High throughput with low latency WAN links between data centers is the most important networking feature for supporting distributed training of large AI models. Distributed training across multiple data centers requires rapid exchange of gradients and model parameters, which demands high-bandwidth, low-latency connections (e.g., InfiniBand or high-speed Ethernet over WAN). NVIDIA's "DGX SuperPOD Reference Architecture" and "AI Infrastructure for Enterprise" emphasize that network performance is critical for scaling AI training geographically, ensuring synchronization and minimizing training time.
QoS policies (B) prioritize traffic but don't address raw performance needs. Segregated segments (C) enhance security, not training efficiency. Wireless networking (D) lacks the reliability and bandwidth for data center AI. NVIDIA prioritizes high-throughput, low-latency networking for distributed training.

Optimizing cost and performance in cloud-based AI training with NVIDIA GPUs (e.g., DGX Cloud) requires resource efficiency. Autoscaling dynamically allocates GPU instances based on workload demand, scaling up for peak training and down when idle, balancing performance and cost. NVIDIA's cloud integrations (e.g., with AWS, Azure) support this via Kubernetes or cloud-native tools.
High core count (Option A) boosts performance but raises costs if underutilized. Data locality (Option C) reduces latency but not overall cost-performance trade-offs. Reserved instances (Option D) lower costs but lack flexibility. Autoscaling is NVIDIA's key cloud optimization factor.