Apple Researchers Turn Idle Mac GPUs Into a Virtual GPU Cloud for ML Inference

Apple researchers have described Minions, a system designed to aggregate idle Apple Silicon GPUs from Macs into a dynamically scheduled virtual GPU cloud for large-scale machine learning inference. The research was outlined on October 15, 2026.

According to the paper, traditional AI inference often relies on centralized GPU clusters, which can face high costs, slow scaling and long queue times. The researchers suggest that Mac GPUs are frequently idle during everyday office, browsing and creative work, and Minions aims to make use of that unused capacity without requiring additional GPU purchases.

The system consists of four main components: a Minion Server for receiving tasks, splitting subtasks and coordinating scheduling; PostgreSQL metadata storage for recording task and device state; a Minion Executor running on macOS devices to perform local inference with Apple Silicon GPUs; and a Minion Watchdog that monitors devices going offline, sleeping or losing power, and reassigns unfinished tasks to available devices.

Minions uses a pull-based task model, in which devices actively claim executable subtasks from the server, with atomic mechanisms preventing duplicate processing. Inference results can be written to object storage, and a Grafana-based dashboard provides real-time visibility into online, busy and idle devices.

In tests, some inference tasks that previously had to queue for more than 10 hours due to limited centralized GPU resources started executing within minutes under Minions, without additional GPU hardware. The project is currently positioned more as an internal research and engineering effort than a consumer product. The team also noted that data confidentiality and governance are important considerations when using devices across departments.