SK Telecom, Nvidia and Rebellions to Build GPU-NPU AI Systems

Splitting inference between chips

The system relies on disaggregated inference. Nvidia GPUs handle the prefill stage, which analyzes the user's question and is compute-heavy. Rebellions NPUs handle the decode stage, which generates the answer step by step and depends more on how quickly data can be fetched from memory. Matching each stage to suitable silicon is meant to raise throughput and power efficiency as agentic AI drives up inference demand.

SK Telecom CTO Jung Seok-geun said the goal is for Korea to produce tokens more cheaply than anywhere else, so that AI is used widely in the country. He described the business as "sovereign tokens": SK Telecom would operate GPU-NPU data centers as a specialist provider, so that government bodies, enterprises and consumer service operators can use tokens through an API without managing hardware. He argued that asking customers to buy Korean NPUs and build their own data centers would be a barrier to entry.

Next steps

SK Telecom plans to deploy the system in its data centers to verify performance and stability, and to study real-world use and commercialization. It intends to connect its A.X K2 model and public-facing services such as "Everyone's AI" to the infrastructure, offering a full-stack sovereign AI package. The three firms say the work fits the science ministry's sovereign AI and domestic AI chip policies.

Jung said domestic chipmakers like Rebellions have secured funding and talent, but selling chips to general companies or exporting them remains difficult. Meaningful local reference deployments and customer support, he argued, are needed. Ashish Karandikar, Nvidia's vice president for NVLink Fusion, said the technology lets Korean ASICs combine with Nvidia's AI factory platform.