My local AI stack, and why I run my own infrastructure Draft
The goal
Do serious AI work — models, agents, image generation — on hardware I own, so the work continues no matter what a subscription costs next year.
Hardware and cost
- Mac Studio (M2 Ultra, 64GB unified memory) — model serving over the home network
- 2017 iMac (Omarchy/Arch Linux) — always-on homelab server
Exact specs and cost notes pending fact-check.
Software stack
- Ollama — local model runner
- LM Studio — model serving with an OpenAI-compatible API
- Open WebUI — chat interface for local models
- Stable Diffusion — local image generation
Why self-hosted
The honest reason: it's a hedge. If pricing ever forces me off the paid tiers, the work doesn't stop — the models are already here, on my machines, under my control. Independence isn't a hobby; it's the backup plan.
What broke and how it was fixed
The real war stories go here after the author's fact-check — what failed on the first try, and what the fix was.
Verdict
Pending fact-check: who should copy this setup, and who shouldn't.