SK Telecom is teaming up with global semiconductor design corporations Arm and domestic AI Semiconductor startup Rebellions to develop next-generation inference servers for artificial intelligence data centers (AI DC). As the center of gravity in the large language model race rapidly shifts from training to inference, the move is seen as a bid to break away from an all–high-power graphics processing unit (GPU) structure and secure a low-power, high-efficiency infrastructure.
SK Telecom said on the 10th that it signed a strategic memorandum of understanding (MOU) with Arm and Rebellions to jointly develop a server solution that combines Arm's data center "Arm AGI CPU" with Rebellions' inference NPU "RebelCard." The completed servers will be actually deployed in SK Telecom's AI data center to verify performance and stability. Arm AGI CPU is the first data center CPU that Arm has directly unveiled in its 35-year history, and RebelCard is scheduled for release in the third quarter of this year.
The core of this collaboration is a heterogeneous computing architecture that divides roles between the CPU and NPU within a single server. The CPU handles general-purpose tasks such as data input/output, memory management, network processing, and job scheduling, while the NPU is dedicated to AI inference computation. The industry expects that as Generative AI services expand into an inference phase that runs 24/7, the importance of this architecture—which can reduce both power efficiency demands and total cost of ownership (TCO)—will grow.
Arm and Rebellions already demonstrated in real time an agentic AI service based on GPT OSS 120B by combining the two chips at the "Arm Everywhere" event on the 3rd. This time, they aim to push a step beyond the proof stage and verify commercial viability in a telecom AI data center environment.
SK Telecom is also considering running its AI foundation model "A.X K1" on the servers. Beyond a simple chip collaboration, the strategy appears to be to strengthen AI DC competitiveness with a "full-package" business that bundles inference infrastructure with its in-house model.