Rebellion expands Korean-style Sovereign AI base with NVIDIA and SKT

Rebellion, along with NVIDIA and SK Telecom [017670], will begin to advance Sovereign AI infrastructure based on domestic neural network processing units (NPUs).

Rebellion announced this plan for cooperation between the three companies at the 'NVIDIA AI Day' held at COEX in Gangnam-gu, Seoul on the 9th.

It is a method that combines NVIDIA's open infrastructure platform, Rebellion's domestic AI semiconductors, and SK Telecom's AI data center, model, and service capabilities based on 'NVIDIA NVLink Fusion' technology.

Rebellion connects the next-generation NPU 'R200' to the NVIDIA platform through NV Link Fusion and develops a rack-level inference system that applies MGX rack architecture, NVIDIA's standard rack (server cabinet) design.

SK Telecom plans to apply this system to the actual environment based on its AI data center operation experience and its own AI model and service capabilities.

The three companies are also introducing a disaggregated inference method that divides the AI ​​inference process into stages and entrusts each to the appropriate semiconductor.

The prefill stage, which requires large-scale calculations, is handled by the GPU, and the decode stage, which generates answers one by one, is handled by the Rebellion NPU, which has high power-efficiency.

Through this, the goal is to lower total cost of ownership (TCO) by simultaneously increasing throughput and power efficiency, and ultimately complete the 'Sovereign AI Token Factory'.

This cooperation, which was concluded with support from the Ministry of Science and ICT, presents a new type of sovereign AI cooperation model that combines domestic AI semiconductors and models, global infrastructure technology, and data center operation know-how.

The three companies plan to pool their respective development, design, and operation capabilities to speed up commercialization and expand the foundation for providing AI services tailored to Korean and domestic data and industry-specific demands on a large scale.

Based on this partnership, Rebellion plans to speed up the development and commercialization of the next-generation NPU 'R200', which is optimized for low-latency and high-efficiency inference of Frontier AI, such as the Mix of Experts (MoE) model with trillions of parameters.

Development will continue with the goal of introducing the R200 system in 2028 by using NV Link Fusion to directly connect the NPU inside the rack to NVIDIA's high-speed scale-up network to maximize the inference performance of proprietary semiconductors.

Park Seong-hyeon, CEO of Rebellion, said, "This announcement is an important turning point in expanding Korea's Sovereign AI ecosystem. Based on our proprietary technology, we will design the next-generation AI infrastructure together with NVIDIA and SKT and ensure that Korea-created technology grows into a key axis that moves the world's AI."