You can have very large MoE models run on small amounts of ram (or vram), if you can transfer the experts you need on the fly from ram to disk.

SSDs are good, but an improved Optane would've been better.