GPUs are still software programmable.

An "LLM chip" does not need that and so can be much more efficient.