As the core focus of the artificial intelligence (AI) market shifts from 'training' to 'inference,' domestically-produced neural processing units (NPUs) are rapidly emerging as the key driver that will determine the Republic of Korea's AI sovereignty and industrial competitiveness. The government is undertaking large-scale NPU transition projects in the public sector, while domestic companies are successfully applying domestic NPUs combined with large language models (LLMs) to real-life services, opening a new era of AI supply chain independence.

As of today (July 19, 2026), the center of gravity in the AI market has shifted beyond model training to 'inference' for actual service implementation. NPUs are specialized in this AI inference task and have the advantage of lower power consumption and higher cost efficiency compared to graphics processing units (GPUs). This presents a new alternative to the existing AI hardware ecosystem centered on Nvidia GPUs.

The government recognizes domestic NPUs as core hardware for securing AI sovereignty and is accelerating their deployment in public sectors. In particular, the government is pursuing a project to convert approximately 50,000 public CCTV cameras to an NPU-based AI monitoring system over the next five years starting this year, and is also implementing a naval CCTV replacement project in the defense sector. Kim Eun-ju, head of the Korea Intelligence Information Society Promotion Agency (NIA), emphasized on July 14 that "domestic NPUs are the core foundation of AI sovereignty for developing and operating AI infrastructure with our own technology," and that leading adoption in the public sector will serve as a priming pump for the early market.

In the private sector, successful cases of combining domestic NPUs and LLMs are increasing the possibility of AI independence. On July 15, domestic AI company Upstage (LLM 'Solar'), NPU developer FuriosaAI (NPU 'Renegade'), and portal 'Daum' operator AXZ announced their collaboration results. Running the 'Solar' LLM on the 'Renegade' NPU and applying it to Daum's real-time search summary service resulted in processing approximately 500 million tokens per day while achieving performance similar to Nvidia GPUs. This proves that token processing costs can be significantly reduced and demonstrates the economic viability and efficiency of domestic AI infrastructure.

[Photo=Yonhapnews]

Baek Jun-ho, CEO of FuriosaAI, stated that "we have achieved performance equivalent to Nvidia products," and emphasized that prices will be reduced to less than half in the future. He added that the 'Renegade' NPU currently in mass production can supply approximately 10,000 units by the end of the year, raising expectations for resolving the issue of dependency on the Nvidia ecosystem.

Furthermore, the government has begun building a 'Hyper-AI Network' in preparation for the era of physical AI. This is a next-generation infrastructure combining 5th generation mobile communication standalone mode (5G SA) and AI-based wireless access network (AI-RAN). A consortium led by SK Telecom and KT has been selected as the lead organization and plans to demonstrate various physical AI services including four-legged patrol robots and humanoid low-power modes.

The synergy between domestic NPUs and LLMs, combined with the construction of the 'Hyper-AI Network,' is expected to be a decisive opportunity for the Republic of Korea to secure leadership in the global AI competition beyond AI independence. However, experts point out that accompanying growth across the industry is essential, including strengthening the software ecosystem, securing specialized talent, and promoting private investment.