A quiet but significant transformation is taking place in the world of artificial intelligence. While all attention is focused on the race for Nvidia GPU clusters, leading labs have begun buying up hardware of a completely different kind. This refers to the tens of thousands of Apple Mac mini and Mac Studio computers that OpenAI has acquired over the past few months.

My analysis of the situation shows: this is not just a purchase of office equipment. It is a strategic maneuver aimed at creating a new class of computing infrastructure. According to my data, these machines are being used to train so-called computer-use agents—neural networks designed to learn how to independently interact with desktop operating systems: clicking, typing, and managing the interface. This is a fundamentally different approach to reinforcement learning, where the environment is not a simulation but a real OS.

Why Apple, not Nvidia?

OpenAI's competitor, Anthropic, is solving a similar problem in a different way, renting Mac mini capacity from Amazon Web Services. However, the essence remains the same. The fact is that training agents requires thousands of isolated virtual "workstations." And here, Apple offers a unique combination of compactness and performance.

Of particular interest is the recently unveiled lineup based on the M5 Ultra and M6 chips. I conducted a detailed analysis of the technical specifications and see a clear division of roles:

  • Mac mini M6. Thanks to its compact form factor, these devices are ideal for deploying thousands of independent environments. An AI agent needs a "sandbox" with a real operating system to learn from its mistakes.
  • Mac Studio M5 Ultra. Support for up to 512 GB of unified memory is a key factor. This configuration allows the heaviest language models to be run and hosted locally without resorting to expensive server solutions.

Let me give a clear example: the flagship Mac Studio M5 Ultra with 512 GB of memory and 1.2 TB/s bandwidth can fit the open-source LLM GLM-5.3-Flash with 320 billion parameters (up to 300 GB depending on quantization) entirely in RAM. On a classic desktop, this would require a setup of ten flagship RTX 5090 graphics cards, which would mean enormous energy costs and a loss of bandwidth on interconnects.

It is telling that back in April 2026, Apple's management publicly acknowledged a shortage of Mac mini and Mac Studio, directly linking it to growing demand for agentic AI tools. The company's marketing has also shifted: new devices are positioned as solutions for "always-on local agentic computing."

It is important to understand: Apple hardware will not replace the Nvidia GPU clusters on which base models are trained. It forms a different, no less important infrastructure layer—a virtual "office" where trained neural networks hone their skills on real tasks. Notably, neither OpenAI nor Anthropic discloses the exact number of devices or purchase amounts, which only underscores the strategic importance of this direction.

My conclusion: we are witnessing the birth of a new market—infrastructure for AI "practice." And Apple, without expecting it, may become its key beneficiary, transforming from a consumer brand into an important player in the supply chain for AGI developers.