[Edaily Reporter Kim Hyun-ah ] South Korean AI semiconductor company Rebellion has successfully run a hyper-large AI model with 500 billion parameters (500B) using a server based on a domestically developed neural processing unit (NPU). This achievement is seen as proof of the competitiveness of domestic infrastructure in addressing cost and power efficiency issues—key challenges in operating large-scale AI services.

Rebellion (CEO Park Seong-hyun) announced on the 23rd that it had successfully run SKTelecom(017670)’s proprietary AI foundation model, “A.X K1,” using its own AI semiconductor server, “RebelServer.”

This demonstration marks the first instance of processing a large-scale AI model with over 500 billion parameters in a single-server environment. Rebellion explained that it has reduced the number of servers and the burden on data center resources required to operate ultra-large AI models, while securing the infrastructure efficiency necessary for the commercialization of AI services.

Ultra-large AI models have traditionally been built around high-performance GPU servers, as they require massive amounts of computation to improve performance. However, the costs of acquiring GPUs and increased power consumption have been cited as obstacles to the expansion of AI services. Rebellion emphasized that this demonstration shows that domestically developed NPU-based infrastructure can serve as an alternative for operating large-scale AI services.

This achievement is the result of combining SKTelecom’s AI model architecture with Rebellion’s software optimization technology. A.X K1 applies a MoE (Mix of Experts) architecture—used for the efficient operation of large AI models—to enhance computational efficiency even with 500B-class models.

Rebellion optimized AI computations for the NPU using its proprietary software technology and applied techniques for parallel processing of various types of tasks. This enabled a single server to reliably handle AI requests from multiple users.

In particular, this proof-of-concept also confirmed the potential for real-world service deployment. Through a demo of a “map-based Agent AI service,” Rebellion demonstrated how the AI generates responses and delivers results in real time, even in an environment where multiple users ask questions simultaneously.

For example, when a user requests everyday information such as “pharmacies open on the weekend,” the AI analyzes the data and provides map-based results. This case demonstrates the feasibility of implementing an Agent AI service for a large user base using only domestically developed AI models and NPU infrastructure.

Building on this achievement, Rebellion plans to accelerate the expansion of its “Sovereign AI” ecosystem, which combines domestically developed AI models and NPUs. The company is also participating in the Ministry of Science and ICT’s Independent AI Foundation Model project as part of the SKTelecom consortium, where it is conducting a pilot of AI infrastructure based on domestically developed NPUs.

Furthermore, drawing on its experience of stably operating the “A.dot Call Summary” service—which generates massive traffic—for about a year, the company intends to contribute to the expansion of AI services for the general public, such as the government’s “AI for All” initiative.

Park Seong-hyun, CEO of Rebellion, stated, “This pilot demonstrates that large-scale AI services can be operated solely by combining a domestically developed hyper-large AI model with a domestically developed NPU.” He added, “Having proven both the infrastructure’s efficiency and its practical applicability, we will contribute to strengthening the competitiveness of the Korean AI ecosystem.”