A school district DC finds UPS runtime too short for graceful training checkpoint during power events. What should the AI facility evaluation prioritize?

A school district DC that finds UPS runtime too short for graceful training checkpoint during power events should prioritize sizing power protection so GPU jobs can checkpoint or drain before hard loss of power. AI densification raises draw and makes abrupt kill expensive: unfinished epochs, corrupted checkpoints, and restart waste all hit civic budgets. Adequate UPS (and related protection) runway is the facility evaluation that matches that risk.

Expanding battery autonomy for network core and management systems for training continuity protects only infrastructure, not the GPU training jobs themselves—you still lose in-flight epochs and cannot coordinate a graceful drain across the AI hall. Installing row-level automatic transfer switches to alternate between utility feeds for ride-through only swaps feed sources; it adds no battery autonomy, so it still does not give GPU jobs enough runway to checkpoint or drain in a coordinated way across the whole hall during a longer utility event. Upgrading rack power supplies to higher efficiency models to extend backup duration during events may shave draw slightly, but efficiency alone does not create the battery runtime needed for GPU jobs to checkpoint or drain before hard cutoff.

Exam caveat: power-protection sizing must cover IT load plus the time needed for scheduler-driven checkpoint or drain policies; a UPS that only covers "a few minutes of lights" may still fail AI graceful-exit requirements under Domain 2.4-style review.

Operational check: time a controlled drain/checkpoint of representative district training jobs against measured UPS autonomy at current GPU load; raise battery/runtime or lower concurrent draw until graceful exit completes before hard cutoff.