Who cares about a few ms more latency on an LLM API? Maybe for voice, but most other use-cases are quite latency insensitive.