Water-treatment plants need pump anomaly detection on-site, while model training will run centrally in the city data center. Which pattern best describes this AI infrastructure design?

Edge inference at water-treatment plants with centralized on-premises training in the city data center correctly describes the utility AI infrastructure pattern. Local anomaly detection must run on-site for timely pump alerts, while heavier model training uses central city-DC GPUs where data science capacity and governance live. Separating plant-local inference from central training matches latency, availability, and operational control needs in SCADA-adjacent environments.

Centralized inference in the city DC with decentralized training at the plants fails because it reverses the required split, placing latency-sensitive pump detection away from the plant and moving training off the central GPUs where it belongs. Cloud-based training with cloud-based inference for streaming pump telemetry data fails because it depends on continuous cloud or WAN paths for on-site detection and shifts training out of the required on-premises city data center. Edge training and edge inference at the plants with nightly model sync to the city DC fails because it conducts training at the edge rather than centrally in the city DC as specified.

Exam caveat: edge-to-DC model promotion needs integrity checks so corrupted or unapproved weights are not pushed to plants. Operational check: confirm plant boxes infer offline/on-site, confirm training jobs run in the city DC on approved datasets, measure alert latency at the plant, and reject centralized-inference, cloud-streaming, or plant-trained substitutes for this SCADA-adjacent design.