A wastewater plant uplink calculates stranded power because PDUs are unbalanced across GPU racks. Which design correction best improves power distribution efficiency?

When a wastewater plant uplink calculates stranded power because PDUs are unbalanced across GPU racks, the design correction that improves distribution efficiency is rebalancing PDU and rack feeds so available facility power is usable across AI racks. Stranded power is capacity the plant paid for but cannot schedule onto GPUs because some feeds sit near trip while others remain lightly loaded. Rebalance restores usable headroom without buying a new utility service.

Load-balancing GPU training jobs across multiple PDUs at the scheduler level is an operational tactic but does not fix the infrastructure root cause: facility power is delivered in unbalanced phases, so some PDUs still trip near capacity while others sit lightly loaded no matter how jobs are scheduled onto them. Increasing upstream transformer capacity to supply more power to AI racks does not correct the unbalanced phases; stranded power persists because individual feeds still cannot be fully utilized even with more upstream headroom. Deploying branch-circuit metering to visualize PDU loading across AI racks only makes the imbalance visible; it does not rebalance the feeds or turn stranded capacity into usable GPU power.

Exam caveat: AI facility power evaluation must treat usable power—not nameplate alone—as the capacity that matters for GPU expansion; imbalance is a design defect, not an accounting curiosity.

Operational check: inventory phase/feed loading per wastewater AI rack, move or rehome loads until headroom is comparable across PDUs, then re-run the stranded-power calculation and confirm more of the facility allotment can host concurrent GPU jobs.