While the whole world watches the arms race in Nvidia GPU clusters, a new, quieter, but no less significant infrastructure trend is forming on the periphery. My analysis of market flows and corporate purchases indicates that OpenAI has, over the past few months, carried out a massive acquisition of tens of thousands of Mac mini and Mac Studio computers. This is not just a replacement of the working fleet—it is a strategic move aimed at creating a specialized computing base for training so-called "computer-use agents."

This is about a fundamentally different type of task. Base LLMs are trained on clusters, but to teach a neural network to interact with an operating system interface—clicking, typing, managing windows—a real environment is necessary. This is where Apple hardware becomes indispensable. Compact Mac minis are ideal for deploying thousands of isolated virtual "offices," where agents master desktop environments through trial and error. This is also confirmed by Tim Cook's recent statement about a shortage of Mac mini and Mac Studio, for which he directly cited agentic AI tools as the cause.

Notably, OpenAI's main competitor, Anthropic, is solving this problem differently—by renting Mac mini capacity through Amazon Web Services. This divergence in strategies shows that the market has not yet settled, and each player is seeking its own optimal model of cost and control.

Hardware Analysis: M6 and M5 Ultra

Apple's presentation on August 25, which introduced the M5 Ultra and M6 chips, sheds light on the technical logic behind these purchases. Analysis of the specifications reveals a clear division of roles:

  • Mac mini M6: Compact form factor, ideal for creating large-scale agent training environments.
  • Mac Studio M5 Ultra: Support for up to 512 GB of unified memory—a critical resource for locally hosting heavy language models. For instance, the GLM-5.3-Flash model with 320 billion parameters can take up to 300 GB of RAM, making such a computer a full-fledged replacement for a bundle of ten flagship RTX 5090 GPUs with their colossal power consumption.

This is not an attempt to replace Nvidia in training foundational models. It is the creation of a new infrastructure layer—a "proving ground" for honing practical AI skills. Apple hardware is becoming a virtual office where neural networks learn to work, not just generate text.

My comment: OpenAI's strategy looks forward-thinking. While competitors measure teraflops in data centers, purchasing hardware for edge computing could provide a decisive advantage in developing autonomous agents. This is a signal to the market: the next battle in the AI industry will unfold not in the cloud, but on the desktop. However, it is worth noting that none of the companies disclose exact purchase volumes, leaving room for speculation about the true scale of this "gray" fleet of hardware.