SK Hynix's board has approved about 54.3 trillion won ($38.3 billion) to build two new memory chip factories in South Korea, one in Yongin and one in Cheongju (Reuters). The company announced the investment on August 7, 2026, pointing to explosive demand for memory chips driven by the construction of AI data centers (Engadget).

The bigger of the two commitments is 35.2 trillion won ($24.9 billion) for the Yongin Y2 factory, which will make a type of memory called HBM and other DRAM chips. DRAM is the kind of memory your computer uses to hold data it is actively working on. HBM is a faster, more advanced version stacked in layers and placed right next to the processors that power AI systems. A further 19.1 trillion won ($13.2 billion) goes to the Cheongju M17 facility, which will make NAND storage chips, the kind of memory used to store files permanently, like in a phone or a USB drive. The first production line across the two sites is targeted to start as early as June 2029 (Engadget).

SK Hynix described the 54 trillion won commitment as securing a long-term production base for AI memory demand (SK Hynix News). The split between the two types of memory is notable. The Yongin factory, focused on HBM and next-generation DRAM, absorbs roughly two-thirds of the total spend. HBM has been the hardest-to-get memory in the current AI boom, and SK Hynix has positioned its HBM3E product line at the center of that demand.

The Cheongju site has a longer investment history at SK Hynix. In April 2024, the company announced plans for the M15X factory there, committing over 20 trillion won to expand HBM and next-generation DRAM production (SK Hynix News). In July 2026, SK Hynix outlined a 100 trillion won plan for Cheongju that includes the M17 NAND factory and an advanced packaging plant, with the goal of establishing the region as a semiconductor hub (SK Hynix News). The M17 NAND factory announced this week falls within that broader plan.

The two new factories add capacity that will come online years from now, not during the current shortage. A first production line starting in June 2029 means initial chips are roughly three years out, with full output extending further. By that point, the AI memory demand that justified the investment will look different than it does today. Memory investment cycles have historically been long and uneven. Factories announced at the peak of one demand wave sometimes enter production as conditions shift.

SK Hynix is the second-largest RAM and NAND chip manufacturer globally, behind Samsung (Engadget). The two Korean firms together dominate the supply of high-bandwidth memory that AI system designers depend on, and Samsung has its own pipeline of factory investments in South Korea. Capacity decisions by either company shape the supply outlook for the entire AI hardware industry.

In practical terms, the Yongin factory's focus on HBM and DRAM directly targets the memory bottleneck in AI systems. The speed of HBM, not the raw processing power of the chip next to it, has been the limiting factor on how fast AI accelerators can run. More HBM capacity coming online in 2029 and beyond would ease that bottleneck, assuming demand forecasts hold. The NAND investment in Cheongju addresses a different need: AI data centers require substantial storage for training datasets, saved models, and running workloads, and high-density NAND is the cost-effective choice for that job.

The broader context here is the scale of the bet. The combined $38.3 billion through 2031 places this among the largest single-company semiconductor investments in recent memory. It is a multi-year wager that AI-driven memory demand has lasting depth, not just a short-term spike. Whether that bet pays off depends on factors no factory announcement can resolve: how AI models continue to evolve, competition with Samsung and other memory suppliers, and the possibility that changes in AI hardware design could shift what kind of memory is needed.

For now, the commitment is concrete. Board approval, a defined timeline through 2031, and named facilities with specified products give this more weight than a preliminary agreement. The first tangible output, a production line running in mid-2029, is the milestone to watch.