The South Korean memory chip maker's Cheongju plant will produce NAND flash, the bulk storage chips inside phones, SSDs, and AI data centers, with output not due until the first half of 2029 — a horizon that has humbled every major memory maker in past downturns.
SK Hynix is wagering roughly $66 billion (KRW 80 trillion) that the AI industry's appetite for data storage will still be hungry enough in 2029 to justify a brand-new NAND flash factory in South Korea, a bet on demand that the company itself will not be able to test for at least four years.
The investment, announced in Cheongju, ranks as one of the largest single-fab commitments in SK Hynix's history. Construction on the new NAND facility is set to begin in 2027, with production targeted for the first half of 2029. NAND flash is the storage technology that holds data in everything from smartphones and solid-state drives to the data centers powering large AI models. It is distinct from DRAM and HBM, the high-bandwidth working-memory chips that have powered much of SK Hynix's recent growth.
The bet is a deliberate split. SK Hynix has spent the last three years positioning itself as the indispensable HBM supplier for AI accelerators. Yet the new capacity is almost entirely NAND, the slower, cheaper, bulk-storage cousin that data centers use to warehouse training corpora, model weights, and inference logs. The company is, in effect, telling the market that AI will not just need more high-speed working memory; it will need a parallel expansion of long-term storage capacity that mirrors the buildout.
Why now: SK Hynix and other memory makers have flagged a NAND market that has swung from multi-year lows into a tightening phase, and AI demand for training data, model weights, and inference logs is now pulling on storage supply that competitors including Samsung and Micron have already begun flagging as tightening. By committing to production four years out, SK Hynix is locking in physical capacity ahead of any near-term competitor buildout, but it is also asking investors to underwrite a demand forecast that will not be testable until 2029 at the earliest.
That horizon is where the bet gets risky. NAND prices have historically swung in two-year boom-bust cycles, and the last downcycle forced every major memory maker to cut capex. A facility financed near a market peak, scheduled to come online when the next cycle could already be turning, exposes the company to the same fate. The risk note is real, even if its timing is debated: past cycles have shown that hyperscale buyers can swing from shortage-panic to inventory-buildout in a matter of quarters, leaving memory makers with capacity they cannot fill.
The Cheongju plant is not the only capacity SK Hynix is committing. Alongside the NAND fab, the company said it will spend an additional KRW 20 trillion, roughly $12.85 billion, on a new advanced packaging facility in the same city, slated to be operational by late 2027. Advanced packaging is the layer where memory chips, including HBM stacks, are bonded to logic dies, a bottleneck that has been the focus of multi-billion-dollar investment from SK Hynix, Samsung, and TSMC over the past 18 months.
The two Cheongju plants are part of a broader KRW 1,100 trillion investment roadmap that also includes SK Hynix's HBM mega-cluster in Yongin and additional sites in southwest Korea. They also sit inside a wider KRW 392 trillion (~$252.5 billion) Chungcheong-region semiconductor package that includes Samsung Electronics' separate commitment to build HBM packaging fabs in the same corridor, a level of state-backed co-location that turns the inland province into the country's most concentrated AI-memory cluster.
That co-location is why South Korean President Lee Jae-myung attended the announcement event alongside SK Hynix CEO Kwak Noh-jung. The president's presence, at what would otherwise read as a routine corporate capex disclosure, signals that the buildout is now part of national industrial policy, not just a company balance-sheet decision. State backing does not eliminate demand risk, but it changes the political cost of walking the capacity back if conditions change.
For now, the company is asking the market to take the demand call on faith: $66 billion of physical capacity, four years of construction, and the assumption that AI data centers in 2029 will still be hungry for the slow, dense, cheap storage that NAND provides. The wire services will report the figure. The harder question is whether the cycle cooperates.