MIT Technology Review reports on a visit to Samsung's new semiconductor manufacturing complex in Pyeongtaek, South Korea, described as a massive undertaking reflecting the scale of investment memory makers are pouring into supporting generative AI. Three fabs are already running around the clock, and a fourth is being brought online, signaling a rapid ramp-up in capacity.
The expansion addresses a concrete bottleneck: generative AI models run on GPUs that rely on stacks of specialized memory chips packaged directly alongside the processor. This high-bandwidth, layered memory has become as critical a constraint in the AI hardware supply chain as the GPUs themselves. Samsung, alongside South Korean rival SK Hynix, is racing to secure its share of a market where demand continues to outstrip available supply.
Pyeongtaek's selection as the site for this buildout is no accident — the city is already one of Samsung's key industrial hubs in South Korea, equipped with the logistics and energy infrastructure needed to support fabs operating nonstop. Bringing a fourth production line online within a relatively short timeframe reflects pressure from major customers, including GPU makers and hyperscalers seeking to lock in supply over the medium term.
The on-the-ground account is a reminder that the AI race is not fought only in research labs or data centers, but also in semiconductor plants whose output capacity directly shapes how quickly the most compute-hungry models can be deployed. For industry players, the ability to produce high-performance memory at scale is becoming as competitive an advantage as access to graphics chips themselves.