News

Startups

Artificial Intelligence

Asia

SK Telecom and Panmnesia Partner to Revolutionize AI Data Centers with CXL Technology

South Korea's leading telco and AI link specialist Panmnesia are collaborating to build high-efficiency, disaggregated AI infrastructure.

South Korea's leading telco and AI link specialist Panmnesia are collaborating to build high-efficiency, disaggregated AI infrastructure.

NewDecoded

Published Mar 7, 2026

Mar 7, 2026

3 min read

Image by Panmnesia

SK Telecom and Panmnesia announced a strategic partnership at MWC26 in Barcelona to develop a next-generation AI data center architecture based on Compute Express Link (CXL). The collaboration aims to improve cost efficiency and performance by moving beyond traditional monolithic server designs. Both companies intend to build and validate a functional CXL-Based AI Rack by the end of 2026.

Current AI infrastructure faces limitations where CPUs, GPUs, and memory are statically bundled, leading to resource waste and high operational costs. To solve this, the partners are proposing a disaggregated architecture that separates computing resources into flexible pools. Using CXL technology allows these components to connect as if they were part of a single system, maximizing utilization rates across the entire rack.

Panmnesia will provide the foundational link solutions, including CXL Fabric Switches and Link Controllers that support PCIe 6.4 and CXL 3.2 standards. These components enable direct communication between devices without the traditional software overhead or data copying required by standard Ethernet networks. SK Telecom will lead the architectural design, drawing on its extensive experience in large-scale AI data center operations.

Suk Geun Chung, Head of AI CIC at SK Telecom, noted that AI competitiveness now depends on system-level optimization involving data flow and memory. Myoungsoo Jung, CEO of Panmnesia, added that their link semiconductors will define the future of high-efficiency AI models on a global scale. The partners hope this new standard will alleviate the structural bottlenecks currently hindering large-scale AI training and inference. The companies plan to run real-world AI models to evaluate latency and throughput performance throughout the current year. Following this validation phase, they aim to conduct proof-of-concept deployments in hyperscale environments to prepare for commercial expansion. Interested partners can already request samples of the CXL Fabric Switches and Link Controllers directly from Panmnesia for their own hardware development projects.


Decoded Take

Decoded Take

Decoded Take

This partnership represents a significant shift in how the industry handles the growing "Memory Wall" problem. While traditional data centers are limited by rigid server configurations that leave expensive GPUs underutilized, the move toward CXL-based disaggregation allows for a truly flexible resource pool. By integrating Panmnesia’s CXL 3.2 technology, SK Telecom is positioning its upcoming 1GW-class data centers to compete on economic efficiency rather than just raw hardware count. This move signals that the next phase of the AI arms race will be won through system-level architectural optimization rather than simply stacking more chips.

Share this article

Related Articles