Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

OpenAI buys thousands of Mac minis to train AI agents, turning Apple's Macs into training infrastructure

OpenAI has bought tens of thousands of Mac mini and Mac Studio systems for reinforcement learning and computer-use agent training, and is still trying to secure more units, according to a report from The Information. Anthropic is renting Mac minis through AWS for similar work, as Apple's desktop hardware unexpectedly becomes AI training infrastructure.

Published
OpenAI 购入数千台 Mac mini 训练 AI 智能体,苹果 Mac 意外成 AI 训练新主力
Image source: apple.com

OpenAI is turning Apple's desktop line into a training ground for its AI agents. According to a report from The Information cited by India Today, OpenAI has purchased tens of thousands of Mac mini and Mac Studio systems, and is still trying to secure more, while Anthropic is renting Mac minis through Amazon Web Services for similar work.

The move lands as OpenAI pushes into custom chips and memory, and now adds Apple's desktop hardware to its compute mix. The Macs are being used for reinforcement learning, where AI learns through trial and error, and for training computer-use agents that navigate software interfaces, carry out multi-step workflows and interact with computers in a human-like way.

Apple recently refreshed both machines: the Mac mini now comes with the M6 chip, while the Mac Studio offers M5 Max or M5 Ultra chips. The devices have become popular for running AI models locally, but the new reporting shows top AI labs are also using them for training workloads.

The trend reflects developers turning to Macs for specialized AI workloads instead of relying only on large Nvidia GPU clusters. Apple's Unified Memory Architecture and thermal design let the Mac mini and Mac Studio run multiple local sessions and large models on a single system, without the latency and cost overhead of GPU clusters.

The surge is showing up in Apple's financials: Mac revenue rose nearly 29 per cent year on year to $10.3 billion in the latest quarter, making it Apple's fastest-growing hardware line. It also appears to have pushed Apple into an unusually early refresh — the M6 Mac mini and the new Mac Studio chips were announced in late August rather than the typical autumn window.

Even so, the enterprise rush caught Apple off guard. The report said Apple lacked a dedicated engineering team for business customers, staff focused on developer relations, and an enterprise AI strategy. With a global memory shortage under way, many high-end Mac mini and Mac Studio configurations have been out of stock for months.

The report also framed Apple as a growing threat to Nvidia in local AI, with Nvidia responding through the DGX Spark. The next question is whether Apple will formalize an enterprise AI strategy — and whether OpenAI's buying spree keeps high-end Macs scarce.

Why it matters

Apple's Mac line is becoming real AI training infrastructure for OpenAI and Anthropic, challenging Nvidia's dominance in local AI compute while straining Apple's own supply chain amid a global memory shortage.

OpenAIAppleAI AgentMac mini
Back to realtime news

Nearby Updates

All

08/31, 11:01

AQuA:让量化研究 Agent 持续进化,也让回测结果经得起检验

AQuA:让量化研究 Agent 持续进化,也让回测结果经得起检验. 自主研究系统的上限,不只取决于模型有多聪明,也取决于系统能否分辨什么是新证据,什么只是一次偶然的高分。

08/31, 10:59

China's AI Model Call Volume Leads for 18 Consecutive Weeks; Zhipu's NiuLai Tops Chart, DeepSeek-V4-Pro Falls Off

Sina reports that Zhipu's NiuLai model has jumped to the top of China's latest weekly AI large-model API call-volume ranking, while DeepSeek-V4-Pro fell off the list. The report says China's AI large-model call volume has led the field for 18 consecutive weeks, a sign that momentum among the country's top model providers is shifting.

08/31, 10:51

New Entrant StartLux-27B Local Model Reportedly Beats DeepSeek V4 Flash

Sina Finance reports that a new player, the locally deployable StartLux-27B model, has outperformed DeepSeek V4 Flash in a head-to-head comparison. The claim, if verified, would highlight how local deployment models are increasingly competing with cloud APIs on capability and cost.

08/31, 11:51

办公Agent抢位战,谁已获得了下阶段的入场券?|ChatGPT Work|Anthropic|Claude Code|OpenAI|Google 新浪财经

办公Agent抢位战,谁已获得了下阶段的入场券?|ChatGPT Work|Anthropic|Claude Code|OpenAI|Google 新浪财经. 办公Agent抢位战,谁已获得了下阶段的入场券?|ChatGPT Work|Anthropic|Claude Code|OpenAI|Google 新浪财经