A preprint posted on arXiv on 9 October by Apple researchers introduces Minions, a system that aggregates idle Apple Silicon GPUs across macOS devices for machine-learning inference tasks. It dispatches workloads to available resources and describes fault tolerance, flexible task execution and separate executor pools scoped to teams or organizations. The authors say the approach uses existing devices without buying additional GPU hardware.

The paper reports internal use across multiple teams and says that, when centralized GPU resources were constrained, some jobs that previously waited in queues of more than ten hours could begin within minutes. It presents before-and-after operational indicators, not an independent performance benchmark, total throughput measure or broadly transferable cost analysis. Minions is designed to address data-governance concerns through team- or organization-specific PostgreSQL databases and isolated executor pools. These are results reported from Apple’s own deployment experience, not a generalized cost promise for commercial cloud services; the work is a preprint that has not yet undergone peer review.