SK Telecom, Nvidia, Rebellions join forces to build Korea-specific AI infrastructure

SK Telecom is teaming up with Nvidia and Rebellions to accelerate the development of Korea-specific AI infrastructure.

The three companies plan to apply a "disaggregated inference" approach using Nvidia's NVLink Fusion technology, with the goal of boosting processing speed and power efficiency.

SK Telecom announced Monday that it signed a memorandum of understanding with Nvidia and Rebellions at the "Nvidia AI Day Seoul" event at COEX in Seoul to jointly develop an integrated GPU-NPU system for building Korea-specific sovereign AI infrastructure.

Under the agreement, the three companies will use Nvidia's NVLink Fusion technology to connect Rebellions' NPU under development, the Rebel R200, to Nvidia's platform. They also plan to apply Nvidia's MGX server rack reference design standard to develop a rack-level integrated GPU-NPU system.

NVLink Fusion extends NVLink — Nvidia's high-speed interconnect technology for data transfer between Nvidia GPUs — to AI chips made by other companies.

The move comes as the rapid spread of agentic AI has driven a surge in inference requests, making processing performance, power efficiency and operational cost competitiveness increasingly critical for AI data centers.

The inference process is broadly divided into two stages. The "prefill" stage, which analyzes a user's input query, is computationally intensive and well-suited to high-performance chips such as GPUs. The subsequent "decode" stage, which generates a response word by word, relies more on the ability to retrieve data quickly from memory than on raw computation. The three companies will apply a disaggregated inference approach to their integrated system, assigning each stage to a different chip: Nvidia GPUs will handle the prefill stage for query analysis, while Rebellions NPUs will take on the decode stage for response generation, with the aim of improving both processing speed and power efficiency.

SK Telecom plans to deploy the integrated system in its AI data centers to validate performance and stability, and will assess its viability for use in live services and commercialization. The company also intends to connect the system to its proprietary AI model and its mass-market AI service "Modu's AI" to deliver a full-stack sovereign AI offering spanning infrastructure, models and services.

"This collaboration is significant in that it validates the potential of a Korea-specific sovereign AI infrastructure that combines the strengths of Nvidia GPUs and Rebellions NPUs to enhance inference performance and power efficiency," SK Telecom Chief Technology Officer Jeong Seok-geun said. "We will strengthen our full-stack sovereign AI competitiveness and build a sovereign AI infrastructure ecosystem, drawing on our experience in AI data center construction and operation, our proprietary AI model A.X K2, and capabilities such as Modu's AI."