The hottest piece of AI infrastructure right now isn’t some exotic server rack — it’s Apple’s Mac mini. According to The Information, OpenAI and rival labs have scooped up tens of thousands of Mac minis and Mac Studios to train computer-use agents, the kind of AI that can click through menus, fill out forms, and chain together multi-step tasks without a human babysitting every move. It’s a quiet but significant shift in how frontier labs think about hardware.
The appeal is straightforward. Apple’s unified memory architecture lets the CPU and GPU share the same pool of RAM, which matters enormously when you’re running models that need to hold large contexts in memory. Add solid cooling and a small footprint, and a Mac mini becomes a surprisingly competent workhorse for long-running inference workloads. Anthropic is reportedly renting Mac minis through AWS rather than buying outright, which suggests this isn’t just an OpenAI quirk — it’s becoming an industry pattern.
There’s a catch, and it’s an awkward one for Apple. The most powerful Mac configurations have been sold out for months, throttled by a memory chip shortage. OpenAI reportedly wants more machines than it can get. That supply constraint is telling: Apple built the Mac mini for consumers and prosumers, not as an enterprise AI training box, and its supply chain wasn’t designed for bulk orders from AI labs.
The grassroots scene is moving in the same direction. A tool called Exo lets users cluster multiple Macs over a local network to run large models that no single machine could handle. Peter Voell, formerly of OpenAI’s computing infrastructure team, is building Mount Thor, an Apple-based cloud service that packages this exact approach for others. The market is responding — Apple’s Mac revenue jumped nearly 29 percent to $10.4 billion in the June quarter. Nvidia’s DGX Spark offers a compact alternative with dedicated GPU power and CUDA cores, but it’s a fundamentally different bet: raw compute versus Apple’s memory-centric design. For training agents that need to see and interact with desktop interfaces, the Mac’s architecture may simply be closer to the real-world environment those agents will eventually operate in.
