The cheap mini PC that runs my local AI lab
You do not need a $2,000 GPU rig to run useful local AI. Here is the small, quiet, under-$300 box doing most of the work in my homelab — and exactly what to put in it.
Every “build a local AI rig” post online seems to start at a four-figure GPU. Mine started at a mini PC that cost less than a decent pair of headphones, and honestly it handles the bulk of what I throw at it. Here’s the box and the parts that matter.
The honest expectation-setting
This is not a machine for training models or running the largest LLMs at speed. If that’s your goal, you genuinely do need the GPU. But for the 80% — running small-to-mid local models, orchestration, home automation, a stack of self-hosted services, an always-on agent — a modern low-power mini PC is shockingly capable and sips power doing it.
Mine idles low enough that leaving it on 24/7 doesn’t show up on the power bill, which for an always-on lab matters more than peak benchmarks.
The box
I went with an Intel N100-class mini PC. The N100 is the sweet spot right now: cheap, passively-or-quietly cooled, ~6W TDP, and enough grunt to run a real workload. You can find them well under $200.
What to actually look for: dual NVMe slots if you can get them, 2.5GbE networking (worth it the moment you have a NAS), and at least one SODIMM slot you can upgrade rather than soldered RAM.
The two upgrades that matter
The base config is fine; these two make it genuinely good:
- RAM — go to 16GB or 32GB. This is the single biggest lever for running more services and bigger local models at once. Don’t skimp here.
- A real NVMe SSD. The bundled drive is usually slow and small. A decent 1TB NVMe makes the whole thing feel like a different machine.
That’s it. Box, RAM, SSD — comfortably under $300 all-in if you shop the N100 deals.
What it’s running
Right now this little box handles a local model for quick automation tasks, my Telegram ops agent, a handful of containers, and assorted glue — quietly, in the corner, on barely any power. The heavy GPU jobs get sent elsewhere; everything else lives here.
Should you buy one?
If you’ve been telling yourself local AI needs a loud, expensive tower — this is permission to start small. Get the cheap box, max the RAM, drop in a real SSD, and see how far it gets you. For most homelab use, that’s further than you’d think.
Prices and availability change constantly — always check current listings before buying.