The artificial intelligence market continues to deliver surprises, and one of the most intriguing trends of recent months is Apple's rapid expansion as a provider of infrastructure for training AI agents. This is not about server racks, but rather the familiar desktop computers Mac mini and Mac Studio, which OpenAI is purchasing in bulk.

According to my data, we are talking about tens of thousands of devices acquired over the past few months. These machines are not used for training base models—that still requires Nvidia GPU clusters. Instead, Mac computers form a new, critically important layer of infrastructure: a virtual "office" where AI agents learn to interact with desktop operating systems. These are so-called computer-use agents, which master skills like clicking, typing text, and managing interfaces through trial and error, i.e., through reinforcement learning.

Notably, OpenAI's main competitor, Anthropic, is solving a similar problem differently—by renting Mac mini capacity from Amazon Web Services. This confirms that the need for "real" operating environments for agents has become an industry standard, and Apple, whether consciously or not, has found itself at the epicenter of this transformation.

Why Mac specifically?

An analysis of the technical specifications of Apple's latest devices, presented in late August, sheds light on this choice. The division of roles is obvious:

  • Mac mini M6. The compact form factor is ideal for deploying thousands of isolated working environments. Agents need full-fledged operating systems to learn from real scenarios, not simulators.
  • Mac Studio M5 Ultra. Support for up to 512 GB of unified memory allows running and hosting the heaviest language models locally. For example, the flagship model with a bandwidth of 1.2 TB/s can fit the open-source LLM GLM-5.3-Flash with 320 billion parameters entirely in RAM, which, depending on quantization, takes up to 300 GB.

For comparison: on a classic desktop, such a volume of video memory would require a setup of ten flagship RTX 5090s with monstrous power consumption and performance loss. This makes the Mac not just convenient, but an economically rational solution for agent training tasks.

Notably, back in April 2026, Tim Cook, who held the position of Apple's CEO, publicly acknowledged the shortage of Mac mini and Mac Studio, directly linking it to the boom in agentic AI tools. The company's marketing quickly adapted: new devices are positioned as solutions for "always-on local agentic computing."

Neither OpenAI nor Anthropic discloses the exact number of purchased devices or contract amounts. However, it is clear that we are witnessing the birth of a new infrastructure market where Apple, thanks to the unique architecture of its chips and unified memory, holds a dominant position. This is not a replacement for GPU farms, but the creation of a fundamentally different layer—an environment for "practicing" already trained models.

My view: this trend is underestimated by the market. If agentic computing becomes mainstream, and everything points to that, Apple could gain a significant source of revenue from corporate purchases, with no direct competitor in this segment. The only question is how soon other manufacturers, including Qualcomm and AMD, can offer comparably efficient solutions.