$AAPL IS BUILDING A LOCAL AI STACK OUTSIDE THE CLOUD Apple is turning the Mac into a local AI workstation since new Mac mini with M6 delivers up to 4x faster AI performance, while the new Mac Studio with M5 Max or M5 Ultra scales to 512GB of unified memory, 1.2TB/s of bandwidth and an 80-core GPU. Apple is also adding Thunderbolt 5 with RDMA so multiple Mac Studios can be linked together for roughly 3x faster distributed inference. The more interesting part is that Apple is competing on memory capacity rather than raw compute. Inference is often memory-bound, so 512GB of unified memory lets a single desktop hold models that would otherwise require multiple data center GPUs, while a four-system M5 Ultra cluster comes to roughly $22,000 and creates a compelling setup for startups or research teams that want private inference without recurring cloud costs or data leaving the building. That is already showing up in the business. Mac revenue grew roughly 29% to $10.35B last quarter as developers increasingly used Mac minis and Studios as local inference machines, and these refreshes push directly into that demand. High-margin memory upgrades also help mix, while the same architecture increasingly looks like a desktop version of the AI server strategy Apple is reportedly building toward longer term. The limitations are still real. Thunderbolt RDMA remains better suited for inference than training, software support across MLX, llama.cpp and other frameworks still needs to mature, and the 512GB configuration does not arrive until late October. Add in sharply higher DRAM prices and worsening supply constraints and the next question becomes whether Apple can build enough of these memory-heavy systems to meet demand without pressuring margins.