Background
AI infrastructure is moving toward a more mixed model of computing. Heterogeneous computing combines central processing units (CPUs), graphics processing units (GPUs), and neural processing units (NPUs) according to the demands of particular services. South Korea’s Ministry of Science and ICT (MSIT) places this model within changes in the global AI-semiconductor market as generative AI and agentic AI increase demand for inference. The ministry describes the approach as a way to pursue cost efficiency while optimizing performance and power efficiency.
AMD is central to the agreement because its portfolio spans CPUs and GPUs. According to the MSIT announcement, the company recorded $34.6 billion in annual revenue in 2025, including $16.6 billion from its data-center business. MSIT describes AMD as a major participant in the AI-computing market, with demand for server CPUs and GPUs contributing to its data-center business.
The company also brings software and standards capabilities to the partnership. AMD’s ROCm platform is open-source-based software for AI and high-performance computing, designed to support GPU-based development and application implementation. AMD is a founding member of UALink, an open standard for high-speed interconnection between AI accelerators. The release presents these capabilities as part of AMD’s effort to expand an open AI-computing ecosystem.
MSIT describes Korean AI-semiconductor NPUs as having strengths in inference. The proposed cooperation would integrate those Korean NPUs with AMD CPUs and GPUs in heterogeneous AI-computing infrastructure. The release presents this as a planned area of development and demonstration.