tl;dr: Meta built a custom CXL 2.0 ASIC called "Vistara" to repurpose DDR4 DIMMs from decommissioned servers into new DDR5-based "MemServer" machines, exposing the older memory as a CPU-less NUMA node via tweaked Linux kernel drivers. The approach sidesteps off-the-shelf CXL limitations (no DDR4 support, bundled DRAM, high cost) and is already running across millions of servers. Meta claims a 25% server reduction for disaggregated ML inference workloads and a 33% drop in OOM-related job failures.