OpenAI bought tens of thousands of Mac minis to train AI agents

The hottest AI training hardware of the season is the small desktop Apple sells to students.

What The Information reports

OpenAI and other AI labs have bought tens of thousands of Mac minis and Mac Studios, according to The Information. The machines are not a substitute for GPU clusters. They serve a different job entirely: reinforcement learning for computer-use agents, the kind that operate a real desktop the way a person does.

That training loop needs thousands of cheap, independent, real computers running in parallel, each one a desktop session where an agent clicks, types, waits and gets scored on whether it finished the task. A Mac mini is a small, power-efficient, real computer that stacks by the thousand in a rack.

Anthropic reached the same conclusion with a different balance sheet: it rents the same machines through AWS instead of buying them.

The consumer side effect

The most powerful configurations have reportedly been sold out for months amid the ongoing memory shortage. A week earlier, Apple’s new Mac mini lineup arrived with a base price of $899, up from $599 two generations ago, with Apple pointing at memory costs. The queue for the box has an unusual front row.

What it signals

The purchases say something concrete about where the labs think agents are going. Training a model to use a computer like a person requires an enormous fleet of actual computers, and the cheapest suitable fleet turned out to be consumer hardware. The desks the agents are learning on look exactly like the ones they will eventually work at.