A regional consortium shares one cloud tenant for joint model training, but each city must run production inference inside its own municipal data center. Which infrastructure pattern is this?

Shared cloud training with local on-premises inference in each city is the consortium pattern that amortizes joint training in one cloud tenant while keeping each municipality’s production inference inside its own data center. The correct design works because training cost and collaboration happen centrally in the shared tenant, but live answers and local control remain with each city’s on-prem inference path—matching both shared economics and per-city production residency/control.

Shared cloud training with local on-premises retraining in each city fails because it substitutes retraining for the required production inference in each municipal data center, so live serving and local control are not preserved despite keeping training shared. Shared cloud training with local on-premises inference in joint cluster fails because it consolidates inference into a single joint cluster rather than keeping production inference inside each city's own municipal data center, violating per-city residency and control. Shared cloud training with local on-premises inference in each city VPC fails because a city VPC is still a cloud network segment, not an on-premises municipal data center, so it does not satisfy the requirement to run production inference on-prem in each city.

Exam caveat: shared training still requires dataset agreements, tenancy isolation, and clear rules for which weights may be exported to each city’s inference DC. Operational check: document the shared training tenant scope, verify each city deploys inference on-prem, enforce data-sharing agreements, and reject shared cloud training with local on-premises retraining in each city, shared cloud training with local on-premises inference in joint cluster, or shared cloud training with local on-premises inference in each city VPC as non-compliant.