NVIDIA GPUs for prefill, Korean NPUs for decode: Rebellions plans R200 inference systems using NVLink Fusion and MGX racks, with SK Telecom planning deployment. The systems are expected to be ready in 2028.