A quiet but landmark infrastructure revolution is underway in the world of artificial intelligence. This isn't about building another data center with thousands of Nvidia GPUs, but rather a completely different approach to training AI systems. I've been closely following this trend, and the latest data confirms: the computing architecture for agentic AI is changing dramatically.
Over the past few months, OpenAI has acquired tens of thousands of Mac mini and Mac Studio computers. This isn't a spontaneous purchase for experimentation—it's a strategic move to train AI agents to interact with desktop operating systems. The key goal is to teach neural networks to perform real actions: clicking, typing, and managing interfaces. This requires not virtual environments, but full-fledged instances of operating systems, and here the compact Mac mini becomes an ideal tool for deploying thousands of independent working environments.
Why Mac, not GPU clusters?
It's important to understand the difference. Training base models still requires the most powerful Nvidia GPU clusters. However, for the reinforcement learning stage and practicing agentic scenarios, different infrastructure is needed. The Mac Studio M5 Ultra with support for up to 512 GB of unified memory is a true breakthrough. For example, the flagship model with 1.2 TB/s bandwidth can entirely fit the open-source LLM GLM-5.3-Flash with 320 billion parameters, taking up to 300 GB, into memory. For comparison: on a regular desktop, this would require a setup of ten flagship RTX 5090s with enormous power consumption and bandwidth issues.
Interestingly, competitors are taking a different path. Anthropic, OpenAI's main rival, rents Mac mini capacity from Amazon Web Services, preferring not to deal with direct purchases. This confirms that Apple's ecosystem is becoming a critically important element in AI infrastructure, even though Apple itself was long considered a laggard in the AI race.
Acknowledging the shortage and a new strategy
It's telling that on April 30, 2026, Tim Cook, then Apple's CEO, publicly acknowledged the Mac mini and Mac Studio shortage, directly citing agentic AI tools as the reason. The company quickly adapted its marketing, positioning the new devices as solutions for "continuously running local agentic computing." And on August 25, Apple unveiled an updated lineup based on M5 Ultra and M6 chips, which fits perfectly into this new role.
Neither OpenAI nor Anthropic discloses the exact number of devices or purchase amounts. However, my analysis suggests that we are witnessing the formation of a new infrastructure layer—a virtual "office" where trained neural networks practice performing real tasks. This isn't a replacement for GPU clusters, but a logical complement to them. And whoever first builds efficient infrastructure for training agents will gain a colossal competitive advantage in the next cycle of AI development.