CXL Consortium Recap: Scaling AI Infrastructure at AI Infra Summit 2026

AI Infra Summit 2026 brought together data center architects, system designers, and infrastructure leaders in Santa Clara, CA, from September 15 – 17. At the event, the CXL Consortium demonstrated how the latest developments in CXL technology help the industry break through the memory wall facing AI workloads.

At the CXL Pavilion, Consortium member companies Liqid, Primemas and VIAVI showcased live demonstrations spanning memory expansion, memory pooling, composable infrastructure, and CXL 3.2 interoperability. Consortium representatives and member experts also took the stage for a Demo Theater session and a panel discussion on the evolving relationship between AI model architecture and hardware.

Missed the show or want a closer look? Read on for session highlights, video interviews with our demo partners, and CXL news from the event.

Demo Theater Session: Boost Your AI Performance With CXL

On Wednesday, September 16, CXL Consortium Board Director Siamak Tavallaei of Samsung moderated a Demo Theater panel, exploring how CXL is transforming AI infrastructure design. Panelists Sandeep Dattaprasad of Astera Labs, Hasan Al Maruf of Meta, and Joe Baco of Primemas explored how CXL memory devices support large language models and retrieval-augmented generation (RAG) pipelines, improve accelerator efficiency, and enable larger model contexts.

The panel also addressed infrastructure bottlenecks in hyperscale and enterprise deployments. Panelists shared real-world deployment considerations, the performance improvements CXL can deliver, and how CXL enables scalable memory expansion, memory pooling, and more efficient resource utilization across CPUs, GPUs, and accelerators.

Panel Session: Scaling the AI Memory Wall – Where Model Architecture Meets Hardware

Later that afternoon, CXL Consortium Board Chair Dr. Debendra Das Sharma of Intel led a panel with Nuwan Jayasena of AMD, Karthik Ganesan of Micron and Andy Mills of Penguin Solutions. The group explored how memory technology roadmaps must keep pace with growing model sizes, longer context windows, and the shift toward inference and agentic AI.

Video Interviews: CXL Pavilion

We connected with member companies demonstrating CXL solutions across the show floor. Watch the interviews below to see how the CXL ecosystem is putting memory expansion, pooling, and sharing to work for AI infrastructure.

Liqid

Liqid demonstrated its software-defined memory and GPU pooling infrastructure. The disaggregated memory pooling solution offers up to 160 TB of DDR5 as an addressable pool that Liqid Matrix software can dynamically allocate across servers.