Apple pools unused Macs for AI calculations
A newly described internal system uses idle Apple chips as temporary shared computing power, but the research has not yet been peer-reviewed.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
Apple researchers have described a system that combines unused Macs within the company into a virtual GPU cloud. According to the researchers, the system can shorten queues for AI calculations, but it is still a preprint, not an independently tested breakthrough.
The system is called Minions and uses the graphics processing cores of Apple Silicon devices that are not being fully utilised at the time. The aim is to distribute AI inference tasks across existing Macs, so that Apple has to wait less often for centrally available GPU capacity.
In the paper on arXiv, six Apple researchers write that some tasks previously spent more than ten hours in a queue. According to their own measurements, Minions often allowed those tasks to start within minutes. The authors describe the system as a dynamic, distributed infrastructure for large-scale inference.
An independent account by Tom’s Hardware says that, for now, the Macs are used only as internal overflow capacity within Apple. The system is therefore not a public cloud service that consumers or external companies can log into. The researchers do mention as a future possibility that multiple organisations could securely link their unused computing power.
This raises an important technical and organisational question. AI calculations can process data that must not leave a team, department or company. Tom’s Hardware writes that employees are concerned about data being processed on devices outside their own group.
The study comes at a time when computing power is a bottleneck for AI development. Apple writes in its financial reporting that it depends on sufficient computing capacity and is dealing with limited availability, longer delivery times and higher costs. Making use of existing devices may therefore be attractive as a temporary supplement.
For now, Minions focuses on inference: running a trained model. The authors cite extending it to distributed model training as future work. That is technically more complicated, because training processes must continuously exchange data and intermediate results between devices.
The research was posted as an arXiv preprint on 9 October. No peer review or independent replication is yet available. The results therefore show that the system works within the Apple environment described, but do not prove that the same performance can readily be achieved at other companies or on a large public scale.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The description stays close to the preprint and makes clear that this concerns internal use and a non-peer-reviewed study. Claims about future global scale are presented as a future proposal by the authors.
- confirmed Minions pools unused Apple Silicon GPUs on Macs for AI inference. — This appears in the title and summary of the arXiv preprint. source
- confirmed The researchers report that some queues fell from more than ten hours to a few minutes. — Tom’s Hardware’s independent account summarises the paper’s results in this way. source
- confirmed The study has not been peer-reviewed or independently replicated. — arXiv lists the document as a preprint; no peer-review status or replication is mentioned. source
- confirmed For now, the system is used only internally at Apple. — Tom’s Hardware describes it as internal overflow capacity. source