Meta's Vistara Tech Reuses Retired DDR4 Memory in AI Servers via CXL Controllers

In 2026, memory prices rose significantly, with DDR5 memory prices climbing more than fivefold from the same period in 2025, and some models seeing increases of up to tenfold. Demand from AI servers for high-specification memory has led to supply tightness and rising costs; equipping all servers with high-spec memory is not only constrained by insufficient supply but also makes cost control difficult. Server makers had previously scrapped large quantities of DDR4 memory, and these modules are now being redeployed in AI servers.

Meta has developed Vistara technology, which uses CXL controller chips to connect DDR4 memory to platforms that require DDR5 buses, allowing retired DDR4 modules to be put back into service. Servers typically have a service life of four to five years, whereas DDR4 memory can last 10 to 12 years, so reusing these modules lowers costs and reduces the need for DDR5 memory. Although DDR4 offers lower bandwidth and latency than DDR5, many AI workloads place greater emphasis on memory capacity, and expansion through the CXL interface can stack two to 2.5 times the memory capacity.