Apple's AI researchers found a cheap fix for GPU shortages: borrow your coworkers' Macs while they're idle.
A new paper describes 'Minions,' an internal system that splits AI tasks into subtasks and farms them out to idle Apple Silicon Macs on the company network. A Minion Server breaks a job into pieces, Mac-based Minion Executors pull work via REST APIs with no central scheduler, and a Minion Watchdog tracks five-minute heartbeats to reclaim stalled subtasks if a machine sleeps or disconnects. The system handles two job types: Hugging Face embedding models and Ollama-run large language model text generation. Apple says jobs that once sat in GPU queues for over 10 hours now finish in minutes.
This matters because it shows Apple quietly admitting that GPU capacity, not just chip design, is a bottleneck even for the company that makes its own silicon. Apple's July quarterly filing already flagged constrained GPU supply and rising memory costs, and its Private Cloud Compute servers now lean on Nvidia GPUs rented from Google Cloud. Minions is the cheap, internal-facing answer: repurpose the Mac Studios already sitting on desks instead of buying more racks.
The approach itself isn't new. It's SETI@home logic applied to office hardware, and running a full model locally on unified memory is exactly what Mac mini hobbyists already do at home. Apple just gave it a Slack bot and a PostgreSQL database.