OpenAI is reported to have purchased tens of thousands of Apple Mac minis and Mac Studios for use in reinforcement learning and training computer-use agents, according to U.S. media outlet The Information.
Reinforcement learning is a method of teaching AI through trial and error and rewards rather than by providing every correct answer one by one. A computer-use agent refers to an AI that handles tasks by looking at the screen and using a mouse and keyboard, just like a person.
The challenge is that this type of agent must repeatedly learn by opening document programs, logging into websites, and filling out forms. It requires real, running computers, not GPU servers that only perform calculations. To scale up training, thousands or tens of thousands of computers must be run at once.
Mac minis are only slightly larger than a palm, so they can be stacked tightly on shelves, and they consume little electricity thanks to Apple silicon chips. Under Apple’s licensing policy, macOS can run only on Apple hardware. That means securing actual Macs is the only way to train agents in a Mac environment.
Anthropic is said to be using Mac mini cloud instances provided by AWS instead of buying the machines. Industry observers see this as a split between the two companies’ strategies: direct purchase to gain control with a large upfront outlay, and rental to reduce initial costs.
So far, AI infrastructure investment has focused heavily on Nvidia GPU data centers. If The Information’s report is accurate, major AI companies are now beginning to build training facilities on the scale of tens of thousands of general-purpose computers in addition to GPUs. It is also an example of a consumer product, the Mac mini, being mass-sold as enterprise infrastructure equipment.
Markets are watching how such large-scale equipment spending will affect OpenAI’s valuation and fundraising.
The need for training infrastructure that reproduces real-world usage environments, in addition to expensive GPUs, applies equally to domestic AI companies as they develop agents. By combining small computers with cloud rentals, they can build training facilities at relatively lower cost.