“For large model training where each step takes several seconds, inter-data center latency is not a significant bottleneck, whereas high bandwidth is critical.”