Several Korean companies are designing custom AI accelerator chips: Rebellions (ATOM chip for data center inference), Sapeon (Samsung spin-off building NPUs), FuriosaAI (low-power inference accelerators), and Samsung's own Exynos NPU division. The Korean government's AI Semiconductor Initiative provides R&D funding and guaranteed procurement for domestic AI chips in government data centers.
Korea's AI chip ambitions are partly motivated by supply security — Korean AI companies found themselves competing with global hyperscalers for scarce Nvidia GPUs during the AI boom of 2023-2025. Building domestic AI accelerators ensures that Korean sovereign AI models can train and run on Korean silicon, reducing a critical foreign dependency.
Rebelions' ATOM chip, fabricated at Samsung Foundry on 5nm, is designed for transformer inference workloads and claims competitive performance-per-watt with Nvidia's inference GPUs. FuriosaAI targets edge AI applications in autonomous driving and industrial automation. While none of these chips threaten Nvidia's training GPU dominance, they represent Korea's effort to own more of the AI value chain beyond memory — moving from supplying HBM to Nvidia to building complete AI compute solutions.