Latency isn't really the issue here (lots of inference serving happening across oceans, which is probably 100-150ms RT), but bandwidth ends up becoming an issue for large scale transfers (mostly model weights, not the actual inference tokens/queries).