Korea-made on-device AI semiconductors have reached global-level hardware performance but are failing to clear the commercialisation hurdle because of software barriers. Domestic NPU chips often drop out at the proof-of-concept stage because software development kits, compilers and related tools are not mature enough for customers to run them. Set makers say a lack of frameworks, profiling and debugging tools raises development costs, pushing them back toward Nvidia’s CUDA-based chips.

South Korea cannot achieve AI full-stack self-reliance in the short term because of high foreign dependence across the technology stack, from semiconductors and accelerators to application software, a report said. KISDI urged resetting strategy toward minimum dependence and maximum competitiveness, with near-term 집중 investment in AI models and cloud. A survey of 38 experts found high overseas dependence in six of nine stacks and below-average competitiveness across all areas.