According to The Information, OpenAI has purchased tens of thousands of Apple Mac mini and Mac Studio units in recent months, primarily to support reinforcement learning and to train AI computer-use agents that can autonomously operate computers within real operating system environments. Other AI labs, including Anthropic, have been renting Mac mini instances through Amazon AWS to conduct similar work. This procurement wave has directly increased demand for high-end Mac configurations, with certain models becoming backordered for extended periods.

Details: purpose of procurement and technical rationale

The report notes these Macs are not being used for large-scale foundation-model pretraining (which still mainly relies on Nvidia GPU clusters); rather, they are being used to train and evaluate computer-use agents. Such agents need to practice repeatedly in real OS environments: opening apps, clicking interfaces, filling forms, browsing web pages, and handling multi-step tasks. The Mac’s unified memory architecture (Unified Memory), strong single-device performance, and stable cooling make it well suited for long-running, interactive, and memory-intensive reinforcement-learning workloads.

OpenAI has focused on purchasing Mac mini and Mac Studio units with large unified-memory configurations to locally emulate full computer environments, enabling agents to learn through trial-and-error and policy optimization. Anthropic has been renting Mac instances via AWS to flexibly obtain compute. Apple CEO Tim Cook noted on a recent earnings call that demand from AI and agent tools has exceeded expectations for Mac mini and Mac Studio, and supply tightness for some products could persist for months.

On August 25, 2026, Apple refreshed the Mac mini (adding M6 and other chips) and Mac Studio lines ahead of schedule, a move that has been partly attributed to surging AI-related demand. High-end configurations, especially large-memory variants, had already experienced supply shortages.

Technical background: why Macs?

Traditional AI training depends on GPUs for massively parallel compute, but computer-use agents emphasize:

interaction with real GUI environments
longer context windows and state retention
stable, reproducible single-machine operation
relatively efficient memory access

Apple silicon’s unified memory design allows CPU and GPU to share large, high-speed memory, offering distinct advantages when running medium-sized models, simulating environments, and performing reinforcement learning. At the same time, Macs are compact with good power and thermal characteristics, making them suitable for large-scale deployment as “agent training farms.” Open-source tools (such as Exo) also support clustering multiple Macs to enhance local inference and collaborative capabilities.

Potential impacts: shifts in hardware landscape and AI training paradigms

1. Impact on Apple

Mac unexpectedly benefits from the AI wave. Mac revenue grew about 29% year-over-year in Q2 2026, reaching $10.4 billion. Purchases by AI labs and developers are shifting Mac’s traditional positioning—from primarily creative and consumer use toward becoming an important platform for agent training and local AI deployment. If Apple better serves enterprise and lab customers, it could unlock additional growth opportunities.

2. Impact on AI training infrastructure

Nvidia GPUs remain dominant for large-scale pretraining, but the rise of computer-use agents, reinforcement learning, and local agent execution is creating demand for single machines or small clusters that are high-memory, stable, and easy to deploy. Macs compete differentially with potential small-form-factor Nvidia systems (such as DGX Spark variants), leading to more diverse hardware choices.

3. Impact on AI agent development

Trainin