As the Big Three Memory Makers Pull Back from In-House Development, the CXL Value Chain Is Being Redrawn

On July 20, reports emerged that Samsung Electronics, SK hynix, and Micron were adjusting their CXL controller development and commercialization strategies. Micron has reportedly moved to an external controller solution, SK hynix has reassigned some of its engineering resources to processing-in-memory, or PIM, and Samsung is said to be retaining advanced research while no longer pursuing its in-house controller as a standalone commercial product. None of the three companies has issued a formal announcement, so the details still require confirmation. Yet viewed against CXL’s development over the past several years, the shift is not particularly surprising.

This article starts with that news and moves on to the more important questions. What problem does CXL actually solve? Why is AI inference, rather than training, likely to be its first meaningful use case? How far have Meta, Google, Microsoft, AWS, Alibaba Cloud, and ByteDance progressed? Why are NVIDIA and Huawei unlikely to place CXL at the center of their interconnect architectures? And if the three major DRAM vendors step back from in-house controller development, what opportunities are left for Astera Labs, Marvell, Montage Technology, and the rest of the ecosystem?

We previously published a dedicated analysis of CXL:

CXL in the AI Era: Revolution or Hype? How Tech Giants Are Betting

CXL Is Not Being Abandoned. In-House Controller Development Is.

The headline can easily be read as saying that the three largest memory vendors are abandoning CXL. A more accurate interpretation is that they are scaling back the internal development and external commercialization of CXL controllers. Their future CXL memory products may rely more heavily on third-party controller silicon.

Those are fundamentally different things.