OpenAI bought the Mac minis behind the shortage

OpenAI bought tens of thousands of M5 Pro and M6 Mac minis for computer-use agent training, per Cult of Mac's report. It bought M5 Max and Ultra Mac Studios too, the same hardware independent developers order for local inference, and Apple's supply couldn't keep up.

Anthropic reached the same conclusion about the training requirement and rented Mac mini capacity through AWS rather than buying.

The constraint is simple. An agent that operates macOS needs macOS, and macOS runs on Apple hardware, full stop. No virtualization trick gets you a fleet. The training base for this kind of agent is a consumer desktop, bought at retail and racked in a data center.

OpenAI's purchase means capex, physical logistics, and a depreciating asset tied to Apple's release calendar. Anthropic's rental means a per-hour cost that never turns off. That cost also depends on AWS's EC2 Mac capacity, which runs into Apple's own 24-hour minimum dedicated-host rule.

Computer-use agents are hardware-bound in a way language agents aren't. That means the cost curve for this capability will fall slower than the rest of AI. Price a desktop-automation product against Apple's supply chain, not against token costs.

A Mac Studio order that slipped this quarter has a reason: independent developers are now bidding against two frontier labs for the same SKUs.