A quiet but significant shift is taking place in the world of artificial intelligence. My analysis of market flows and corporate purchases shows that OpenAI has acquired tens of thousands of Mac mini and Mac Studio computers over recent quarters. This is not just a hardware refresh — it is a strategic foundation for a new generation of agentic systems.

Why do AI labs need Apple hardware?

The key task OpenAI is tackling is training AI agents to interact with desktop operating systems. This concerns computer-use agents that must click, type, and control the interface as naturally as a human would. For this, real, isolated environments with a full OS are needed, not emulators.

Significantly, the main competitor — Anthropic — is moving along a similar path but rents Mac mini capacity from Amazon Web Services. This confirms that the Apple ecosystem is becoming the de facto standard for training agentic models, revealing a new infrastructure layer.

Hardware analysis: division of roles

If we break down the technical specifications of the recently announced devices (Apple unveiled the lineup on M5 Ultra and M6 chips in late August), an engineering division of labor becomes obvious:

  • Mac mini M6. The compact form factor allows deploying thousands of independent "sandboxes" for parallel agent training.
  • Mac Studio M5 Ultra. Support for up to 512 GB of unified memory means the ability to host the heaviest LLMs locally without relying on cloud GPUs.

Here is a concrete calculation: the flagship Mac Studio M5 Ultra with 1.2 TB/s bandwidth can entirely fit the open-source GLM-5.3-Flash model with 320 billion parameters, which takes up to 300 GB of RAM. On a classic desktop, this would require a setup of ten RTX 5090s — with monstrous power consumption and a bus bottleneck.

Apple's new role in the AI race

Notably, back in April 2026, Apple CEO Tim Cook publicly acknowledged the shortage of Mac mini and Mac Studio, directly citing agentic AI tools as the reason. The company's marketing now positions these devices as solutions for "continuously running local agentic computing."

Of course, Apple hardware will not replace Nvidia GPU clusters for training foundation models. But it is forming a new infrastructure layer — a virtual "office" where neural networks undergo practical training. Neither OpenAI nor Anthropic discloses the exact number of devices or purchase amounts, but the scale is evident.

My conclusion: we are witnessing the birth of a market for "inference farms" based on ARM chips. This could become a serious challenge for Nvidia in the medium term, as agentic workloads are growing exponentially, and Apple offers a unique combination of energy efficiency and unified memory.