OpenAI Bought Tens of Thousands of Mac minis for Agents as Unified Memory Wins

APPSO · wechat · 2026-08-31

OpenAI has purchased tens of thousands of Mac minis and Mac Studios for RL training and Computer-use Agents, frequently pressing Apple for delivery, The Information reports; Anthropic is renting Mac compute at scale via AWS.

Why Mac? Apple Silicon's unified memory architecture (UMA) lets CPU, GPU and NeuralEngine share one memory pool, avoiding PCIe shuffling between discrete VRAM and system RAM. With M5 Max at 128GB and M5 Ultra at 512GB, a single Mac can load quantized models far beyond typical 24/32GB consumer GPUs. Mature tooling (MLX, LM Studio, Ollama) turns a big-memory Mac mini into a 24/7 local agent node, and projects like EXO Labs cluster multiple Macs to run models too big for one machine.

Ripple effects: Mac revenue grew 29% YoY to $10.4B last quarter, outpacing iPhone; Nvidia now sees Apple as its top rival in local AI (hence DGX Spark); DRAM/NAND shortages stretched delivery of high-memory configs, pushing some buyers to DGX Spark, while Mac mini pricing in China climbed from ¥4,499 to ¥6,999 in two years. Startups like MountThor sell Apple Silicon AI cloud, and Namespace Labs even racks MacBook Pros to serve macOS CI and coding-agent environments.

Related event: OpenAI Reportedly Buys Tens of Thousands of Mac Minis to Train Agents(4 posts)→

Original post →

More from Companies & People

Companies & People channel →