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.