When Schedulers Meet Converters: A Case for Energy-Readiness-Aware AI Workload Placement

Large AI training jobs create synchronized, multi-megawatt power ramps as accelerators move between compute, communication, and checkpoint phases. Existing responses largely react to measured or imminent power events through local feedback, storage, GPU power control, or workload shaping. We argue that the compute stack holds a complementary resource: advance knowledge of workload intent. We propose energy-readiness-aware orchestration, in which the power domain exports a bounded dynamic-load contract, rather than internal control-loop parameters, and the compute domain uses placement, admission, and limited timing flexibility to reduce anticipated contract stress. An advisory intent channel may pre-position storage or converter operating modes before a predicted event, while local converter control and hardware protection remain authoritative. We formalize this as a robust placement-and-offset problem spanning baseline load, prediction error, resource feasibility, network locality, ramp limits, and an operator-defined stability-risk boundary. A reduced-order synthetic experiment, evaluated with perfect realized-profile knowledge as an oracle upper bound, shows lower ramp-related stress metrics for repetitive, partially decorrelatable workload transitions. It does not establish SST stability, suppress the most severe checkpoint-synchronized transient, quantify performance cost, or compare against optimized autonomous power-side conditioning. The contribution is a cross-layer problem formulation, an explicit separation of control authority, and a falsifiable evaluation agenda linking AI workload scheduling with solid-state-transformer and DC-distribution design.