By combining a Korean-made AI chip and a large language model, it has succeeded in delivering infrastructure performance comparable to global GPU servers on a single server. Rebellions said on Wednesday it ran SK Telecom’s proprietary AI foundation model A.X K1 on its in-house AI semiconductor server, Rebel Server, and presented an agent service demo handling simultaneous requests from multiple users.

The demonstration verified infrastructure efficiency needed to operate large-scale AI. Rebellions stably ran the country’s largest language model, with more than 500 billion parameters, on a single Rebel Server. It is a way to reduce the burden of building and operating data centres by processing computation without waste within one server. The company said this configuration alone secured infrastructure capabilities comparable to global high-performance GPU servers.

A.X K1 adopts the Mixture of Experts, or MoE, structure, which has a high adoption rate globally, to secure computational efficiency even for large models. Rebellions developed dedicated software to match it, distributing various computations in a way optimised for semiconductor chips. The technology enables real-time processing without bottlenecks even when multiple requests come in at the same time.

Rebellions also demonstrated practical usability with a “map-based agent service” in the demo. Even if multiple users connect at the same time and each asks a question such as “pharmacies available over the weekend”, A.X K1 processed them without delay and immediately displayed results on a map. It is an example showing that public AI agent services used simultaneously by multiple users can be handled using only Korean-made AI technology.

Rebellions is participating as a member of an SKT consortium in the Ministry of Science and ICT’s project for a proprietary AI foundation model. Based on its experience running large-scale real-time AI services for about a year, including “A.Dot phone call summary”, it plans to contribute as a Korean-made NPU infrastructure partner in the government’s “AI for Everyone” project as well.

Rebellions CEO Park Seong-hyeon (박성현) said the demonstration was a practical example of sovereign AI, showing that large-scale AI services can be operated solely by combining a Korean-made hyperscale model and a Korean-made NPU, and that it proved both infrastructure efficiency and practical usability. He said it would continue to work with SK Telecom to build a Korean-style AI infrastructure that anyone can use in everyday life.