The construction site stretches across the hills south of Seoul, where cranes now outnumber the trees. CNBC toured the complex on Aug. 13, and the scale of what SK Hynix is building there is hard to convey in numbers alone: a cluster of memory fabs that will, by 2034, triple the company’s production capacity, funded by an investment of $720 billion.
The figure makes this the largest semiconductor factory construction program in Korean history, according to company executives and industry analysts. SK Hynix is betting that the AI boom’s appetite for memory, particularly the high-bandwidth memory used in Nvidia’s accelerators, will last for a decade. The bet is large enough that its outcome will help decide Korea’s position in the global chip industry for a generation.
Group Chairman Chey Tae-won addressed the obvious question during the tour: what happens if the company’s largest customer slows down? He said he is not worried about dependence on Nvidia, pointing to the numbers. In the first half of the year, sales to Nvidia exceeded 17 trillion won, roughly 13 percent of SK Hynix’s total revenue. That concentration would alarm investors in most industries, but Chey’s argument is that the relationship is mutual: Nvidia’s products cannot ship without SK Hynix’s memory.
The expansion plan covers both high-bandwidth memory, the stacked chips that sit next to Nvidia’s processors, and conventional DRAM and NAND. Executives said the new fabs will be built with flexibility, able to switch production between memory types as demand shifts. That flexibility matters in a business where misreading the cycle by a year can produce billion-dollar write-downs, as SK Hynix itself demonstrated during the last downturn.
The company’s position today is the strongest in its history. It leads the high-bandwidth memory market with a share that analysts estimate above 50 percent, and its technology advantage in the stacked architectures HBM requires has held through multiple product generations. Its customers are effectively locked in: switching memory suppliers mid-generation is not practical for chip designers, which gives SK Hynix pricing power it has rarely enjoyed.
The $720 billion program is also a national project. Korea’s government has designated semiconductors a strategic industry and has backed the expansion with tax incentives, fast-tracked permitting and infrastructure spending on power and water for the fab complex. The country’s economic planners view the memory industry as a pillar that cannot be allowed to erode, and SK Hynix’s investment is the largest single expression of that policy.
The financing question is real. SK Hynix generated strong cash flow in the first half as memory prices recovered, and its balance sheet has room for debt after years of deleveraging, but $720 billion over eight years is a sum that will require continuous access to capital markets. The company has said it will fund the program through operating cash flow, borrowing and, if needed, equity. Analysts said the plan is feasible only if the AI memory boom holds; a downturn would force the company to stretch payments and slow construction, as it did during previous cycles.
Competitors are watching closely. Samsung, the other Korean memory giant, is spending heavily on its own expansion, and Micron in the United States is adding capacity with government support. All three are preparing for a decade in which AI servers consume memory in volumes the industry has never produced. If they are right, the world will need every fab they build; if they are wrong, the industry faces the deepest overcapacity glut in its history.
Chey’s public confidence reflects a longer view. He told reporters during the tour that the memory industry’s problem has never been too much demand, only too little supply discipline, and that the AI cycle is different from the smartphone cycles of the past because the demand comes from data centers with year-round consumption. The company’s own forecasts show HBM demand growing at triple-digit rates for the next several years.
The company’s engineering plan is as ambitious as its financial one. Each new fab in the cluster is designed to be the most automated in the industry, with robots moving wafers between tools and artificial-intelligence systems managing yield in real time. SK Hynix executives said the complex will eventually employ tens of thousands of workers, with housing, schools and hospitals planned nearby, turning the site into a company town in the tradition of Korean industrial conglomerates.
The tour ended with the chairman standing in front of a site map showing dozens of planned buildings. The company says the first new fabs will begin production before the end of the decade, with capacity coming online in waves through 2034. By then, SK Hynix expects to produce three times as much memory as it does today, most of it feeding AI systems.
For Korea, the project is a statement of national industrial strategy made in concrete and steel. For the global chip industry, it is a supply forecast with the power to set prices for a decade. And for Chey Tae-won, it is a personal conviction, stated plainly: the AI era will need the memory, and SK Hynix intends to build all of it.
