South Korea's Moreh claims AMD MI355X delivers over twice Nvidia B200's inference throughput

This is a software-stack story disguised as a hardware benchmark. Moreh's inference layer is the actual throughput unlocker, striking at the weakest seam in Nvidia's CUDA moat: the inference side, where over 80% of production AI compute is consumed. The strategic read for Seoul is unambiguous—pairing AMD silicon with domestic software is a geopolitical hedge against US export-control volatility while simultaneously building sovereign AI capacity. Nvidia's counterplay is predictable: B200 price compression, accelerated CUDA inference optimizations, and direct lock-in deals with Korean hyperscalers. Over the next 12-24 months, three shifts will define the landscape. First, inference—not training—becomes the primary competitive axis. Second, "local software plus foreign silicon" solidifies as the default Asian AI infrastructure pattern, echoing how Korea has historically balanced US, Japanese, and Chinese tech dependencies. Third, AMD's ROCm ecosystem receives its first credible large-scale enterprise validation. The critical risk for AMD is not silicon performance. A single benchmark win does not build an ecosystem. Whether ROCm can sustain the software iteration velocity that Moreh demonstrated will determine if AMD transitions from "acceptable alternative" to "default choice" in inference. Seoul's government endorsement transforms Moreh from a startup into a de facto reference architecture—and that is the real competitive weapon in this equation.