The fleet, in full
Four machines, sorted by GPU class: the big card thinks, the middle one speaks and paints, the small one watches: and a fourth, with no GPU at all, keeps the code.
The full hardware picture behind Alix: a workstation that thinks (and games), an always-on server that runs the stack and speaks and paints, an older server that watches over everything, and a GPU-less box that holds version control and CI. The rule behind every model swap: a model that fits entirely in the card beats a bigger one that spills onto the CPU.