OpenAI has recently acquired tens of thousands of Mac mini and Mac Studio computers for specific agent-related tasks, according to a report by The Information.

Reports suggest the tech giant is utilizing these devices to train AI agents in autonomous interactions with desktop operating systems (known as computer-use agents) and for reinforcement learning tasks.

In contrast, its main competitor, Anthropic, is tackling similar challenges by renting Mac mini units from Amazon Web Services.

On August 25, just days before the report from The Information, Apple unveiled a refreshed lineup featuring devices powered by M5 Ultra and M6 chips.

Source: Apple.

An analysis of the stated technical specifications indicates a division of roles for computers in AI labs:

  • Mac mini M6. Its compact design makes it ideal for deploying thousands of independent working environments. AI agents require real operating systems to learn clicking, typing, and managing interfaces;
  • Mac Studio M5 Ultra. With support for up to 512 GB of unified memory, it provides the necessary capacity for local execution and hosting of the most demanding language models.

For instance, the flagship Mac Studio M5 Ultra model with 512 GB of memory and a bandwidth of 1.2 TB/s can accommodate the open LLM GLM-5.3-Flash, which boasts 320 billion parameters. Depending on usage goals and quantization options, it may require up to 300 GB of RAM.

On a typical desktop setup, such a volume of video memory would necessitate installing a combination of 10 top-tier RTX 5090 graphics cards with 32 GB each, leading to substantial energy costs and bandwidth loss.

On April 30, 2026, Tim Cook, who was then the CEO of Apple, announced a shortage of Mac mini and Mac Studio units, directly attributing it to demand from AI agent tools. The company has since adjusted its marketing strategy, promoting the new devices as suitable for "always-on local agent computing."

Apple's technology is not intended to replace Nvidia's GPU clusters, which are primarily used for training foundational models. Instead, it could create a new infrastructure layer—a virtual "office" where trained neural networks practice completing real-world tasks.

Neither OpenAI nor Anthropic has disclosed the number of devices, the model breakdown, or the total expenditure on these purchases.