Ai2 published 30 days of production scheduling data, reporting 98% delivery of owed GPU hours at 98% occupancy.
Ai2 reallocates scarce GPUs with budget-weighted fair sharing
2026-10-09Verified announcementWATCH
Ai2 replaced static GPU priority queues with hierarchical, budget-weighted fair sharing, minimum-runtime contracts and simulation on historical traces. Its 30-day production data reports delivery of 98% of owed GPU hours, 98% cluster occupancy and a drop in debug-task P90 queue time from two hours to 30 seconds.
Availability: architecture article; no public scheduler or simulator found
What this means for you
For research platforms with persistent GPU scarcity, this shows fairness, utilization and interactive latency can be balanced in one scheduling model instead of manual preemption. The scheduler and simulator are not public, so reuse the metrics and contract ideas rather than copying the implementation.
Evidence & related material
Technical details & scores
| Demand | 2-3x available GPU supply |
|---|---|
| Policy | hierarchical budget-weighted fair share plus minimum-runtime contracts |
| Production Window | 30 days |
| Reported Results | 98% owed GPU hours delivered; 13 of 15 allocations at least 95%; 98% occupancy |
| Debug Queue P90 | 2 hours to 30 seconds |
| Largest H100 Job | median 5 minutes to 24 seconds; P90 2.8 hours to 1.8 hours |
| Human Repairs | down 74% |
| Availability | architecture article; no public scheduler or simulator found |
Technical impact 3
Engineering value 5
Adoption signal 3
Maturity 4
Verification cost 4
Risk 2