AI machines for local LLMs
A buying reference for prebuilt machines that run large open-weight models locally - unified-memory appliances and large-VRAM desktops - with starting prices and where to buy.
In short: To run 70B-class models locally without building a multi-GPU rig, a unified-memory machine is the simplest path. The NVIDIA DGX Spark (128 GB) and AMD Strix Halo mini-PCs (128 GB, ~96 GB usable) handle 70B at Q4; an Apple Mac Studio (96 GB unified) is the quietest option; and the NVIDIA DGX Station (GB300, 748 GB) targets trillion-parameter work. The table below has current starting prices and vendors - figures are indicative, so verify before buying.
Want a single GPU instead? See Compare GPUs or What fits my GPU? For datacenter cards (H100, B200, MI300X), each GPU page has a "How to buy" section with quote, cloud-rental, and integrator options.