The Allen Institute for AI’s (Ai2) infrastructure team replaced the priority-based scheduler on its GPU clusters, which serve about 150 researchers, with a system built on GPU time budgets, hierarchical fair-share allocation, and a “scheduling contract.” Under the new approach, workloads declare a minimum protected runtime before they can be preempted, and can later be requeued so the scheduler can rebalance capacity. In a 30-day test described in an October 9, 2026 blog post, teams received 98% of the GPU hours they were owed, and the p90 queue wait for debug workloads fell from about two hours to 30 seconds.
