Reserving GPU spares can increase schedulable rack capacity

SkyPilot explains why allocating every node in an NVL72 rack can reduce usable capacity when hardware fails. In a cited NVIDIA model at two percent node unavailability, sixteen-node slices with two spares provide 27% more usable capacity than eighteen-node slices. Rack-aware placement preserves the fast internal network, while spare nodes let recovery stay within the rack; the percentage is modeled capacity, not measured application throughput.

SkyPilot · Oct 29, 2026

Why it matters

Cluster utilization depends on failure tolerance and placement constraints as well as the number of idle GPUs.

More context

The October 29 article describes SkyPilot Platform’s topology-aware scheduling, node replacement and defragmentation. Defragmentation moves workloads that can resume safely from checkpoints. Operators considering it should map rack topology and recovery behavior first; platform access and pricing need separate confirmation.

Keep in mind

The capacity result assumes a particular cluster model and failure rate. SkyPilot’s reported operating experience does not establish the same benefit for every workload, rack size or scheduler.

Original read