SI Agent News
DevelopingChipsBusiness

NVIDIA and SK Telecom partner with Rebellions for sovereign AI infrastructure

A new partnership aims to build power-efficient AI systems in South Korea by combining NVIDIA's interconnects with domestic chip designs.

Risograph poster: picture for the story 'NVIDIA and SK Telecom partner with Rebellions for sovereign AI infrast'
AI-made risograph poster for this story. No real people or products are shown.
2028Target for commercial deployment

The short version

Why it matters

This collaboration allows South Korea to develop its own AI infrastructure while still leveraging NVIDIA's dominant ecosystem. It provides a blueprint for how nations can balance domestic hardware production with global technology standards.

A new sovereign AI alliance

NVIDIA, SK Telecom (SKT), and the South Korean chipmaker Rebellions have signed a memorandum of understanding to build sovereign artificial intelligence infrastructure. The South Korean government's Ministry of Science and ICT (MSIT) is providing support for this project to develop local computing capabilities. The goal is to create a system that allows the nation to process its own data domestically while reducing reliance on foreign hardware.

The partnership was officially announced during NVIDIA AI Day Seoul in November 2026. It brings together three distinct entities: a global hardware leader, a major telecommunications provider, and a domestic semiconductor startup. This collaboration is intended to provide a blueprint for research institutions and university computing facilities to deploy frontier models.

By combining these resources, the partners aim to create a full-stack sovereign AI ecosystem. SK Telecom will provide the data centers and proprietary models, while Rebellions provides the specialized silicon. NVIDIA provides the high-speed interconnect technology that allows these different components to work together seamlessly within a single rack.

How the hardware works

The technical core of this partnership is a disaggregated inference architecture. This means the system splits the workload of a large language model into two different stages based on what the hardware does best. During the prefill stage, the system processes the initial input prompt in parallel, which is a compute-heavy task well-suited for NVIDIA GPUs.

The second stage is the decode stage, where the model generates output tokens one by one. This part of the process is limited by memory bandwidth rather than raw compute power. To solve this, the partners will use Rebellions’ Rebel R200 neural processing units (NPUs) to handle the token generation. This specific split helps maximize the efficiency of both types of processors simultaneously.

To connect these different chips, the group will use NVIDIA NVLink Fusion. This is a scale-up interconnect that allows custom third-party chips to communicate directly with NVIDIA GPUs at high speeds. By using this technology, the system eliminates the latency bottlenecks typically found in standard PCIe connections. The entire setup will be housed in NVIDIA MGX modular server architectures to ensure compatibility with standard data center power and cooling.

Impact on research and academia

This new infrastructure model has significant implications for university computing centers and national laboratories. These institutions often face strict limits on power consumption and cooling capacity. By offloading the token generation to energy-optimized NPUs, researchers can reduce the power used per token while maintaining low latency for conversational models.

For AI researchers, the use of NVLink Fusion provides a new environment for studying heterogeneous memory hierarchies. It allows them to experiment with mixed-precision execution and KV-cache transfers without being locked into a single proprietary host bus. This could lead to more diverse methods for training and running large-scale Mixture-of-Experts (MoE) models.

From a governance perspective, sovereign infrastructure allows universities to keep sensitive data secure. This includes private medical files, biological data, and unique patent information. Serving models on domestic sovereign clouds helps these institutions comply with strict data protection mandates while still accessing high-performance computing resources.

Roadmap to 2028

The consortium has established a clear four-stage roadmap to bring this technology to market. The first phase, spanning 2026 and 2027, focuses on silicon co-design and interconnect validation. During this time, Rebellions and NVIDIA will finalize the Rebel R200 design and test it inside prototype MGX chassis.

In 2027, SK Telecom will deploy these prototypes in its domestic data centers. This phase is dedicated to evaluating the system's thermal stability and power efficiency when running the A.X K2 model. These tests will ensure the hardware can handle real-world workloads before it is scaled up for public use.

The final commercial launch is targeted for 2028, where the Rebel R200 rack systems will begin supporting public services. Following the domestic rollout, the consortium plans to market this sovereign AI infrastructure package to global telecom operators. This move aims to provide other nations with a similar blueprint for building their own independent AI capabilities.

What to watch

Try it

Look up the 'AI for All' project by SK Telecom to see how they plan to deploy these sovereign models to the public.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news