Above $5,000 the question stops being what you can afford and becomes what you actually need. Two paths to running 70B models locally, the engineering that makes them work, and a blunt account of who should not build one.
Most AI build guides start with the graphics card and run out of budget before the power supply. This one starts from what actually runs models reliably for hours - full parts list, what to skip, and the honest ceiling.
Building is cheaper, and that is not always the point. What a prebuilt AI workstation premium actually buys, the three tiers worth considering, and the middle path most comparisons miss.
A machine that runs AI for you is more useful than one you use for AI - and almost nobody builds it. Full parts list specified for efficiency and remote management, plus the stack that makes it worth having.
At $2,500 a local AI machine stops involving compromise. Full parts list built around 32GB of VRAM, what 32B models actually change, and how to know whether you need this tier at all.
