Common-Pool Bandwidth: Governing Shared Compute in Research Consortia

Shared GPU clusters exhibit the classic commons dilemma: individually rational over-reservation degrades collective throughput. Drawing on allocation logs from six research consortia (2022–2026), we show that Ostrom-style graduated sanctions outperform both market pricing and administrative quotas, raising effective utilization from 47% to 81% while reducing queue-time variance. We derive design principles for compute commons and test them prospectively in two institutions.