Borrowing The Night Reclaiming Idle Inference Gpus For Research

Runway lends production GPUs to research when customer demand falls, then returns them before demand rises. Its capacity-controller account explains why this requires advance planning: moving hardware between its separate production and research environments takes 20 to 60 minutes.

The controller builds schedules from recent demand and checks proposed allocations with queue simulations. Research jobs using borrowed capacity can be interrupted when production reclaims it, while dedicated research capacity stays in place. This extends the company's earlier sharing system, which used reserved quotas and interruptible borrowing of idle GPUs within a cluster.