Why OpenAI Bought Tens of Thousands of Mac Minis

Why did OpenAI buy tens of thousands of Mac minis and Mac Studios? I did the math: renting one cloud Mac mini on AWS for a month costs about $898 — a brand-new M6 Mac mini costs $899. Plus what each Mac can actually run for local AI, and whether you should buy one now. At the end of August, The Information reported that OpenAI bought tens of thousands of Mac minis and Mac Studios for reinforcement learning of computer-use agents (AI that operates a computer by screenshots and clicks); Anthropic reportedly rents Mac minis through AWS instead. The key reason is Apple's macOS license: macOS may only run on Apple hardware, with at most two extra virtual copies per Mac — so every agent "practice desktop" needs a real Mac. That's also why AWS rents Mac instances for a 24-hour minimum. The craze started with OpenClaw in January; in April Apple stopped taking orders for high-memory Mac minis and Mac Studios, and the cheapest Mac mini went from $599 to $799 to $899 (M6). For local AI: the M6 Mac mini reads memory at 153–170 GB/s, M5 Pro 307, Mac Studio M5 Max up to 614, M5 Ultra ~1.2 TB/s. Speed ≈ bandwidth ÷ model size (real life 60–75% of that). Qwen 3.8 27B (4-bit, ~17 GB) needs 24 GB; 70B runs 4–6 words/s on a 64 GB M5 Pro and 12–18 on a 128 GB M5 Max Mac Studio; GLM-5.3-Flash (320B, 18B active) fits a 256 GB M5 Ultra. Fleet numbers (20,000 Macs) are my illustration — the report only says "tens of thousands". Subscribe for daily breakdowns of local AI with real numbers — and tell me in the comments: would you buy a Mac for AI now, or wait? Thumbnail photos: "Mac mini 2024 M4" by Seasider53 (CC BY 4.0), "Mac Studio (2022) front" by Yasu (CC BY-SA 3.0), Wikimedia Commons, background removed. Chapters: 0:00 Intro 1:02 What OpenAI actually bought 3:00 Why Macs, not Nvidia 5:36 The math: buy vs rent 7:45 How a Mac runs AI 9:40 What each Mac can run 12:35 Should you buy one now? 14:13 The receipt #MacMini #OpenAI #LocalAI