SK Telecom, AI startup Panmnesia partner to rewire data centers

SK Telecom and Panmnesia, an AI infrastructure link solution provider, signed an MOU at MWC 2026 in Barcelona to jointly develop a CXL-based AI data center architecture.

The goal is to improve resource efficiency and performance.

The partners plan to use Compute Express Link (CXL) to disaggregate CPUs, GPUs, and memory at rack-level using CXL Fabric Switches.

This will allow resources to be dynamically composed to reduce waste and improve cost efficiency and performance.

They also plan to replace network-based GPU collective operations with CXL links and Link Controllers to cut data copies and software intervention.

The companies say this will improve processing efficiency and throughput without adding GPUs.

🔗 Source: Panmnesia

🧠 Food for thought

Implications, context, and why it matters.

This partnership pairs a national AI champion with a specialized local hardware innovator

The deal supports SK Telecom’s push to become an AI data center (AIDC) provider, including plans for a 1 GW-scale site plus purchases of thousands of NVIDIA GPUs 1.
The teams plan to build a “CXL-Based AI Rack” using Panmnesia’s CXL Fabric Switches. Panmnesia calls its PCIe 6.4/Compute Express Link (CXL) 3.2 fabric switch silicon (PANSWITCH) the first CXL switch silicon with port-based routing (PBR), and says it fully implements the CXL 3.2 standard 2.
PBR supports flexible, non-hierarchical layouts like mesh in Panmnesia’s framing. That design can pool disaggregated CPUs, GPUs, and memory across a rack, which matches the goal SK Telecom and Panmnesia set out 3.
The announcement also mentions “Link Controllers.” These hardware components can be built into CPUs, GPUs, AI accelerators, and memory devices to enable direct communication over CXL, which cuts software involvement plus data copies compared with conventional network-based interconnects 4.

The project is a blueprint for challenging Ethernet’s dominance in AI

The effort tests CXL as the main interconnect for AI clusters, a job the companies say often falls to general-purpose networks such as Ethernet for GPU collective operations 4, 3.
If it works, CXL becomes a low-latency option for heavy workloads and it could reshape how AI data centers are built.
The work puts “composable infrastructure” into practice by tackling stranded resources. Idle GPUs or memory in one server often cannot be used by another 4.
Results at scale could speed up the CXL ecosystem by creating demand for CXL-native CPUs, accelerators, and memory modules from more vendors 5.