OpenAI Is Buying Tens of Thousands of Mac Minis for Agents

Somewhere in a data hall, a Mac mini is filling out a spreadsheet no human will ever read. It misclicks, recovers, and tries again, thousands of times a day. According to The Information, OpenAI has bought tens of thousands of Mac minis and Mac Studios to run exactly this kind of drill, at industrial scale, for its computer-use agents.

Why real Macs instead of simulators

Computer-use agents are the class of model that looks at a screen and drives the mouse and keyboard the way a person would. Training that skill runs on reinforcement learning: observe, act, see the result, get rewarded for progress, repeat millions of times. You can do some of that in simulation, but a simulator is too polite. A real operating system throws up notifications at the wrong moment, opens windows late and hangs occasionally, and an agent that will face that chaos in production needs to meet it in training.

There is also a licensing angle to the hardware choice: macOS only runs legitimately on Apple hardware, so an agent that needs to learn the Mac has to learn it on Macs. The mini's size and price make racking thousands of them practical, and Apple silicon gives the labs dense, power-efficient nodes for a workload that needs realistic desktops rather than giant matrix math.

Buying versus renting

The same report says Anthropic is training similar workloads without buying anything, renting Macs through AWS instead. Two procurement strategies, one conclusion: the big labs treat computer-use agents as the next product wave and are committing infrastructure money now. The report even ties Apple's unusually early Mac mini and Mac Studio refresh on August 25, weeks ahead of the usual autumn schedule, partly to AI labs buying faster than expected.

One honest caveat: this all traces back to a single outlet, The Information. Other publications are relaying the same story rather than confirming it independently.

What it signals if you run a business

The products already exist; Anthropic shipped a computer-use capability in its API back in late 2024, and OpenAI's Operator drives a browser. What has been missing is reliability, and hardware purchases at this scale are aimed squarely at that gap. When these agents mature, their first targets will be the screen-bound work every company has: data entry, repetitive form-filling, legacy applications that have an interface but no API. A useful exercise this quarter is simply an inventory. Which of your processes produce hours of screen work each week? That list is where these tools will land first.

Sources: Business Today (reporting The Information)