SK hynix to spend $39 billion on AI memory chip factories in South Korea

SK hynix put a date on the AI memory race. The board approved a $39.42 billion capex on Friday, split between a DRAM and HBM megafab in Yongin and a NAND flash plant in Cheongju, both in South Korea. Groundbreaking for the Yongin Y2 line is scheduled for July 2027, with the first cleanroom ready by June 2029.

Each fab targets a different layer of the AI memory stack. DRAM is the working memory every AI server reads and writes during a training run. HBM, or high-bandwidth memory, is the stacked variant that sits next to the AI accelerator on the same package and feeds it data fast enough to keep thousands of cores busy. NAND flash is the persistent storage behind enterprise SSDs, the tier that holds model weights and training data. SK hynix is committing to all three on a defined schedule through the rest of the decade.

The Yongin Y2 fab eats 35.2 trillion won (about $25.70 billion) of the total and covers roughly 1.13 million square meters, about 280 acres, with completion slated for October 2031. The Cheongju M17 line takes 19.1 trillion won (about $13.94 billion) and is sized for enterprise SSD demand, the kind that follows hyperscaler AI buildouts. Both sites sit inside a much larger blueprint: SK hynix announced in June 2026 a 600 trillion won (about $438 billion) Yongin Semiconductor Cluster commitment and a 100 trillion won (about $73 billion) Cheongju production center. This $39.42 billion is the first dated tranche of that plan.

Omdia projects combined DRAM and NAND demand will compound at 19 percent annually through 2030. SK hynix reads the curve as structural rather than cyclical, and stated the case directly: "In the AI era, technological competitiveness alone is not enough and the ability to supply the required volume at the exact moment customers need it is the ultimate competitive advantage." That is a thesis statement, not a marketing line, and it sets the bar SK hynix is now publicly trying to clear.

The Yongin Y1 cleanroom is scheduled to open in February 2027, with utility installation 99 percent complete, so the first wave is not theoretical. But the AI memory market has a cycle history the company does not control. Memory has historically been the most cyclical segment in semiconductors: capacity additions take two to three years from groundbreaking to volume output, and demand has historically caught up or overshot by the time those lines ramp. SK hynix's claim that supply timing is the new competitive advantage is testable, and the test is whether AI training demand plateaus, hyperscaler memory orders normalize, or peer capacity from Samsung and Micron absorbs the same wave. The 600 trillion won Yongin cluster is a multi-decade buildout, and multi-decade buildouts carry multi-decade cycle exposure.

Samsung and Micron are watching the same Omdia curve and making their own capacity calls. If all three expand on the same timeline and AI training plateaus in 2028 or 2029, the segment that delivered the strongest margins of the cycle will also deliver its sharpest correction. The $39.42 billion does not change that math; it only commits SK hynix to it.

The Y2 cleanroom date, June 2029, is where the structural-versus-cyclical argument meets the production line. By then, two years of execution will sit against the calendar SK hynix just published, and Omdia's 19 percent curve will have produced its own verdict on AI demand.