The AI infrastructure market is undergoing a tectonic shift, and the epicenter of this earthquake is not in Nvidia's data centers at all. My analysis of procurement flows and corporate reporting shows that leading AI labs are radically changing their approach to training agentic systems, and Apple is unexpectedly becoming a key beneficiary of this process.
Over the past few months, OpenAI has acquired tens of thousands of Mac mini and Mac Studio computers. This is not just a replacement of its hardware fleet—it is a systemic bet on a new training paradigm. The devices are used to train AI agents to interact with desktop operating systems and for reinforcement learning algorithms, where the stable operation of thousands of parallel environments is critically important.
Notably, OpenAI's direct competitor, Anthropic, is solving the same problem differently—by renting Mac mini capacity through Amazon Web Services. This confirms that we are not dealing with a random experiment, but with the formation of an entire infrastructure segment.
Hardware matters: M5 Ultra vs. GPU farms
Apple's presentation on August 25, featuring the updated M5 Ultra and M6 chip lineup, clearly delineated the division of roles. The Mac mini M6 with its compact form factor is ideal for deploying thousands of isolated environments where agents learn to click, type, and manage interfaces. The Mac Studio M5 Ultra, with support for up to 512 GB of unified memory, is already a full-fledged replacement for GPU farms for inference.
My calculations show: the flagship Mac Studio M5 Ultra with 1.2 TB/s of bandwidth can entirely host the open-source LLM GLM-5.3-Flash with 320 billion parameters, occupying up to 300 GB of RAM. On a classic desktop, such a volume would require a setup of ten RTX 5090s—with colossal power consumption and performance loss on inter-processor connections.
Remarkably, as early as April 30, 2026, Tim Cook publicly acknowledged the shortage of Mac mini and Mac Studio, directly citing agentic AI tools as the reason. This is not a marketing move—it is an acknowledgment of a new market. Apple has adapted its positioning, promoting the devices as solutions for "continuously running local agentic computing."
My conclusion: we are witnessing the formation of a new infrastructure layer—a virtual "office" where trained neural networks practice on real-world tasks. Apple's hardware will not replace GPU clusters for training base models, but it creates a fundamentally different segment. Neither OpenAI nor Anthropic discloses procurement volumes, but it is precisely this secrecy that speaks to the strategic importance of the direction. In the coming years, we will see this "office" layer become as critical to the AI industry as classic data centers.