Memory bandwidth is the binding constraint in AI compute. A GPU can add faster logic and more cores, but if the memory interface cannot move weights and activations quickly enough, throughput stalls. High-bandwidth memory, the stacked-die architecture that sits directly beneath AI accelerator chips, solves that problem at the hardware level. SK Hynix, the leading maker of high-bandwidth memory, is committing $720 billion to expand memory factory capacity in South Korea. An exclusive first look at the facilities shows a buildout now large enough to visibly reshape the country's industrial footprint.
The constraint driving that number is specific: HBM production is not interchangeable with standard DRAM manufacturing. Stacked-die packaging requires dedicated lines and specialized process capability, which is why volume supply remains concentrated in a small number of producers. SK Hynix holds the leading position. Its factory output sets the practical ceiling on HBM availability to the AI accelerator market globally.
Capital commitments at $720 billion scale are responses to contracted demand. AI hardware companies secure multi-year HBM supply through purchase agreements, and factory investment is sized to those agreements. A buildout of this scale, characterized as taking over South Korea's industrial footprint, is a measure of the volume AI hardware demand currently requires.
Whether South Korea's power infrastructure and construction capacity can sustain expansion at this pace is a question the buildout raises without yet answering. The $720 billion committed to memory factories is the current visible fact.