Grace Blackwell racks waste GPUs unless the scheduler keeps jobs rack-local

Grace Blackwell needs a new scheduler: 27% more usable GPUs from the same racks

NVL72 puts all 72 GPUs in a rack on one NVLink fabric, but naive schedulers scatter multi-node jobs across racks and push critical all-reduce and all-to-all traffic onto InfiniBand, where throughput can drop 43%. SkyPilot measured about 2% node unavailability on production clusters and found that 16-node slices with two spare nodes per rack yield 27% more usable capacity than full-rack allocations, while in-place recovery and defragmentation keep jobs rack-local as hardware fails.