South Korea: SK Telecom, Rebellions, and NVIDIA Aim for Sovereign AI

The Korean government supports an alliance between NVIDIA, Rebellions, and SK Telecom to build a full-stack AI infrastructure. The Rebellions NPU R200 is expected to be deployed in 2028, featuring an architecture that combines NVLink Fusion and MGX racks. The partners plan a "disaggregated" inference that distributes pre-filling across NVIDIA GPUs and decoding on Rebellions NPUs, which SK Telecom aims to integrate into its public models and services, including "AI for All."

A Displayed Timeline and State Sponsorship

Rebellions' R200 systems are announced for deployment in 2028. The Ministry of Science and ICT of Korea supports the collaboration between NVIDIA and Korean NPU companies, starting with Rebellions and SK Telecom, to strengthen the country's sovereign AI infrastructure. Bae Kyung-hoon, Deputy Prime Minister and Minister of Science and ICT, believes this agreement allows NPUs developed in Korea to access the computing infrastructure market alongside NVIDIA through NVLink Fusion. He emphasizes that the state will play an active role in this process to achieve tangible results and advance cooperation, particularly in physical AI and securing access to GPUs. The companies involved seek to lay the foundations for sovereign AI services for both public and private sectors, linking national technologies to NVIDIA's infrastructure. The ministry provides official support for this partnership.

A Distributed Inference Between GPUs and NPUs

The partners envision a disaggregated inference approach, where different stages of response generation would be assigned to the most suitable type of processor. They plan to use NVIDIA GPUs for pre-filling and Rebellions NPUs for decoding, aiming to enhance inference throughput and energy efficiency. The MGX rack architecture is designed to allow GPU and NPU systems to share the footprint, power distribution, and cooling of the same data center. This organization is presented as a way to support infrastructure planning in parallel with the development of Rebellions' processors and systems.

SK Telecom Aims to Integrate and Operate in Its Own Data Centers

SK Telecom highlights its experience in building and operating AI data centers, as well as in developing proprietary foundational models and public services. The operator recalls a long-standing collaboration with NVIDIA in the field of AI and sovereign infrastructure. It plans to connect these capabilities and deploy them in its dedicated data center, relying on MGX racks, and anticipates the deployment of systems resulting from this collaboration. These projects are part of a comprehensive sovereign AI strategy, including connecting the infrastructure to public models and services, particularly those from the "AI for All" program. Jung Jai-hun, President and CEO of SK Telecom, cites SKT's proprietary models, including A.X K2.

Rebellions Prepares the R200 and Aims for Integration via NVLink Fusion and MGX

Rebellions is developing its R200 generation of NPUs, designed for low-latency and energy-efficient inference, including with multi-trillion parameter models. The company plans to connect these NPUs within a rack via the NVLink network, leveraging NVLink Fusion, and to use the MGX rack architecture and its reference designs for system development. The MGX ecosystem is expected to open up an existing supply chain of original design manufacturers and suppliers. Rebellions also indicates plans to adopt NVLink Fusion to connect its R200s to NVIDIA's open infrastructure platform and to leverage MGX for rack-scale systems. Sunghyun Park, co-founder and CEO of Rebellions, states the goal of achieving low-latency and energy-efficient inference for advanced models, aspiring to position Korean technology as a central pillar of global AI.

NVLink Fusion and MGX, Technical Foundations Claimed by NVIDIA

NVLink Fusion is described as an extension of NVIDIA's AI platform for custom XPUs, allowing the combination of specifically designed silicones with its high-bandwidth, low-latency network, as well as with an already established infrastructure ecosystem. According to NVIDIA, this method is likely to enhance performance, mitigate deployment risks, and shorten time-to-market. The announced collaboration brings together NVIDIA's open infrastructure platform, Rebellions' AI processors, and SK Telecom's data center, model, and service capabilities, aiming for AI services tailored to language, industries, and national priorities, as well as a link between local semiconductors and operational systems for large-scale AI. Jensen Huang, founder and CEO of NVIDIA, emphasizes Korea's potential and states that NVLink Fusion can connect custom ASICs to NVIDIA's AI factory platform. He also mentions working with the MSIT, Rebellions, and SK Telecom on a computing foundation for Korea to produce its own intelligence. Jung Jai-hun, President of SK Telecom, believes the agreement validates the potential for next-generation sovereign AI infrastructure, combining global technologies and Korean semiconductors.