Researchers present a way to pool idle Mac GPUs for AI inference

A new research paper presents a way to turn idle Macs into a shared pool of AI computing capacity. Minions, submitted to arXiv on October 9, 2026, aggregates Apple Silicon GPUs across macOS devices for machine-learning inference—running trained models to produce outputs. The aim is to let teams run workloads on existing hardware without buying additional GPUs.

The paper is by Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter. Apple’s machine-learning research site also lists it as an October 2026 publication. The arXiv record describes a three-page paper with one figure and one table, under distributed computing and machine learning.

Capacity already sitting on desks

The authors frame Minions as a response to three problems with centralized GPU infrastructure: high cost, limited ability to adjust capacity, and long waits when resources are constrained. Their argument concerns access as well as expense. A team needs computing resources available when it has inference work to run, not simply hardware somewhere in a centralized system.

Their alternative starts with a different observation: modern Apple silicon Macs have powerful graphics processors that remain largely idle during normal use. Minions brings those processors into a distributed system, making multiple macOS devices the source of capacity for large-scale inference workloads.

Our experience shows that idle macOS devices can be transformed into a practical distributed GPU inference platform in environments where centralized GPU resources are constrained.

The Minions paper’s authors, in the abstract

Flexible tasks and fault tolerance

The authors say the system supports flexible execution of inference tasks, fault tolerance and efficient resource use. Those capabilities address how the pool operates, not just which processors it contains. Fault tolerance means a system can cope with failures; it does not mean that individual machines never fail. Resource utilization concerns how effectively the available computing capacity is used.

The abstract gives a qualitative assessment rather than measured throughput or cost savings, so it does not establish how pooled Macs compare with dedicated GPU infrastructure on performance or total expense.