South Korea unveiled a 10-year, $880 billion plan to dominate AI-chip production, with Samsung Electronics and SK Hynix steering the effort. The sheer size of the spend and the fact that the first new fabs will not ship silicon before 2028 make the initiative a potential choke point for the global AI-hardware supply chain.

The plan in numbers

  • Semiconductor manufacturing: $518 billion earmarked for four new wafer-fab complexes in the country’s southwest.
  • AI data centers: $356 billion to build massive compute facilities that will host the next generation of generative models.
  • Robotics: Strategic funding to embed AI into logistics and factory machines; the exact amount is not disclosed.

The money comes from a mix of government backing and private-sector commitment, but the two Korean chip giants drive the project.

Why AI chips need more than GPUs

Modern AI accelerators from Nvidia, AMD and others rely on high-bandwidth memory (HBM) to feed data at terabytes per second. HBM stacks several memory dies vertically and connects them with microscopic interposers, delivering far more throughput than standard DDR memory.

Current market share figures show:

  • SK Hynix controls roughly 50-56 % of the HBM market.
  • Samsung holds about 35-40 %.
  • Micron accounts for the remaining 5-10 %.

All three suppliers have announced that their HBM capacity is sold out through 2026, meaning new AI workloads will keep competing for a shrinking pool of memory chips.

Timeline and infrastructure bottlenecks

Even with $518 billion slated for fab construction, the earliest silicon from the new plants will not appear before 2028. Building a state-of-the-art fab requires years of civil work, equipment installation, and a lengthy qualification phase.

The bigger, less discussed hurdle is utilities. The Yongin chip cluster, one of the sites, will need 15-16 GW of electricity at full capacity, yet the local grid supplies only 1.9 GW. Water, essential for wafer-washing and cooling, faces a similar shortfall. Without massive upgrades to power lines and water-treatment plants, the factories cannot run at design speed.

Who stands to gain, who could lose

  • Samsung and SK Hynix: A larger slice of the AI-hardware pie could translate into billions of dollars in future royalties and higher valuations.
  • South Korean economy: The plan bets on becoming the backbone of AI hardware.
  • Cloud providers and AI startups: With HBM scarce and new chips delayed, GPU rental prices will likely stay elevated through 2027, squeezing margins for firms that rely on cheap compute.
  • Competing memory makers: Micron’s modest share may stall expansion without new capacity, and any newcomer will face high entry costs in a market already dominated by the two Korean firms.

What to watch

  • Utility upgrades: Progress on new transmission lines and water-treatment capacity will indicate whether the fab schedule stays on track.
  • HBM supply trends: Any announcement of additional memory capacity, from the Korean firms or a third party, could ease the current “sold-out” status.
  • Custom silicon designs: Companies may accelerate in-house chip projects that integrate memory on-die, sidestepping the external HBM bottleneck.
  • GPU pricing: Cloud GPU prices will stay high through 2027 due to scarcity.

Success depends on building more than just factories; the nation must build the power grids to support them.