Samsung and SK Hynix Push a New CXL Memory Tier as AI Inference Outgrows HBM

Korea's memory giants are commercializing Compute Express Link modules — a layer between DRAM and storage — that they say can hold huge AI models close to GPUs at near-DRAM speed and lift inference throughput.

Jul 20, 2026Source: The Korea Herald

Samsung and SK Hynix are pushing a new class of memory into commercial use, betting that AI inference is outgrowing the expensive high-bandwidth memory stacked beside GPUs. The technology, Compute Express Link (CXL), adds a tier between conventional DRAM and slower SSD storage — a way to give AI systems far more addressable memory without relying solely on scarce, costly HBM.

What CXL changes

Large models need to keep enormous parameter sets and context close to the processor. HBM is fast but limited in capacity and supply; system DRAM is cheaper but constrained by how much a server can physically hold. CXL lets memory be pooled and shared across devices over a high-speed interconnect, expanding the effective memory a GPU can reach — directly relevant to serving big models at scale.

The demos

SK Hynix showed a second-generation 256GB CMM-DDR5 module on CXL 3.2, and its researchers reported throughput gains of up to 35.7% on inference workloads. Samsung tested a 1TB CXL memory pool that reached about 92% of DRAM-level inference performance across eight GPUs — close enough to full-speed memory to be usable in production while adding far more capacity.

Timing and dependencies

Samsung is targeting year-end 2026 production of its CMM-D module, though it could slip to 2027 because CXL adoption depends on delayed Intel and AMD server platforms that support the standard. Yim So-jung, an analyst at Eugene Investment & Securities, said CXL "really takes off from 2026," with mainstream CXL-capable servers arriving by 2028.

A second front in the memory war

For the Korean giants, CXL is a way to sell more memory into AI systems beyond the HBM they already dominate — and to address the capacity ceiling that HBM alone cannot lift. It positions memory, not just GPUs, as a lever on how large and how cheaply AI models can be served.

Why It Matters

Memory capacity, not just compute, is emerging as a hard limit on serving ever-larger models, and HBM is supply-constrained and expensive. A viable CXL tier at near-DRAM speed would change the economics of inference — and hands Samsung and SK Hynix a new growth lane tied to the AI buildout beyond the HBM they already sell.

Filed underAI Chips #AI inference #CXL memory #HBM #SK Hynix #Samsung