CXL memory expanders are server-only - no consumer path in 2026

Shipping CXL expanders (Marvell Structera X 3406/3506 CXL 3.2, Astera Leo, Samsung CMM, XCENA MX1) are E3.S/E3.C modules for hyperscaler/server platforms; a card in a consumer board enumerates as nothing.

CXL 2.0+ needs CPU/root-complex support that consumer Intel/AMD platforms don't enable. Even where it works, one PCIe 5.0 x16 CXL link is ~64 GB/s - WORSE than a single DDR5 channel pair - and adds ~200-400ns latency: a capacity play, not a bandwidth play, so it does not help decode speed at all.

None today. If you already own a CXL-capable server platform (Sapphire Rapids/Genoa+) you have 8-12 DDR5 channels and don't need the expander for speed. Revisit when consumer chipsets ship CXL root ports (~2027 Nova Lake/Medusa-class, optimistic).

The future use case is cheap far-memory so a 400B model LOADS while a GPU decodes from fast memory - not relevant to a low-budget build today.