Mechanism: Managers allocate GPU-time budgets through a research hierarchy; fair-share scheduling compares workload occupancy against those allocations, while a declared minimum runtime grants bounded preemption protection. Unallocated jobs can use spare capacity but remain preemptible.
Why now: The previous priority system produced priority inflation, GPU squatting and time-consuming negotiation around non-preemptible work. Ai2 says the replacement shifts decisions about relative research value into budget reviews, while the scheduler handles changing job arrivals and occupancy. This separates strategic allocation from day-to-day scheduling without pretending demand can be forecast exactly.
Implications: Budgeting makes idle or strategically inflated reservations consume a project allocation rather than appearing cost-free. A bounded runtime guarantee can protect meaningful progress while enabling later rebalancing and requeueing. The reported 30% additional compute is a researcher’s perception, not an independently measured cluster-wide result.
Evidence boundary: The evidence is Ai2's own operational report; it documents its stated design, internal deployment context and reported problems, but offers no independent evaluation or proof of portability.
Watch: Whether Ai2 publishes measured before-and-after utilization, queue latency, preemption frequency and completed-work data.