Granite or sapphire rapids are very under rated for MoE inference loads. But you need a GPU for the KV cache. Plus many boards also support CXL for RAM expansi

Granite or sapphire rapids are very under rated for MoE inference loads. But you need a GPU for the KV cache. Plus many boards also support CXL for RAM expansion over PCI 5! Source: building a hybrid inference business for regulated industry workloads.