The quad 5090 server rig
Four RTX 5090s: 128GB of VRAM and 7,168 GB/s of aggregate bandwidth, for people who know what a power budget is.
This is the build the difficulty rating exists for. Four RTX 5090 cards mean two power supplies planned against a 2,400W draw, a case with the airflow of a wind tunnel, and tensor parallelism configured by hand. Nothing here is plug in and go: you are assembling a small server. What you get for the pain is the fastest token production per dollar that exists in 2026, and image and video generation that a Mac simply cannot touch. It will not run the 400B+ MoE class: 128GB of VRAM is huge but their weight matrices are bigger at usable quantization.
You know what a power budget is because you have exceeded one before. You want maximum tokens per second, image and video generation at production quality, and you accept that setup day is a project with a parts list. This is the home datacenter tier.
first check: next Monday
The parts
What it runs
Nothing local, honestly - but the 2.8T MoE class is still out of reach, and the power bill is a real line item at 2,400W sustained.
Two power supplies, 2,400W at full load: you will want a dedicated circuit and you will hear it. The build assumes you enjoy cable management the way some people enjoy gardening.