Apple reportedly plans M8 Ultra AI inference servers for enterprises, backed by new CEO Ternus

APPSO · wechat · 2026-09-28

Apple is reportedly discussing a return to the enterprise server market (abandoned since 2011's Xserve) with machines packing two to four M8 Ultra chips, possibly linked via NVIDIA's NVLink Fusion, to run pre-trained models for inference — not training. New CEO John Ternus reportedly backs the project.
The piece frames this via the local-AI boom: OpenClaw-style always-on agents made high-memory Mac minis scarce, while 175,000 misconfigured Ollama instances exposed on the public internet show the risks of DIY local AI. Enterprises want private inference without building NVIDIA-based stacks from scratch.
Apple's unified memory (up to 512GB, 1.2TB/s on M5 Ultra) makes it well suited to turnkey inference boxes; it sells machines, not cloud compute.
Ternus, Apple's first hardware-background CEO in 30 years, is doubling down on deciding where compute happens: on-device first, then Mac, then private servers, with cloud as overflow — while Apple leases Gemini for Siri at $1B/year.

More from Companies & People