OpenAI Is Buying Tens of Thousands of Macs to Train Its Agents

OpenAI has spent the past several months quietly buying tens of thousands of Apple Mac mini and Mac Studio computers, according to a report from The Information relayed by Digitimes. The machines are not there to serve as glossy office hardware. They are being wired into racks and used to train reinforcement-learning systems and the computer-use agents that are becoming OpenAI's next product frontier.

This is a different kind of buying than the headline-grabbing GPU orders. The graphics clusters OpenAI rents for pre-training a large model do the heavy mathematical lifting. The Macs do something more specific: they give an AI agent a real computer to practise on. To train a model that can open an app, click through a workflow and finish a multi-step task, you need a lot of actual machines running actual software, and Apple's compact desktops turn out to be a cheap, power-efficient way to supply them.

An accidental hardware win for Apple

Apple did not court this. The company has no large enterprise AI sales operation, and by most accounts it was caught out by the demand. The buying is widely blamed for the shortages that pushed Apple to refresh both the Mac mini and the Mac Studio on 25 August, weeks ahead of the autumn timing it usually keeps. A firm that spent years watching Nvidia define the AI hardware market has found a slice of that market landing on its own product line, unbidden.

The appeal is the silicon. Apple's chips put processor, graphics and memory on one package with a shared pool of unified memory, which suits the kind of steady, mid-sized inference work that running thousands of agent sessions demands. Racks of Mac minis draw less power than comparable server gear and slot neatly into a data hall. For a workload defined by breadth rather than raw peak throughput, they fit.

OpenAI is not alone

Anthropic is reported to be renting Mac mini capacity through Amazon Web Services for a similar purpose, which suggests the tactic reflects where agent development is heading rather than one company's quirk. If the near future of AI is software that drives other software, then training that software cheaply means owning, or borrowing, a great many ordinary computers.

A few caveats are worth keeping. Neither OpenAI nor Apple has confirmed a unit count, a price or a contract, so the "tens of thousands" figure comes from reporting rather than a filing. What is clear is the direction. OpenAI has been diversifying its compute base for a while, from renting GPUs to designing its own inference chip, and even chipmakers are now reworking their processors around agent workloads. Buying a warehouse of Mac minis is the same story told in the least glamorous hardware imaginable.