SK Telecom is moving to develop an artificial intelligence (AI) inference server solution that combines a graphics processing unit (GPU) and a neural processing unit (NPU).

SKT said on Thursday it signed a strategic partnership memorandum of understanding with global chip designer Arm and AI chip startup Rebellion to drive innovation in next-generation AI infrastructure.

The three companies will jointly develop a solution to boost AI inference performance by installing Arm’s newly launched Arm AGI CPU and Rebellion’s Revel Card, scheduled for release in the third quarter, in the same server. They plan to validate the solution at SKT’s AI data center (DC).

SKT pointed to a shift in the AI industry paradigm as the background to the cooperation. It said the key task for AI infrastructure is changing from massive computing power for training to how quickly and cheaply AI services can be provided.

AI inference, unlike training, involves processing relatively light computations quickly and repeatedly. GPUs can be used for inference, but they consume excessive power and are costly. As a result, the industry is seeing inference-optimised NPUs emerge as a new alternative.

Arm AGI CPU is a data center processor that Arm is producing itself. It is optimised for AI inference services. Rebellion’s Revel Card is also an NPU specialised for large-scale AI inference.

SKT will install the two chips in a single server. Arm and Rebellion in March demonstrated in real time an agentic AI service based on OpenAI’s language model by combining their chips at an Arm Everywhere event held in Britain.

SKT plans to install servers using the solution at its AI DC to verify performance and stability. It will also review operating its independently developed AI foundation model, A.Dot X K One (A.X K1). SKT plans to secure low-power, high-efficiency AI inference infrastructure and strengthen competitiveness in its AI DC business.

Jae-shin Lee (이재신), head of AI business development at SKT, said, "By providing a full package that combines inference-optimised infrastructure with our proprietary foundation model A.X K1, we will further strengthen our AI DC competitiveness."