GPU memory is the ceiling. A language model or image model has to fit in the graphics card's memory (VRAM) to run at full speed. If it doesn't fit, the software either refuses to load it or spills into system RAM and crawls. So I spec an AI build from the card's memory class first and everything else second. Tom's Hardware's Stable Diffusion testing shows it plainly: at higher resolutions, cards short on memory fail to produce an image at all while the same GPU family with more memory finishes the job.
System RAM is the workspace. Model files stage through RAM on the way to the card, datasets load into it for cleanup, and some runtimes split a model between card and RAM when it doesn't fit. AI builds get more RAM than gaming builds, running at rated speed.
Fast NVMe storage. Model files are large, you'll end up with a folder of them, and loading one from a slow drive is a coffee break. The operating system and active models go on fast NVMe, with a second large drive for the archive when the job calls for it.
CPU cores for data prep. The GPU does the inference; the CPU does everything around it, from cleaning tables and resizing images to running the scripts that feed the card. A current high-core-count desktop CPU keeps the card busy.
A power supply with headroom. AI workloads pin the graphics card at full draw for hours, a different load than a game that spikes and settles. I use ATX 3.x supplies with a native 12V-2x6 connector and real margin above the card's draw. Skimp here and you get the shutdowns and melted connectors I see on the repair side.
Cooling for sustained load. A case that breathes, enough intake to feed the card, and a CPU cooler sized for the chip's real sustained output. Quiet and cool are compatible with the right case and fans; see my quiet PC build page.
Gaming PC and AI PC: same parts, different priorities. A gaming build chases frame rates, so the money goes to raw GPU speed. An AI build chases capacity, so the money goes to GPU memory, RAM and storage. An AI workstation games very well; a gaming tower often has too little card memory for the models people want to run.