Samsung plans Q4 launch of AI server memory modules

Samsung Electronics plans to send samples of its CXL 3.1 memory modules for AI servers in the third quarter.

The company aims to begin mass production in the fourth quarter as demand grows for more flexible data center memory.

The CMM-D 3.1 module combines DRAM and a CXL controller on one board and supports up to 1 terabyte with bandwidth of 72 gigabytes per second on PCIe 6.0.

The report said Samsung earlier sent CXL 2.0 samples to more than 40 companies including Microsoft and Amazon while Yole Group projects the global CXL market will rise from US$2.1 billion in 2026 to US$16 billion by 2028.

🔗 Source: The Korea Herald

🧠 Food for thought

Implications, context, and why it matters.

Hyperscalers turn to CXL for memory limits

Compute Express Link (CXL) is gaining ground as AI systems run into memory limits. Large language models can need about 80 to 120 gigabytes of memory per graphics processing unit (GPU) for key-value cache, which stores earlier tokens to speed replies 1.
CXL adds about 70 nanoseconds of latency versus direct-attached DRAM. Engineers at Meta and Google say that delay works for colder data that is accessed less often 2.
Support is spreading across the stack. Server central processing units (CPUs) from AMD and Intel now support CXL, and Microsoft plans a public cloud preview of CXL-attached memory in November 2025 1.
Microsoft found that CXL memory pooling could trim server costs by 5% by lowering total memory needs by about 10% 1.

CXL fits the AI data center expansion

Samsung’s CXL work lines up with a huge AI buildout. Meta says it is building multi-gigawatt clusters, or very large groups of connected servers, including Prometheus in 2026 and Hyperion, which it expects to reach 5 gigawatts over several years 3.
That scale raises pressure to use hardware better. CXL can add memory capacity and improve use of existing systems without constant upgrades 1.
Heavy demand has given more leverage to suppliers such as Samsung and SK hynix, a South Korean memory chipmaker. Analysts call it a structural change rather than a normal cycle 4.
Memory suppliers are also shortening contracts with large technology companies and using post-settlement pricing to reset prices more often in a tight market 5.