AI / ML

AI / ML workstation

Local ML lives or dies on VRAM and thermals. Buy the most VRAM you can, then plan cooling and power around it.

Design target

What this setup solves

A local ML box: enough VRAM to run real models, cooling that survives sustained load, and storage that keeps up.

Fit at a glance

Good forLocal LLM · diffusion · fine-tune
PriorityVRAM · thermals · fast storage
LookBoring, quiet, always-on

Gear options

Gear picks · 27 options across 9 categories.

Every row shows a Starter, Sweet-spot, and Long-term choice with rough prices. Skip any row that already works in your room.

A local ML box: enough VRAM to run real models, cooling that survives sustained load, and storage that keeps up.

Starter · lower-cost exampleMiddle · mid-tier exampleHigher tier · higher-cost example

GPU (AI / compute)

VRAM is the constraint that actually blocks work. Buy more VRAM before more cores.

Case + cooling

A GPU-loaded box needs airflow more than aesthetics. Mesh front + push-pull first.

PSU

ATX 3.1 12V-2×6 connector matters for modern GPUs. Get 20% headroom over rated load.

RAM

System RAM has to be ≥ 2× VRAM if you offload layers to CPU.

Storage

Model weights are big and read-heavy. NVMe on a real Gen4 lane matters.

Monitor

One good 27" 4K is enough. Save GPU frames for compute, not pixels.

Chair

You will sit here while jobs run. Do not skimp.

Network / switch

A wired 2.5GbE link to the NAS makes dataset loads sane.

Power + cable management

A UPS + PSU-side surge protection is not optional for a $2k+ GPU.

Price bands are rough, not live quotes. Amazon search links are paid affiliate links; internal links are our own writeups. Verify the exact model, fit, and current price before checkout.

More setups