Oversubscription
Provisioning less network bandwidth than full non-blocking would require; 1:1 is non-blocking, 3:1 is oversubscribed.
Current numbers
~3% training / ~21% inferencepower-budget headroom oversubscription, training vs inference fleets
~21% vs ~3%power-oversubscription headroom: inference (uncorrelated per-request peaks) vs training (synchronous peaks)
~31%back-end fabric cost avoided with 2:1 oversubscription (virtual-modular-switch design) vs the non-oversubscribed reference — where the workload allows (contested — single-source)