Apple researchers cut AI job wait times from 10-plus-hour GPU queues to as little as minutes by distributing stalled workloads to idle Macs, according to an arXiv paper describing an internal system called "Minions." The system, also published on Apple's Machine Learning Research site, forms what the authors call a "virtual GPU cloud" and runs inference only, on Apple Silicon.

A request can start with a Slack message. The paper's authors observed a common concern among teams: their data running on devices outside their group.

How does the Minions architecture work?

The system has three components:

• Minion Server — splits submitted tasks into subtasks stored in a PostgreSQL database.
• Minion Executor — runs on each Mac, pulls the next subtask over REST APIs; there is no central scheduler in this "pull-based execution model."
• Minion Watchdog — tracks heartbeats that active Macs send every five minutes and requeues unfinished subtasks when a machine sleeps, drops off the network, powers off, or the owner stops the executor.

Each Mac reads inputs from Amazon S3, fetches the model if needed, runs the subtask on its own GPU, and returns the result to S3. The paper does not suggest any one model is split across machines — each Mac runs a whole model on its own hardware with unified memory.

The system handles two kinds of jobs: embeddings using open-source Hugging Face transformer models, and text generation from large language models through Ollama, installed on each participating Mac. Owners can prioritize certain teams' or projects' jobs through "dynamic task binding," or fall back to the default first-in, first-out scheme. The architecture echoes volunteer-computing projects such as SETI@home and Folding@home, which distributed small work units to idle machines.

Takeaway: a pull-based design with five-minute heartbeats lets Apple harvest spare Mac capacity without a central scheduler.

What does Minions save Apple?

The paper cites three gains beyond shorter queue times:

• Requests no longer require new GPU allocations.
• Infrastructure cost drops to zero (electricity costs and hardware wear are not counted).
• Scaling shifts from manual effort to organic growth as more Macs join.

Participation is informal. Team members, the authors write, may "send Slack messages requesting others to temporarily run the Minion Executor and target a specific project or task." Apple mitigates confidentiality concerns through team- and organization-scoped deployments, each with its own database and separate pools of Macs.

Takeaway: the savings come from reusing hardware Apple already owns, not from new infrastructure.

Why now? Compute is the bottleneck

Apple acknowledged the pressure in its July quarterly filing, stating 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.

Other markers of strain: Apple's Private Cloud Compute, the server side of Apple Intelligence for complex requests, now also runs on Nvidia GPUs in Google Cloud, first announced at WWDC in June. Then-CEO Tim Cook said in April that Mac mini and Mac Studio supply could take months to catch up after sudden demand for local AI. The July earnings call also cited advanced-node constraints and rising memory costs.

What comes next?

Future work will extend Minions to distributed model training. The authors write that if idle Macs across organizations "or potentially across the global community" can be "securely federated," they could form a shared GPU pool. For now, the Macs serve as overflow AI capacity only inside Apple.

Inside one company, Minions already turned a compute shortage into minutes-long queues; federated globally, it remains a research proposal.