A visit to Samsung’s massive new semiconductor hub
Oct 20, 2026, 2:00 AM · MIT Technology Review
The AI boom has a physical address in a Korean farming city, where thousands of construction workers are pouring concrete for the memory chips every GPU needs.
Why it matters
Every frontier model runs on GPUs, and those GPUs depend on stacks of specialized high-bandwidth memory packaged right next to the processor. Samsung's answer to soaring demand is a sprawling complex in Pyeongtaek, south of Seoul, where three fabs already run around the clock, a fourth is coming online, and a fifth — paused during the 2024 memory slump — has been restarted on the back of AI demand.
Samsung expects memory demand to keep surging through 2028. The build-out is also a competitive fight: the company has watched domestic rival SK Hynix pull ahead in high-bandwidth memory for Nvidia's accelerators. Who supplies that memory shapes how fast and how cheaply AI capacity can grow.
From the desk
We like dispatches like this because they puncture the idea that AI is weightless. The chatbot on a phone ultimately traces back to excavators, tower cranes and shuttle buses at dawn in a campus about the size of Central Park. Model releases get the headlines; fabs like these set the ceiling on how many of those models can actually be served.
The restarted fifth fab is the detail worth sitting with. A project halted in a downturn and revived because of AI tells us how much of the semiconductor cycle now rides on one category of demand. That is good news for capacity and for the workers on site — the safety officer the reporter profiled earns 140,000 won (about $97) a day and expects the boom to keep her busy. It is also concentration risk. If AI spending cools, memory has a long history of brutal swings, and Samsung itself already lived one in 2024. Communities that reorganize around a boom, down to the real estate offices advertising short-term studio leases, absorb the bust too.
We were also struck by how much attention the site gives to safety: banners urging crews not to rush, helmets carrying a pledge not to work unsafely. That culture deserves credit, and it sits in tension with the urgency of a race to reclaim market share. The interesting question is not whether the slogans exist, but whether they hold when a customer schedule slips.
Strategically, a stronger Samsung in high-bandwidth memory would be healthy for the AI industry. A market where one or two suppliers decide how many accelerators ship is fragile, and more competition should mean more supply and better pricing for everyone building on GPUs. I'm watching whether this capacity actually translates into Samsung winning qualification for the most advanced memory parts, because building a fab and winning the design slot are different victories.
Context
High-bandwidth memory is stacked DRAM packaged alongside AI accelerators to feed them data fast enough. SK Hynix has led that segment for Nvidia's chips, while Taiwan's TSMC dominates contract manufacturing of chips designed by other companies. Samsung's Pyeongtaek complex is described as the world's largest chipmaking site.
Who feels it
- AI labs and cloud providers
- More memory capacity coming online through 2028 could ease a key constraint on accelerator supply, if Samsung's output qualifies for top-end parts.
- Chip industry competitors
- Samsung is spending to close the high-bandwidth memory gap with SK Hynix; expect sharper competition on supply and price.
- Construction and local workers
- The boom brings steady, comparatively well-paid work to Pyeongtaek, along with exposure to the memory market's history of sharp downturns.
What to watch
- When the fourth Pyeongtaek fab reaches volume production
- Construction milestones on the restarted fifth fab
- Samsung's share of high-bandwidth memory supply for Nvidia accelerators versus SK Hynix
- Any softening in AI memory demand forecasts before 2028