South Korean chip startup Rebellions plans to plug its next AI processor into Nvidia’s rack-scale hardware, and SK Telecom intends to run the result in its own data centre. The three companies announced the arrangement on 9 November as part of Korea’s push for homegrown AI infrastructure.
The processor in question is Rebellions’ Rebel R200, a neural processing unit (NPU) built for inference, the stage where a trained model answers questions. Through NVLink Fusion, a program that lets outside chipmakers connect custom silicon to Nvidia’s high-speed networking, the R200 will sit in the same racks as Nvidia GPUs. Nvidia says Rebellions’ R200 systems should be deployable by 2028.
Nvidia’s announcement says the racks will follow its MGX design, a shared blueprint for rack size, power delivery and cooling. That gives Rebellions access to the existing supply chain of server makers and component suppliers rather than forcing it to build everything from scratch.
The plan includes a split of labour inside each AI response. Nvidia GPUs would handle prefill, where the model reads the prompt, while Rebellions NPUs would handle decode, where the answer is generated token by token. The companies say dividing the work this way should raise throughput and cut power use, though no benchmark figures have been published.
Software is SK Telecom’s part: it wants the racks tied to its A.X K2 model and to the state-backed “AI for All” services. Jensen Huang said the aim is for Korea to “produce its own intelligence”, and Korea’s science and ICT minister Bae Kyung-hoon said the government will back the work, including cooperation on physical AI and securing GPU supply.
Rebellions already sells an inference chip, the Rebel100, which is built as a chiplet design on open-source software, and says it has commercial deployments in enterprise and government settings. The R200 follows it, and Rebellions is pitching it at models with multiple trillions of parameters.
For Nvidia the deal shows how it is responding to countries that want some of their AI hardware made locally: instead of competing with a national chip, it offers a way for that chip to live inside its platform. Rebellions, backed by investors including Samsung, SK Hynix, Arm and Aramco, gets a route into data centres that already run Nvidia equipment.
The announcement came alongside several other Korea-focused posts from Nvidia, covering physical AI with manufacturers, agent security work with Hyundai Motor Group, and deployment recipes for Korean open models such as SK Telecom’s A.X K2 and LG’s K-EXAONE 2.0. Pricing, rack counts and the scale of SK Telecom’s order were not disclosed, and 2028 is when systems should be deployable, not a date for customers going live.
