Apple turns idle Macs into a GPU cloud it calls Minions

Minions is a way to borrow the graphics chip (GPU) inside a Mac when nobody is using it. The researchers describe it as "a dynamic and distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices."

In plain words, a lot of Macs sitting on desks become one big pool of AI computers. A team that needs to run a model sends the job to that pool instead of waiting in line for the company's central GPU servers.

The paper is by six Apple researchers: Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter. It appeared on arXiv on October 9, 2026, and on Apple's Machine Learning Research site.

Why it matters: the authors say Minions lets teams "leverage existing hardware without requiring additional GPU procurement." That means no new machines to order.

Running AI models takes GPUs, and big central GPU clusters are expensive to grow and often overbooked. The paper says that when they fill up, teams face long queues, slower experiments and higher costs.

Meanwhile, Apple Silicon Macs have capable GPUs that sit mostly idle during normal use. Email and slide decks barely wake them up.

Tom's Hardware adds some context from Apple's own filings. In its quarterly report in July, Apple said it "depends on access to sufficient computing resources" in a market with "constrained supply, extended lead times, and increasing costs."

The outlet also notes that Apple's Private Cloud Compute, the server side of Apple Intelligence, now also runs on Nvidia GPUs in Google Cloud, which Apple announced at WWDC in June. The big picture: compute is tight everywhere, even at a company that designs its own chips.

The system has four parts, each with a job:

• Minion Server: the front desk. It takes requests through a web page and APIs, checks them, and splits each task into smaller subtasks.
• PostgreSQL metadata store: the notebook. A database with seven tables keeps track of every task, subtask and machine.
• Minion Executor: the worker. It runs on each Mac, grabs the next subtask, downloads the inputs from Amazon S3 storage, runs the model on the Mac's GPU and uploads the result.
• Minion Watchdog: the safety net. It watches busy Macs and rescues work from any that disappear.

Here's the clever bit. There is no central boss handing out work. Each Mac asks for the next piece on its own, which the paper calls a "pull-based execution model."

When two Macs try to grab the same piece, the database lets exactly one win. The loser simply picks another subtask. It works like a deli counter where only one person can hold ticket number 42.

This is the part a normal cloud never has to worry about. Personal computers come and go all day.

The paper lists the usual ways a Mac drops out: it loses the network, goes to sleep, gets powered off, or its owner stops the executor because they need the full machine back. The catch: any of these can happen in the middle of a job.

So every busy Mac sends a heartbeat every 5 minutes. If the heartbeats stop, the Watchdog puts the unfinished subtask back in the queue, and another healthy Mac picks it up. Nobody has to step in by hand.

In real life Picture your work laptop running a chunk of a colleague's AI job while you grab coffee. You come back, open a heavy app, and stop the helper. The job doesn't fail. Another idle Mac down the hall quietly finishes that piece.

For now, Minions only does inference, which means running models that are already trained. It handles two kinds of jobs:

• Embeddings: turning text into lists of numbers that capture meaning, using open-source transformer models from Hugging Face. Search systems use these to find similar documents.
• Text generation: answers from large language models through Ollama, an open-source tool for running models locally. Ollama is installed on each participating Mac.

If a Mac doesn't have the needed model yet, it downloads and keeps a copy. Tom's Hardware reads the paper as each Mac running a whole model on its own slice of work, not one model split across many machines.

That is different from another Mac trick. The Register notes that recent Mac mini and Mac Studio models can be clustered with RDMA over Thunderbolt 5 to load a single model too big for one machine. Minions is not about running bigger models; it is about putting idle machines to work.

According to the paper, yes. Minions supports AI workloads across multiple Apple teams. The authors summarize the change in a simple before-and-after table:

By the numbers: "Prior to its adoption, some inference jobs experienced queue times exceeding ten hours," the paper says. "With Minions, tasks can often begin execution within minutes."

Tom's Hardware adds a fair footnote: "zero" cost leaves out electricity and wear on the hardware.

Teams also get Grafana dashboards showing how many Macs are registered, online, busy or idle, and how many tasks are running, finished or failed.

By default, it's first come, first served. But each Mac's owner can tell their machine to favor a certain team or project, a feature the paper calls dynamic task binding. When no such job is waiting, the Mac helps with everyone else's queue.

In practice, the coordination is pleasantly low-tech. The authors write that team members may "send Slack messages requesting others to temporarily run the Minion Executor and target a specific project or task."

There is one real worry. People hesitate to let their team's data run on machines outside their group, even inside the same company. To ease that, Minions can run as separate team or organization deployments, each with its own database and its own pool of Macs.

Tom's Hardware compares the idea to SETI@home and Folding@home, the volunteer projects that sent small chunks of science work to idle home computers.

The authors say future work will extend Minions to distributed model training, not just inference.

They also think bigger. If idle Macs "across organizations, or potentially across the global community" can be "securely federated," they write, they could form a massive shared GPU pool. For now, though, Minions is an internal Apple tool, with no product or public release announced. You can follow more Apple news on our Apple topic page, or see how rivals pitch local AI on a mini PC.

• No download yet: Minions is an internal system described in a research paper. There is nothing to install on your Mac today.
• Your Mac is more than a typewriter: the paper is a reminder that the GPU in an Apple Silicon Mac can run real AI models, the same kind of local work tools like Ollama already offer.
• Watch the "global pool" idea: if Apple ever lets Macs share spare power beyond one company, privacy and security will be the first questions to ask.