I run a 4-node private AI lab out of my home. It handles local LLM inference, email automation, a CMS, and a self-hosted API — with zero ongoing cloud costs. Here's exactly how I built it.
The Stack
Every node has a job:
- Moolah (RTX 5070, 47GB RAM) — AI inference, Stable Diffusion, Whisper transcription
- Neo (Intel i7, 8GB) — Ghost CMS, n8n automation, Supabase
- Uno (Celeron, 8GB) — Listmonk newsletter, Pi-hole DNS, monitoring
- VPS — Cloudflare tunnel gateway, hides home IP
All nodes connect over Tailscale — a WireGuard mesh that makes every machine feel local, regardless of where it is. No port forwarding. No exposed IPs.
The Software (All Free)
- Ollama — runs Llama 3.1, Gemma 3, and Qwen locally. Better than GPT-4o for code on a tuned model.
- Open WebUI — the ChatGPT replacement I actually use daily
- LiteLLM — proxies all models through a single OpenAI-compatible API
- n8n — workflow automation without Zapier's $600/yr bill
- Ghost CMS — this blog (and newsletter) running on a $0 server
- Listmonk — email newsletters, self-hosted, no per-subscriber fees
- Supabase — Postgres + auth + storage, self-hosted on my hardware
The Real Cost
My hardware was bought used. The Celeron box (Uno) was $80. The i7 desktop (Neo) was $120 from eBay. The VPS is $6/mo for the public IP. Everything else is electricity — roughly $15/mo for all four nodes.
Compare that to:
- ChatGPT Plus: $20/mo
- Zapier: $49–$599/mo
- Mailchimp: $50+/mo at 5k subscribers
- Ghost Pro: $36/mo
Payback period: under 3 months.
What I'm Building Here
This blog documents the whole build. Templates, automation workflows, step-by-step guides — all the stuff I wish existed when I started. If you want the full setup blueprint (every config file, every compose stack), subscribe below.