# NVLink Fusion: Rebellions NPUs join Nvidia racks for SK Telecom

> Rebellions will link its R200 NPUs to Nvidia via NVLink Fusion, with SK Telecom deploying the racks. Systems are expected in 2028.

- Published: 2026-11-09
- Topic: [AI](https://futuretechnologyhq.com/topic/ai/)
- URL: https://futuretechnologyhq.com/article/nvlink-fusion-rebellions-npus-join-nvidia-racks-for-sk-telecom/
- Publisher: [Future Technology HQ](https://futuretechnologyhq.com/), free to read. Quote as "Future Technology HQ" with a link to the article.

## Key points

- Rebellions will connect its R200 NPUs to Nvidia's platform through NVLink Fusion and build on the MGX rack architecture
- SK Telecom plans to deploy the systems, with Nvidia GPUs handling prefill and Rebellions NPUs handling decode
- R200 systems are not expected to be ready for deployment until 2028
- Custom silicon inside Nvidia's rack ecosystem keeps Nvidia at the centre even as nations push for sovereign AI

## Article

Nvidia, South Korean chip startup Rebellions and telecoms group SK Telecom have agreed to put Rebellions' next-generation R200 processors inside Nvidia's rack-scale infrastructure, with SK Telecom planning to run the result in its AI data centre. Rebellions' R200 systems are expected to be ready for deployment in 2028.

## What happened

Rebellions will use NVLink Fusion to connect its Rebel R200 neural processing units (NPUs), chips built for AI workloads, to Nvidia's platform. NVLink Fusion extends Nvidia's scale-up networking, the fast links that tie chips together inside a rack, to custom processors that Nvidia itself did not design. Rebellions will also use the Nvidia MGX rack architecture and its reference designs to build the systems.

SK Telecom plans to deploy those systems in its AI data centre, alongside its own foundation models, including A.X K2, and public-facing services, among them those under Korea's "AI for All" project. Korea's Ministry of Science and ICT (MSIT) is backing the collaboration, which supports collaboration between Nvidia and Korean NPU companies, beginning with Rebellions and SK Telecom.

The R200 is aimed at low-latency, energy-efficient inference, meaning running trained models rather than training them, including multitrillion-parameter mixture-of-experts models. Rebellions' current flagship, the Rebel100, is also built specifically for inference and uses a chiplet architecture.

The companies also plan to develop disaggregated inference: splitting the stages of generating an AI response across different hardware. The plan is to use Nvidia GPUs for prefill, the stage that processes the incoming prompt, and Rebellions NPUs for decode, the stage that produces the answer token by token.

## Why it matters

Nvidia arguably gains the most. Jensen Huang's pitch is that Korea's chipmakers can "connect their custom ASICs to NVIDIA's AI factory platform". Custom silicon that might once have competed with Nvidia's GPUs is instead being pulled into Nvidia's networking, rack design and supplier base. Nvidia describes this as a "vertically integrated, horizontally open" platform, though the architecture the custom chips plug into remains Nvidia's.

Rebellions gains a shortcut. Through MGX, it gets access to an established supply chain of original design manufacturers and component suppliers, rather than having to assemble one. Nvidia says NVLink Fusion helps reduce deployment risk and accelerate time to market, for a startup whose backers include Aramco, Arm, Samsung, SK Hynix and SK Telecom.

For SK Telecom, the practical gain is in facilities planning. MGX lets GPU and NPU systems share rack footprint, power delivery and cooling, so a data centre operator does not need a bespoke design for each processor type. That is a concrete operational benefit, not a marketing one.

Seoul gets a headline outcome too. Deputy Prime Minister Bae Kyung-hoon said the deal lets Korean-developed NPUs enter the computing infrastructure market "alongside NVIDIA", and flagged further cooperation on physical AI and securing GPU supply. The second point is telling: sovereign AI here still depends on access to Nvidia GPUs.

## The bigger picture

Sovereign AI, meaning national control over compute, models and services, has often been framed as a way to reduce reliance on a single American supplier. This deal shows the compromise many governments are likely to make: domestic chips and models, but inside a foreign rack and interconnect standard. Korea keeps the language models and the silicon design; Nvidia keeps the architecture they plug into.

The prefill and decode split is also worth watching beyond Korea. The companies say the aim of the prefill and decode split is to improve inference throughput and power efficiency. If it works at SK Telecom's scale, it becomes a template for any operator running mixed GPU and accelerator fleets.

The timeline tempers the excitement. A 2028 deployment date means this is a planning story, not a procurement one. SK Telecom's statements are also forward-looking, and its filing notes technological and operational challenges as risks.

## What it means for UK and US teams

The source gives no UK pricing or availability, and no UK partner is named. The relevance for UK and US infrastructure teams is architectural. Anyone planning AI data centre capacity for the end of the decade should note that MGX is designed to let GPU and NPU systems share rack footprint, power delivery and cooling. UK operators weighing sovereign or domestic AI strategies can read the Korean model as one worked example.

## What to watch next

- **Further Korean NPU companies:** MSIT says the collaboration begins with Rebellions and SK Telecom. Additional chipmakers joining would show NVLink Fusion is becoming a national programme rather than a one-off.
- **Disaggregated inference results:** the GPU prefill, NPU decode approach is planned, not demonstrated. Published throughput and power figures would test it.
- **R200 progress toward 2028:** any slippage matters, given the single stated availability date.
- **GPU supply commitments:** Bae's reference to securing GPU supply suggests the government sees continued Nvidia access as part of the bargain.

## Key takeaways

Rebellions gets a faster route to market, SK Telecom gets standardised racks it can plan around, Korea gets a domestic chip in a global ecosystem, and Nvidia gets custom silicon inside its platform rather than outside it. The deal turns on the 2028 R200 delivery and on whether splitting inference across GPUs and NPUs delivers the efficiency gains claimed. Until then it is a credible plan with a long runway, and a clear sign of how sovereign AI is likely to be built: nationally designed parts inside Nvidia's frame.

## Sources

- [Nvidia Newsroom](https://nvidianews.nvidia.com/news/nvidia-rebellions-and-sk-telecom-announce-collaboration-to-advance-sovereign-ai-infrastructure-with-nvidia-nvlink-fusion)
- [digitimes.com](https://www.digitimes.com/news/a20261109VL228/nvidia-sk-telecom-rebellions-ai-inference-data-center.html)
