GPU-based compute connected by a high-throughput, low-latency network best satisfies the combined training and inference requirements. Model training performs large numbers of parallel matrix and tensor operations, which GPUs execute far more efficiently than general-purpose CPUs. Frequent retraining also requires rapid movement of training data between storage, host memory, accelerators, and-when training is distributed-between GPUs on different servers. High-throughput networking prevents the accelerators from sitting idle while gradients, parameters, or data batches are exchanged. Real-time fraud inference adds a latency requirement: transaction features must reach the model quickly and predictions must return within the application's response window. Cisco's AI/ML data-center blueprint emphasizes predictable, high- performance networking for accelerator clusters, while Cisco UCS GPU platforms are designed for high- throughput inference with minimized latency. CPU-only instances may handle lightweight inference but are inefficient for frequent large-scale training. Local SSDs improve data access but do not provide the required inter-compute fabric. Generic cloud elasticity can add capacity, yet Option C does not guarantee GPU acceleration or the deterministic low-latency east-west connectivity stated in the question. Dedicated multicore CPUs and distributed storage remain limited by CPU parallelism and network overhead. Option B directly couples the correct accelerator architecture with the fabric characteristics needed for scalable training and immediate inference, making it the complete solution. Cisco AI/ML data-center networking blueprint

We are talking about management only, so SSM is out. Since we need to separate between data and control plane, we should use mgmt0 interface, so SVI interface is out. Since we need to use encryption and authentication, this can be done in SNMPv3, while telnet does not support encryption, so telnet is out.