OpenAI Purchases Macs for Agents That Click Like Humans
Frontier Labs is practicing controlling real desktops-though the scope remains unconfirmed, the AWS host list is publicly exposed.
The Information reports that tens of thousands of Mac mini and Mac Studio computers are being used by computer-use agents to operate a system like a human would. OpenAI is reportedly purchasing these machines, while Anthropic is renting them via AWS. Neither company has publicly confirmed the quantities, dates, or costs involved.
Key Takeaways
- OpenAI is said to be buying Macs, while Anthropic is renting them through AWS. The August 30 report claims OpenAI purchased a fleet of Macs for reinforcement learning-fine-tuning based on success and failure in real-time operation-and training computer-use agents. Neither company has confirmed the scale of these operations.
- The agent requires a full operating system. Computer-use agents interact with interfaces and inputs like a human, necessitating thousands of independent macOS instances. A single GPU cluster cannot replicate this setup. AWS lists EC2 Mac hosts up to the M4 Max model.
- Apple’s revenue reflects the demand. In the quarter ending June 27, Mac sales surged nearly 29% year-over-year-but this doesn’t prove a direct link to the labs’ purchases.
Related:AWS AgentCore Brings Memory to GovCloud / Shopify Threatens Claude Code Over a Single File
What The Information Reports
The Information broke the story on August 30, with Decoder and other outlets amplifying it the following day. The core claims are concise: OpenAI reportedly purchased tens of thousands of Mac mini and Mac Studio units over recent months. These machines are used for reinforcement learning and training computer-use agents-models that interact with computers like humans, performing tasks such as clicking, typing, opening files, and executing multi-step workflows. Anthropic, meanwhile, is renting Mac minis via Amazon Web Services for similar purposes. Neither OpenAI nor Anthropic has confirmed the quantities, timelines, or costs involved.
Both companies are reportedly using the same Mac lineup Apple sells as desktops-one buying, the other renting. The effort behind computer-use agents isn’t separately itemized in either company’s financials.
Secondary reports include details that likely originate from The Information’s source. High-end configurations have been scarce for months, with memory chips in short supply. Studios with additional RAM reportedly face delivery delays of two weeks to nearly two months. Apple, OpenAI, and Anthropic have not commented on these claims.
The Agent Needs a Full Operating System
Pretraining large models happens on GPU clusters, but computer-use agents eventually operate within a window. They must open Safari, read a form, locate a version number in Terminal, and set a reminder-tasks requiring a real macOS instance, often with a screen, mouse, and keyboard input. Thousands of independent Mac minis are the logical solution.
Apple’s Unified Memory aligns with this need: CPU, GPU, and Neural Engine share a memory pool. Running thousands of parallel macOS instances demands dedicated RAM and machines-not the same GPU clusters used for pretraining on Nvidia H100 systems. The report explains why Macs make technical sense, though it doesn’t prove the claim. It shows why the assertion isn’t absurd.
What can be independently verified is AWS’s rental listings. The service offers dedicated EC2 Mac hosts, including Intel Mac mini, M1, M2, M2 Pro, M1 Ultra, M3 Ultra, M4, M4 Pro, and M4 Max models. As of August 31, the M4 Max supports up to 128GB of Unified Memory. AWS describes these instances as ideal for building and testing iOS and macOS applications. While the report claims Anthropic rents the same hosts for computer-use tasks, the documentation doesn’t mention M5 or M6 Mac instances-the newer models refreshed on August 25 aren’t listed.
Apple tracks revenue, not agent purchases
On July 30, Apple released its figures for the quarter ending June 27. Mac revenue rose nearly 29% year-over-year. Reuters quoted Tim Cook citing a shortage of advanced-node chips as a key constraint for the Mac lineup-hard evidence of the product line’s challenges rather than an inventory list from the labs.
The report suggests purchases were made in recent months, but the quarter ended June 27. Apple refreshed the Mac mini and Mac Studio on August 25, earlier than the usual fall cycle, with preorders open and deliveries starting September 22. The timing aligns with the report, though no direct proof exists.
OpenAI and Anthropic are reportedly testing the same agent class-once on purchased Macs, once on rented ones. Pretraining still relies on GPU clusters, while fine-tuning happens on desktops. OpenAI is said to buy the desktops and account for them on its balance sheet, while Anthropic would lease them via AWS and cancel as needed. Apple sells the machines; the labs decide whether to own or rent them. Unclear is whether OpenAI, Anthropic, or Apple will disclose volumes or timelines.
Frequently Asked Questions
Have OpenAI or Anthropic confirmed Mac purchases?
No. *The Information* cites unnamed OpenAI sources, but neither OpenAI nor Anthropic has publicly confirmed quantities, timelines, or spending. This remains an unverified report.
What is a computer-use agent?
A model that interacts with a computer like a human-opening windows, clicking, typing, reading files, and executing multi-step tasks. Each parallel session requires an operating system, often macOS. A GPU job alone isn’t enough.
Which Mac instances does AWS list?
AWS’s public list from August 31, 2026, includes Intel Mac mini, M1, M2, M2 Pro, M1 Ultra, M3 Ultra, M4, M4 Pro, and M4 Max. AWS describes them as build and test hosts for Apple software. No M5 or M6 Mac instances appear.
Did OpenAI drive Apple’s Mac quarter?
The quarter ending June 27 showed Mac revenue up nearly 29% year-over-year. The lab report mentions purchases in recent months, but no confirmed source links them causally.
What does Apple say about these purchases?
Nothing. Apple’s June 27 earnings report highlights a 29% Mac revenue increase and cites chip supply constraints-but no customer names. Whether OpenAI or Anthropic contributed remains unknown.
Editor’s Picks
- ChatGPT Now Rarely Cites Reddit-Data Shows Why
- Kubeflow: CNCF Declares AI Platform Mature
- Claude Breaks Math Barrier-But No Proof in Sight
More from the network
More from the MBF Media Network
MyBusinessFutureAgents Cost Providers Five Times More Than ChatDigital ChiefsWhy AI Fails at Master DataSecurityTodayExploitGym: OpenAI Agents Built a Secret BackchannelImage source: AI-generated (September 2026)
Translated from the German original using artificial intelligence. The German version is authoritative.
Also available in

