Three months ago I sat down and worked out whether to buy a Mac mini
The AI agent I run at home needs to be on more or less all the time, and a small, silent, low-power box in the corner of the desk seemed exactly right. I didn't buy one in the end, the Windows laptop is already running twenty-four hours a day, and there was no sense buying a second machine to do the same job.
Last week I found out that somebody bought tens of thousands of them.
The Information reported that OpenAI has quietly acquired tens of thousands of Mac minis and Mac Studios over the past few months. Not for foundation models, for training computer-use agents: the kind that move the pointer themselves, open applications, fill in forms, and work through a task step by step. The buying has been blamed for shortages sharp enough that Apple refreshed both machines ahead of schedule on 25 August, introducing the M6, its first two-nanometre chip.
The caveats matter, because this story is already travelling further than the reporting supports:
The figure comes from The Information. OpenAI has not confirmed it, and no precise number has ever been given. Apple has said enterprise demand is strong but has not attributed it to anyone. And these Macs are not training large models, that still happens on Nvidia racks. The Macs do something much narrower.
What stopped me wasn't the number. It was the question of why it had to be a Mac at all.
Apple has said enterprise demand is strong but has not attributed it to anyone.
If you want to teach an agent to use a computer, you have to give it a real computer. The same cursor, the same windows, the same dialog box that appears out of nowhere and blocks what you were doing. That part can't be derived inside a GPU rack. It needs a great many ordinary desktop machines running an ordinary operating system, one attempt at a time.
Which means what is happening inside those tens of thousands of machines isn't AI learning to think.
It's AI sitting at a computer identical to yours, learning the things you do all day.
I work in architecture, and a large share of my day looks like this: open the file, carry the figure into another table, revise a schedule, move something out of one system and into another. None of it is professional judgement. All of it eats the day. And there are now tens of thousands of desktop machines dedicated to learning precisely that layer.
It is also, very often, the layer school measures. How quickly, how neatly, how few mistakes.
I don't have a clean answer to this. The roughest version I've got is: if a task can be learned end-to-end by a desktop machine watching a screen, it probably isn't the thing worth spending a childhood getting good at.
What looks worth practising is everything that never appears on the screen. Judging whether something is worth doing at all. Negotiating. Standing in the middle of a mess and deciding what the first move is.
Nobody is buying tens of thousands of machines to teach that.
How much of your own working day is doing the clicking, and how much is deciding what to click?
