Researchers Build a GPU Cloud Out of Idle Macs After Inference Jobs Waited More Than Ten Hours in Line

A preprint posted to arXiv on October 9 describes Minions, a system that gathers up the idle GPUs on Apple Silicon Macs and presents them as one virtual GPU cloud for machine learning inference. The Register, which covered the paper, attributes the work to Apple researchers and says Apple uses it to give employees and teams better access to GPU compute. The arXiv listing does not itself state an affiliation, a detail worth noting before anyone gets a mortgage on the news.

The pitch is simple. Somewhere in your office, a number of expensive Macs are, at this moment, doing nothing. Minions proposes to put them to work.

The motivation is a familiar one. According to The Register, before Minions some inference jobs waited more than ten hours in queue because centralized GPU capacity was limited. With the pooled Macs, the paper says, tasks can often start within minutes. The abstract, by authors Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter, cites the “high cost, limited elasticity, and long wait times” of centralized GPU infrastructure, and concludes that idle macOS devices can become a practical distributed inference platform where centralized GPUs are scarce.

Ten hours, it should be said, is a long time to wait for a GPU. It is enough time to eat a meal, attend a meeting, and reconsider your career, in that order. The idea that a perfectly good computer was sitting two desks over, bored, the entire time is the kind of insight that makes a person look up from the queue and see the office.

The architecture has a pleasingly staffed feel. Per The Register, the Minion Server receives requests and breaks them into subtasks. A PostgreSQL metadata store holds task information. The Minion Executor runs inference on a Mac’s GPU, uploads results to S3, and reports back. And a Minion Watchdog monitors for failures, such as an unresponsive Mac, and reassigns the work. Database statistics feed Grafana dashboards for real-time visibility.

A server, an executor and a watchdog. In a normal office, these would be a project manager, a contractor and a person who asks where the contractor is. The system has simply moved the org chart into software and made it fault-tolerant.

Groundy’s analysis of the paper adds the details. Executors send a heartbeat every five minutes, and the watchdog requeues unfinished subtasks when a device goes silent. So the failure mode for a Mac in the Minions pool is that its owner closes the lid, and about five minutes later a database notices and assigns somebody else to finish the job.

That is, to be fair, also how the failure mode works at most companies.

Now, the careful part, and Groundy is careful about it. The paper supports two workloads: embedding generation using Hugging Face transformers and LLM completion using Ollama. It doesn’t claim fine-tuning. It reports no cluster-level throughput or latency figures, no measured recovery from devices dropping in and out, and no cost-per-token comparison. Its evaluation is operational, centered on reducing queue times. Groundy describes it as a single, unreplicated, self-reported preprint from the authors’ own testbed.

There are also some limits that the paper’s design implies. Pooling adds parallel capacity, not faster individual requests, so total throughput is roughly the sum of what each Mac contributes. Models run locally, but task inputs and outputs travel through AWS S3, which matters for anyone with data-residency requirements. And the Neural Engine shouldn’t be counted toward the pool’s capacity, because Ollama and MLX use the GPU.

The Register adds another clarification: Minions is meant to redistribute idle resources on the network, not mainly to run very large models. It separately mentions that recent Mac mini and Mac Studio models support RDMA over Thunderbolt 5 clustering, which allows machines to combine compute and memory for something too big for one device. The article presents this as an unrelated capability.

So this isn’t a rival to a rack of Nvidia hardware, and nobody serious says it is. The Register contrasts Apple’s focus on hardware availability with Nvidia clusters, which are built for throughput, latency and efficient use of GPUs on large models. One approach is a purpose-built cathedral. The other is finding that every desk in the building has a small chapel.

What makes the story endearing is how modest it is. The AI industry has spent three years announcing gigawatt campuses, half-trillion-dollar financing arrangements and orbital data centers. In the same week, a group of researchers noticed that the Mac on the next desk was not busy, and wrote a paper about it.

It is the infrastructure equivalent of discovering a vacant room in your own house. No new land is purchased. No permits are filed. There’s no utility negotiation. There’s just an exhausted, overworked queue, and a lot of computers that will, at a moment’s notice, agree to do some extra work for free, provided they check in every five minutes and the lid stays open.

The Register doesn’t say whether Apple plans to offer Minions externally. For now it is a research paper about getting more from hardware already paid for. In an industry where the default solution to every compute problem is to buy more, the idea of using what you have is almost experimental.

Sources: