Want to share a tech blog, which explains 3 modes of CXL Memory: CXL Memory Explained. Hopefully, it will solve the memory shortage for LLM model.

Most CXL coverage collapses three separate things into one, which is why the technology reads as both imminent and permanently delayed.

Expansion gives a single host extra CXL-attached memory as a second tier. It is in production at hyperscale — 768 GB of local DDR5-6400 alongside 256 GB of reclaimed DDR4-2400 per node, at 0.13 the cost per gigabyte of local DRAM, because those DIMMs were already bought. Meta verified this?

Pooling gives several hosts exclusive slices of a shared device, and does not require a switch: a multi-port device pools across whatever hosts plug into it. NSDI '26 measurements put multi-port paths at 260–300 ns against 490–600 ns switched.

Sharing — many hosts on one region, coherent in hardware — is defined in CXL 3.0 and absent from shipping devices.