How do you implement cost governance and chargeback for ML workloads on AWS?

Enforce a tagging standard through IAM and SCPs, activate cost allocation tags, build per-team and per-model dashboards with Cost Explorer or CUR in Athena, set budgets with alerts, and expose cost per training run and per thousand predictions so teams can act on it.

ML spend is unusually easy to lose track of: bursty training, always-on endpoints, notebooks that outlive their purpose, and Bedrock token charges attributable to no instance at all. Governance means attribution first, then optimisation.

1. Make tags mandatory, not aspirational. Enforce at creation time with an IAM condition:…