OpenAI buys tens of thousands of Macs to train agents

The company is deploying Mac minis and Mac Studios for reinforcement learning while rival Anthropic rents Apple hardware through Amazon Web Services.

OpenAI has purchased tens of thousands of Apple desktop computers to support reinforcement learning and the development of computer-use artificial intelligence agents, according to reporting from The Information.¹

The procurement includes Mac mini and Mac Studio computers, deploying consumer desktop hardware directly into the company’s research infrastructure.¹ While frontier foundation models train primarily on clusters of specialized graphics processors from manufacturers like Nvidia, building software agents capable of operating personal computers requires direct access to operating system environments.

Developing models that interact with graphical user interfaces introduces distinct computational demands. Unlike standard text generation, where models process tokenized language sequences in massive parallel batches, computer-use agents must continuously observe graphical interfaces, interpret visual layout trees, issue precise pointer and keyboard commands, and evaluate the resulting screen state. Training these systems relies heavily on reinforcement learning, where an agent attempts multi-step workflow tasks, receives automated feedback on its success, and iteratively updates its behavioral policy.

Generating the millions of interaction trajectories required for reinforcement learning means running thousands of desktop environments simultaneously. Apple Silicon processors combine central processing units, graphics processing units, and high-bandwidth unified memory on a single die. This architecture allows compact desktop machines like the Mac mini to execute local macOS instances and interface automation with low power consumption and consistent latency.

Competitors are adopting different infrastructure strategies to secure Apple hardware environments. Anthropic, which introduced computer-use capabilities to its Claude model family, rents Mac mini capacity through Amazon Web Services, The Information reported.¹

Amazon Web Services provides macOS environments through specialized EC2 Mac instances, which integrate physical Mac hardware into standard data center server racks. Renting cloud-hosted Mac capacity allows engineering teams to scale macOS instances dynamically without managing physical hardware, power distribution, and thermal cooling in dedicated facilities. Purchasing physical machines directly, however, eliminates recurring cloud rental fees for persistent, high-utilization simulation workloads.

The development of computer-use agents represents an industry-wide effort to expand artificial intelligence from conversational text interfaces to active software operation. Training autonomous agents to book travel, manage enterprise spreadsheets, and write software within standard desktop applications requires continuous exposure to real-world operating systems. Deploying dedicated Apple hardware clusters provides the local execution layer needed to simulate those complex interface interactions at scale.

Amir Efrati of The Information shared the publication’s reporting regarding the hardware deployments on August 31, 2026.¹ Neither OpenAI nor Apple has publicly disclosed the specific terms of the hardware purchases.

Reporting note: this piece draws on reporting published by The Information on August 31, 2026.

Source: The Information via X, August 31, 2026

References

This article is based on 1 source, listed in the order they are cited.

1. 1 H https://x.com/amir announcement · 31 Aug 2026 Report Says OpenAI Is Buying Macs See the source