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

Yeah, the customer support use case is already grinding me even worse than offshoring to India.