Any serious LLM work isn't going to use that. They'll use Metal GPU. No one is going to inference using the NPU on a Mac.

OP said "work stations" which is implying Macbook Pros and Studios.