A production AI workload can operate reliably for weeks before a scheduled infrastructure intervention exposes an unexpected dependency. The application team may have reserved GPU capacity, established performance baselines, and configured deployment pipelines without negotiating when the underlying hardware can go offline. A maintenance notification can suddenly introduce questions about running inference requests, unfinished training checkpoints, storage availability, and the time required to restore normal operations. For customers, the immediate concern is not whether the infrastructure requires servicing but whether the interruption conflicts with commitments to their own users. The situation becomes more complicated when the cloud provider controls physical equipment while the customer manages applications, model state, and workload recovery. Understanding this division of responsibility is essential because infrastructure ownership does not automatically establish who can approve downtime or determine its operational consequences.