Domestic NPUs Strengthen AI Sovereignty: "Building a Full-Stack Sovereign AI Ecosystem"

Upstage Solar: Igniting the Domestic AI Engine
Furiosa AI and AXZ Join Forces for Service Collaboration

As domestically produced neural processing units (NPUs) are increasingly adopted in both Korean companies and the public sector for artificial intelligence (AI) services, their range of applications is gradually expanding. Amid the accelerating movement among big tech firms to move away from Nvidia, the goal is to reduce reliance on foreign-made graphics processing units (GPUs) and further expand the use of high-performing domestic NPUs, thereby realizing sovereign AI in infrastructure as well.

Sung-Hoon Kim, CEO of Upstage, revealed in an online briefing on the morning of July 15, 2026, that “the large language model (LLM) ‘Solar’ is running on high-performance domestic NPUs and now operates in real time on a nationwide portal," confirming the collaboration. The briefing was attended by CEO Kim, Joonho Paik, CEO of Furiosa AI, and Geonsu Lee, CEO of AXZ.

Sung-Hoon Kim, CEO of Upstage, Joungho Baek, CEO of Furiosa AI, and Geonsu Lee, CEO of AXZ (from left), are speaking during an online discussion held on the morning of the 15th. YouTube capture

Through this collaboration, the AI Overview search summary service provided by the portal site ‘Daum’ operates based on Furiosa AI’s NPU, ‘Renegade’ (RNGD). The inference process for AI-based search summarization is handled by the Renegade NPU, while Daum’s AI Overview runs on Upstage’s Solar model.

Joonho Paik, CEO of Furiosa AI, explained, "For the operation of the Solar model powering Daum’s AI Overview, about 24 Renegade NPU chips are deployed, processing approximately 500 million tokens per day," adding, “By jointly designing both hardware and software, our solution achieves superior performance-to-power and cost-to-performance ratios compared to Nvidia’s H200 GPU, while delivering comparable performance.”

Upstage aims to use this opportunity to establish a 'full-stack sovereign AI' ecosystem. This means the company intends to internally secure all core technologies—from NPUs as AI semiconductors, to AI models, and the resulting AI services—independently through domestic technology. Upstage was selected as an elite team in the government's independent AI foundation model project, released the 'Solar Open 100B' model, and passed the initial assessment. In May 2026, Upstage acquired AXZ, operator of Daum, thereby securing a channel to connect AI services with users.

Domestic NPUs are now being widely adopted not only by Upstage but also by other companies and the public sector, broadening their use cases. In addition to its collaboration with Upstage, Furiosa AI also partners with organizations such as LG AI Research, Samsung SDS, LG CNS, and LG Uplus. In some cases, AI models and services are provided using Furiosa AI’s NPUs, while in others, NPUs are provided as a service (NPUaaS). Similarly, Rebellions is working with SK Telecom, applying Rebellions' NPUs to services like A.Dot’s AI-powered call summaries. In the public sector, projects are underway to convert more than 50,000 closed-circuit (CC)TV units to NPU-based AI monitoring systems.

There are increasing calls for greater adoption of domestic NPUs to further strengthen sovereign AI. As the global AI market shifts its weight from model training toward inference for real-world services, domestic NPUs, with their higher power efficiency in the inference phase compared to GPUs, are seen as gaining a competitive edge. The industry expects that by 2030, the market size for inference will surpass that of training.

However, given the continued importance of GPUs for model development, some experts recommend a 'two-track' strategy that leverages both global company GPUs and domestic NPUs. Eun-joo Kim, Director General at the National Information Society Agency (NIA), stated in the keynote address at the ‘2026 K-NPU Tech Wave’ the previous day, “Domestic NPUs are not just a means to reduce costs, but a key foundation for building and operating AI infrastructure with our own technology. While global GPUs should be used for training, a pragmatic parallel strategy is required—expanding the use of domestic NPUs in the inference domain.”

This content was produced with the assistance of AI translation services.