A municipal CIO asks which lifecycle activities consume the GPU training cluster versus the inference farm. What mapping is most accurate?

The accurate mapping for a municipal CIO is that training and experimentation stress the GPU training cluster, while deployed serving uses the inference farm. Lifecycle stage determines which footprint pays for the work: weight updates and trials on training capacity; production classifications and generations on inference capacity. Data curation and feature engineering are upstream data-prep steps rather than the GPU training load, and while live scoring does use the inference farm, this option misses training and experimentation as the defining training-cluster work, so it is not the most accurate mapping. Nightly retraining and validation do stress the training cluster, but registry archiving is model-store governance and storage rather than inference serving, so placing archiving on the inference farm makes this mapping inaccurate. Model tuning and checkpointing are only subtasks of training and experimentation, and offline batch scoring is only one serving pattern rather than deployed serving generally, so while the pools are directionally right, this option is too narrow to be the most accurate mapping. Exam caveat: small environments may colocate training and inference on shared GPUs; the conceptual mapping still holds for capacity planning and isolation even when physical separation is incomplete. Operational check: chargeback and schedule long training jobs on the training cluster, reserve inference SLOs on the serving farm, and prevent experiment sweeps from preempting live municipal services.