AMD expands Korea push after Taiwan supply talks

Lisa Su's Taiwan, China-to-Korea itinerary signals AMD's pivot from wafer procurement to ecosystem construction. NVIDIA's moat was never the GPU—it's CUDA's software lock-in and HBM vertical integration. Technically, the core play is coupling ROCm's open stack with Korean HBM production lines; the Instinct MI300 series depends on 48GB HBM3E, making Korean fabs the chokepoint. The "open computing ecosystem" framing is a direct shot at CUDA's closed model—AMD is selling a de-NVIDIA narrative to Korean AI data center builders. On compliance, tightening US export controls on AI chips make the Korea push both a capacity hedge and a regulatory firewall. However, Seoul's own export review mechanisms are hardening, and US friend-shoring pressure could force Samsung into exclusive advanced-node commitments, compressing AMD's negotiating leverage. Competitively, NVIDIA will likely accelerate HBM4 co-development with SK Hynix and deepen compute bindings with Naver and Kakao. Intel may exploit the window to push 18A node localization in Korea. 12-24 month outlook: AMD will ship a Korea-localized AI accelerator reference architecture targeting Seoul's national AI infrastructure budget. The real battleground isn't silicon—it's whose framework runs the first line of code a Korean developer writes.