Apple runs a distributed system called "Minions" that turns idle Macs across its network into a "virtual GPU cloud," cutting AI job queues that once stretched past 10 hours down to as little as minutes. Researchers at the company described the system in a recent arXiv paper, also published on Apple's Machine Learning Research site.

A single Slack request from an employee can set the system in motion. The paper names three components: a "Minion Server," a "Minion Executor," and a "Minion Watchdog." The system handles one kind of work only — AI inference on Apple Silicon — and splits into two job types:

• Embeddings, run with open-source Hugging Face transformer models
• Text generation from large language models, served through Ollama installed on each participating Mac

The mechanism is deliberately simple. The Minion Server breaks submitted tasks into subtasks, which sit in a PostgreSQL database. Each Mac runs a Minion Executor that pulls the next subtask over REST APIs. There is no central scheduler; the paper calls this a "pull-based execution model." The Mac reads its inputs from Amazon S3, fetches the model if needed, runs the subtask on its own GPU, and returns the result to S3.

What happens when a Mac drops out?

The Minion Watchdog tracks heartbeats that active Macs send every five minutes. When a machine goes to sleep, loses its network connection, powers off, or an owner stops the executor to reclaim the device, unfinished subtasks return to the queue. Owners can also set priorities, preferring one team's or project's jobs through "dynamic task binding," or accept the default first-in, first-out order.

Each Mac runs a whole model on its own silicon for its piece of work. The paper does not suggest that any single model is split across machines. The volunteer-computing pattern will look familiar to anyone who ran SETI@home or Folding@home: small chunks of work distributed to idle machines.

The reported gains go beyond shorter waits. The paper states that Minions requests require no new GPU requests, that infrastructure cost drops to zero, and that scaling shifts from a manual effort to an organic one. The authors do not count electricity costs or hardware wear in that calculation.

Why is Apple scraping for spare compute?

The paper lands at a moment when compute is the bottleneck for AI work. Apple said as much in its July quarterly filing, stating that it "depends on access to sufficient computing resources" in an ecosystem with "constrained supply, extended lead times, and increasing costs." The company flagged the risk that it may be unable to "secure sufficient capacity" even with its own data centers and third-party cloud providers — which makes an internal overflow pool of existing Macs look practical rather than exotic.

The supply picture around the Mac itself has been tight. Then-CEO Tim Cook said in April that Mac mini and Mac Studio supply might take months to catch up after a surge of demand for local AI hardware. The July earnings call also touched on advanced-node constraints and rising memory costs.

On the server side, Apple's Private Cloud Compute — the backend for more complex Apple Intelligence requests — now also runs on Nvidia GPUs in Google Cloud, an arrangement first announced at WWDC in June. Minions does not replace any of this. For now, the Macs serve as overflow AI capacity strictly inside Apple.

Does the data stay safe?

Confidentiality is the standing worry, even internally. The researchers observed a common concern among teams about their data running on devices outside their group. Participation itself is informal: the paper notes that team members may "send Slack messages requesting others to temporarily run the Minion Executor and target a specific project or task."

Apple's mitigation is structural. Deployments can be scoped to a team or organization, each with its own database and a separate pool of Macs.

The authors point to two future steps. They plan to extend Minions to distributed model training, and they write that if idle Macs across organizations "or potentially across the global community" can be "securely federated," they could form a shared GPU pool — a roadmap, not a shipped product, but a signal that Apple sees fleet-scale Apple Silicon as a real compute resource rather than a stopgap.