South Korea's government is purchasing neural processing units (NPUs) from domestic chipmakers FuriosaAI and Rebellions to power new public-sector AI services, marking a deliberate shift away from reliance on Nvidia's graphics processing units (GPUs). The National Information Society Agency (NIA) signed agreements to buy NPU servers worth a combined 10 billion won (approximately $7.2 million USD) on August 26, with 25 AI servers set to be installed across government agencies.

This move reflects a broader strategic calculation: as government agencies accelerate their AI transformation, the power consumption and cost of GPU-heavy systems could become unsustainable. NPUs are specialized processors designed specifically for AI inference tasks, the repetitive computational work that powers chatbots, recommendation systems, and other public-facing AI services. Unlike GPUs, which consume significant electricity, NPUs deliver higher performance per watt, making them more economical for the kinds of routine AI workloads that government agencies typically run.

Why Is South Korea Prioritizing Domestic AI Chips?

The government's first major application will be an AI chatbot for public officials, which is expected to prioritize AI models from South Korea's independent AI foundation model project. FuriosaAI and Rebellions are participating in consortiums developing these homegrown large language models (LLMs), which are AI systems trained on vast amounts of text data to understand and generate human language.

Beyond the immediate cost savings, the government's NPU purchases serve a longer-term ecosystem goal. Once domestic chipmakers can demonstrate real-world performance and reliability in government systems, those case studies become powerful marketing tools for international clients. Governments and enterprises worldwide are increasingly interested in reducing their dependence on a single foreign supplier, creating a global market opportunity for Korean NPU makers.

"Domestic NPU companies are staking everything on building long-term partnerships with overseas clients, based on their track records with domestic public infrastructure and telecom and cloud partnerships," an official in the AI chip industry stated.

AI chip industry official

The timing is strategic. Last month, the Ministry of Science and ICT designated AI chips from Rebellions, FuriosaAI, and DeepX as innovative products from outstanding research and development, signaling government support for the sector.

How Are Korean Companies Building the NPU Ecosystem?

Beyond government contracts, Korean tech companies are taking a comprehensive approach to NPU adoption and development:

• Integrated Solutions: Hancom has partnered with affiliate Hancom Innostream and FuriosaAI to develop an integrated AI transformation (AX) appliance combining Hancom's agentic operating system with FuriosaAI's second-generation RNGD (Renegade) AI accelerator, targeting enterprise and public-sector customers.
• Workforce Development: Hancom Innostream will handle NPU specialist training and education, building the talent pipeline needed to deploy and optimize these systems across organizations.
• Regional AI Infrastructure: DGIST (Daegu Gyeongbuk Institute of Science and Technology) has been selected to build a deep tech startup hub with 3.5 billion won in investment, combining high-performance AI computing infrastructure with startup workspace to help early-stage companies access the expensive computing resources they typically cannot afford.

The partnership model reflects a recognition that NPU adoption requires more than just hardware. Companies and government agencies need training, software optimization, and proven use cases. Hancom CEO Kim Yeon-soo emphasized this integrated approach, stating that the three-way partnership will "support the AX transition of companies and public institutions and accelerate our push into the market".

What Does This Mean for the Global AI Chip Market?

South Korea's government-backed NPU strategy signals a broader global trend: countries and enterprises are diversifying away from GPU-centric AI infrastructure. While Nvidia's GPUs remain dominant for AI training (the computationally intensive process of teaching models), NPUs are increasingly preferred for inference, the deployment phase where trained models answer questions or make predictions in real time.

The government's 10 billion won investment may seem modest compared to global AI spending, but its significance lies in validation and precedent. When FuriosaAI and Rebellions can point to successful deployments powering government chatbots and public services, they gain credibility in international markets where enterprises and governments are actively seeking alternatives to Nvidia.

This strategy also reflects lessons learned from semiconductor history. South Korea built its memory chip dominance by securing domestic contracts first, then leveraging that experience to compete globally. The NPU ecosystem appears to be following a similar playbook, with government procurement serving as the foundation for international expansion.