DIGITIMES chairman: Taiwan-South Korea cooperation beats competition in AI race

Hwang's cooperation thesis is engineering inevitability, not trade diplomacy. When AI accelerator bottlenecks shift from logic die to HBM-advanced-packaging coupling, Taiwan, China's CoWoS capacity and Korea's HBM3E/4 supply become physically inseparable—a 2.5D/3D co-design constraint, not a preference. The chain reaction is concrete: NVIDIA's B200 cadence simultaneously locks TSMC packaging slots and SK Hynix yield curves; a single-quarter expansion miss on either side creates a $20B+ delivery gap. The 'ecosystem' argument is really about OSAT, substrate, and materials supplier density around one fab cluster—something Samsung's vertical model cannot replicate within 18 months. On compliance, US export controls convert regional coordination from commercial choice into regulatory mandate. If Korea substitutes local 2nm logic for Taiwan, China nodes, the EDA and IP cost cliff is non-trivial. Strategic counter-moves: Intel 18A, if on schedule, fractures the TSMC-Korea packaging-storage duopoly; NVIDIA will likely qualify a second packaging source by 2026 as a geopolitical hedge. 12-24 month tail: AI inference demand pulls HBM requirements from training cards into edge and automotive silicon. The cooperation premium migrates down-market, and ecosystem lock-in becomes a more durable pricing moat than any single process node.