Power oversubscription
Provisioning more IT than the power can sustain at full draw, relying on workloads rarely peaking together.
Current numbers
~3% training / ~21% inferencepower oversubscription headroom: synchronous training vs inference (correlated vs uncorrelated peaks)