Apple Silicon as Local AI Infrastructure Play

Apple announced M5 Ultra with up to 512GB unified memory and 1.2TB/s memory bandwidth, alongside M6 in the new Mac mini, and the HN conversation barely touched gaming. The threads on Mac Studio and Mac mini were dominated by one question: when does this become the default local LLM machine? A separate thread asked whether anyone has replaced Claude or GPT with a local model for daily coding, with users reporting real results from Qwen 3 27B and other open-weight models on Apple hardware.

The pattern here: unified memory architecture is what makes Apple Silicon different from a GPU cluster for local inference. You can load a full-fat 70B model into RAM that talks to the CPU at memory bus speeds, not PCIe speeds. The 512GB option arriving in October makes this more interesting for teams who want to run large models on-premise without building a data center.

The counterpoint in the threads is real: token-per-second rates on local hardware are still meaningfully slower than cloud APIs for interactive use, and the tooling to manage model selection is immature. But the cost-per-token math keeps shifting, and several commenters noted that for batch workloads or offline use, local is already competitive.