OpenAI has purchased tens of thousands of Mac mini and Mac Studio computers in recent months for specific agent tasks, according to The Information.
The tech giant is reportedly using these devices to train AI agents to interact independently with desktop operating systems (computer-use agents) and for reinforcement learning.
Anthropic, a main competitor, is addressing similar tasks differently by renting Mac mini resources from Amazon Web Services.
On August 25, just days before The Information’s report, Apple introduced an updated line of devices featuring M5 Ultra and M6 chips.
An analysis of the announced technical specifications suggests a division of roles for computers in AI labs:
- Mac mini M6. Its compact form factor makes it suitable for deploying thousands of independent work environments. AI agents require real operating systems to learn how to click, type, and manage interfaces;
- Mac Studio M5 Ultra. With support for up to 512 GB of unified memory, it provides the necessary capacity for locally running and hosting 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 fully accommodate the open LLM GLM-5.3-Flash with 320 billion parameters. Depending on usage goals and quantization options, it may require up to 300 GB of RAM.
On a typical desktop, such a volume of video memory would require a setup of 10 flagship RTX 5090 graphics cards with 32 GB each, leading to enormous energy costs and bandwidth loss.
On April 30, 2026, Tim Cook, then CEO of Apple, reported a shortage of Mac mini and Mac Studio, directly attributing it to agent AI tools. The company continued to adapt its marketing, positioning the new devices as solutions suitable for “always-on local agent computing.”
Apple’s technology is not intended to replace Nvidia’s GPU clusters used for training foundational models. Instead, it may form a new infrastructure layer—a virtual “office” where trained neural networks practice performing real tasks.
Neither OpenAI nor Anthropic disclosed the number of devices, model breakdowns, or purchase amounts.
