AMD teams up with FuriosaAI, MangoBoost to build Seoul AI cluster

AMD's Seoul play is not a chip sale—it is an architectural wedge against NVIDIA's CUDA moat. Pairing FuriosaAI's sparse-compute fabric with AMD's dense silicon builds a heterogeneous inference stack that breaks the bigger-HBM-equals-better linear scaling narrative. The 2027 delivery window signals infrastructure lock-in, not a product launch: once Korean workloads are baked into this hybrid stack, switching costs become prohibitive. The DEEPX thread is the sharper signal. Low-power AI silicon targeting robotics and industrial edge directly contests NVIDIA Jetson and Intel's edge portfolio. AMD is shedding its datacenter-GPU identity and pushing into the edge tier. Risk calculus: Korea's structural HBM dependency on Samsung is a single-point-of-failure for MI-series cluster delivery. US export-control spillover could further constrain hardware selection in Korean AI infrastructure, inflating compliance costs through 2026. NVIDIA will almost certainly counter within twelve months with a Korea-specific sovereign-AI bundle, leveraging CUDA lock-in and localized services to neutralize AMD's architectural differentiation. Over the next eighteen to twenty-four months, industrial AI clusters will emerge as the second major battleground after datacenters, and heterogeneous will migrate from whitepapers into procurement contracts.