold out through 2029 — a phrasing that signals extensive coverage rather than an absolute, unconditional sellout.

Capital Returns Add a Second Counterweight

The same day brought a shareholder remuneration programme for 2026, comprising cash dividends of more than 30 trillion KRW alongside share buybacks and cancellations worth more than 60 trillion KRW. These are announced intentions rather than completed measures, but they reinforce the picture of a company whose recent news flow diverges sharply from the mood in the wider tech sector.

What investors are left to weigh is a genuine clash of perspectives. The company's own figures and long-term volume commitments speak to current business strength. The market's anxiety is aimed at a different question — whether the industry's AI infrastructure outlays will pay off — and record quarterly results alone cannot settle that industry-wide profitability puzzle.

A US Manufacturing Footprint Takes Shape — on Paper

Running in parallel is a capital question of a different kind. According to a report today in the Seoul Economic Daily, SK Hynix is exploring a substantial build-out of memory production in the United States. Its NAND subsidiary Solidigm is in talks over a site in the state of New York, while a possible DRAM operation at Intel's facility in Ohio is also under discussion. Neither project has been approved; investment terms, financing and long-term customer contracts all remain prerequisites.

The Marcy Nanocenter in Oneida County is seen as the leading candidate for the first US NAND plant, where two fabs could rise in stages. The overall investment could reach as much as $100 billion, the report said, though SK Hynix stressed explicitly that construction and spending have not been decided. Each fab is envisaged at 100,000 wafers per month on a 12-inch basis, with industry sources putting the cost per plant at $40–50 billion.

New York is considering up to $50 million for site preparation. Industry circles point to SK Hynix's own cash reserves, loans and subsidies as possible funding components, though which mix would actually be used is open. The staged approach would spread the expansion across several construction phases, and neither a final start date nor a firm investment volume has been set — the figures describe possible dimensions, not binding spending commitments.

Ohio Would Need a Different Playbook

In Ohio, SK Hynix is reportedly weighing DRAM production at Intel's New Albany plant, with options including leasing part of the facility or forming a joint venture with Intel and US cloud companies. No concrete plan has been confirmed. Intel's planned logic manufacturing cannot simply be swapped for memory output: DRAM requires different tools and processes, and additional investment and retrofitting could be needed. Long-term offtake agreements and a clear cost split would therefore be decisive.

The timeline also limits near-term significance. Intel has pushed completion of its first plant to 2030, with production slated for 2030–2031, making any collaboration a long-horizon capacity project.

What Would Actually Move the Needle

The strategic backdrop is a memory market that is largely spoken for. Including existing and still-discussed customer contracts, DRAM and HBM capacity is practically sold out through 2029, according to the company — a classification that explicitly covers agreements still under negotiation. Long-term contracts provide a foundation for expansion, but they substitute neither for a final investment decision nor for settled financing. With the US projects in particular, the key question is how firmly customers would commit to taking additional capacity.

For investors, the next meaningful step is therefore not another site announcement. Binding investment decisions, clarified funding terms and robust offtake contracts would carry far more weight. Until then, New York and Ohio represent potential additional capacity — not production that is already secured.