The latency table says it all. Cloud-to-cloud is 40ms for 20 turns. Hotel Wi-Fi is 16 seconds. You can halve inference time and still have a broken product on bad connections.