Nvidia, Rebellions and SK Telecom Team Up on Sovereign AI for South Korea
NVIDIA, Rebellions and SK Telecom signed a memorandum of understanding on 9 November 2026 at NVIDIA AI Day Seoul at COEX in Seoul to build Korean sovereign AI infrastructure, linking Rebellions next-generation Rebel R200 neural processing units to NVIDIA platform with NVLink Fusion and deploying the combined system in SK Telecom AI data centre. The R200 systems are expected to be ready for deployment in 2028, not live now, so this is a development and deployment plan, not infrastructure already in operation.
Sovereign AI is AI built on domestic models, data, chips and infrastructure, tailored to a country language, industries and priorities, so that the country does not depend entirely on foreign systems. It does not mean every part must be made at home. Rebellions chief executive Park Seong-hyun said exactly that at the signing, noting that sovereign AI does not mean every bolt and screw must be domestic, and that it would take three to four years and massive capital just to replicate NVIDIA interconnects, switches and rack systems.
The deal was signed at NVIDIA AI Day Seoul, with Park Seong-hyun for Rebellions, Jeong Seok-geun, chief technology officer of SK Telecom, and Ashish Karandikar, vice president of NVLink Fusion at NVIDIA, posing for a photograph after the signing. Lee Do-gyu, director general of the Information and Communications Policy Office at the Ministry of Science and ICT, and Raymond Te, senior vice president at NVIDIA, also attended. The collaboration is supported by Korea Ministry of Science and ICT, which is promoting sovereign AI and domestic AI chips.
Under the agreement, Rebellions will use NVIDIA NVLink Fusion to connect its Rebel R200 neural processing units to NVIDIA open infrastructure platform for rack-scale inference, and will harness the NVIDIA MGX rack architecture for rack-scale system development. NVLink Fusion extends NVLink, NVIDIA high-speed interconnect for data transfer between NVIDIA GPUs, to AI chips made by other companies, letting specialised processors use NVLink scale-up networking designed for high bandwidth and low latency.
 |
| Illustration: A data centre corridor with rows of green-lit server racks |
How the GPU and NPU will work together
The three companies plan to apply a disaggregated inference approach to their integrated system, assigning each stage of AI inference to a different chip. NVIDIA GPUs will handle the prefill stage, which analyses the user query, while Rebellions NPUs will take on the decode stage, which generates the response, with the aim of improving both processing speed and power efficiency. An NPU, or neural processing unit, is a chip designed specifically for AI inference, optimised for low latency and energy efficiency.
That split is the technical heart of the deal. Prefill and decode have different needs: prefill is a large, parallel computation over the input, while decode is a step-by-step generation of output tokens, where latency and power matter most. By giving each stage to the chip best suited to it, the partners hope to cut the cost per token, which is the cost of generating each piece of output, and to make Korean AI services cheaper to run at scale.
Rebellions is developing the R200 with the goals of low-latency and energy-efficient inference using frontier models, including multitrillion-parameter mixture-of-experts models. Those are design objectives for a processor still under development, not performance measurements from deployed systems. A mixture-of-experts model is a large model made of many specialist parts, of which only a subset is active for each task, and the R200 is designed to run that kind of model efficiently.
SK Telecom plans to deploy the integrated systems in its AI data centre, drawing on its AI data centre operations expertise, proprietary models and services, and to validate performance and stability before assessing viability for live services and commercialisation.
 |
| Illustration: A close-up of a computer chip on a circuit board |
What SK Telecom brings to the deal
SK Telecom contribution extends beyond housing the hardware. It plans to build a full-stack sovereign AI system spanning infrastructure, models and services, connecting its proprietary AI model, including A.X K2, and public-facing services such as AI for Everyone.
Jeong Seok-geun, SK Telecom chief technology officer and head of AI CIC, said the company wants to set a computing standard in Korea first and use it to take on the global market. Park Seong-hyun said the goal is to make Korea a token exporter, meaning a country that can produce AI tokens at competitive prices and sell that capacity abroad. SK Telecom also plans a sovereign token service that lets companies and public institutions use the system without buying equipment directly.
The timing matters. The R200 systems are expected to be ready for deployment in 2028, which is more than a year away. Until then, the work is development, integration and validation, not a live service. Readers should not mistake the memorandum of understanding for a finished product, and they should not expect Korean sovereign AI services on this hardware in 2026 or 2027.
 |
| Illustration: Business leaders shaking hands at a signing ceremony in an office |
Why Korea wants its own AI stack
Korea has a strong semiconductor industry, a large domestic market, and a government that has made sovereign AI a policy priority. The Ministry of Science and ICT is supporting collaboration between NVIDIA and Korean NPU companies, beginning with Rebellions and SK Telecom, and the three companies say they will establish a Korean sovereign AI infrastructure model and lay the foundation for sovereign AI services across Korea public and private sectors.
The model is a partnership, not a go-it-alone effort. NVIDIA provides the interconnect technology and rack design that ties together multiple servers, Rebellions provides the domestic NPU, and SK Telecom provides the data centre, the model and the services. That division reflects the reality that no single Korean company can replicate the whole NVIDIA stack, and that the fastest path to sovereign AI is to combine domestic innovation with global infrastructure.
For the wider AI industry, the deal is a test of NVLink Fusion as a way to open NVIDIA platform to other chipmakers. If Rebellions R200 works well inside an NVIDIA rack, other NPU makers may follow. If the integration proves difficult, it will show the limits of mixing chips from different vendors in a single system.
For Korea, the risk is that the 2028 date slips, or that the cost per token does not fall as much as hoped. A memorandum of understanding is not a binding contract to deliver a product, and the R200 is still under development.
| Partner | Role in the deal |
|---|---|
| NVIDIA | NVLink Fusion interconnect and MGX rack architecture, open infrastructure platform |
| Rebellions | Rebel R200 NPU, connected via NVLink Fusion, for low-latency and energy-efficient inference |
| SK Telecom | Deploy systems in its AI data centre, validate performance, connect proprietary models and AI for All services |
| Ministry of Science and ICT | Government support for sovereign AI and domestic AI chips |
| Inference stage | Chip | Job |
|---|---|---|
| Prefill | NVIDIA GPU | Analyse the query |
| Decode | Rebellions Rebel R200 NPU | Generate the response |
| Deployment | Combined system | Expected ready in 2028, not live now |
The NVIDIA investment picture is covered in our Nvidia will invest 3 billion in Meta report, and the national AI strategy angle is covered in our Canada is buying into Cohere article. The open-model race is covered in our Kimi K3.1 Pro is here report, and the market context is covered in our Is AI bubble bursting article, the hub of this cluster.
Key Takeaways
- NVIDIA, Rebellions and SK Telecom signed a memorandum of understanding on 9 November 2026 at NVIDIA AI Day Seoul to build Korean sovereign AI infrastructure.
- Rebellions will connect its next-generation Rebel R200 NPU via NVIDIA NVLink Fusion and MGX rack architecture, with NVIDIA GPUs handling prefill and R200 NPUs handling decode.
- SK Telecom will deploy the combined system in its AI data centre, validate performance and stability, and connect it to proprietary models and Korea AI for All services.
- R200 systems are expected to be ready for deployment in 2028, not live now, and the deal is a development plan, not finished infrastructure.
- The collaboration is supported by Korea Ministry of Science and ICT, and aims to cut the cost per token and make Korea a token exporter.
Conclusion
The NVIDIA, Rebellions and SK Telecom deal is Korea clearest sovereign AI bet yet: foreign scale, domestic silicon and a national carrier, combined in one rack. The technology split is sensible, the government backing is real, and the 2028 date gives the partners time to build it properly. The test will be whether the R200 delivers the promised efficiency when it finally reaches SK Telecom data centre.
Sources
Unite.AI, Rebellions to Use NVIDIA NVLink Fusion for Korean Sovereign AI, 9 November 2026
Comments
Post a Comment