Self-hosted AI: the complete guide
Run AI on your own terms: which open models and tools to use, and whether to use your own hardware, a VPS or a rented GPU.
Updated
This guide is being written. Below is the outline and our current price data; test results will be added once our testing is complete.
What “self-hosted AI” means
Running the model, the app around it, or both, on hardware you control instead of a hosted chat service.
The building blocks
- Models: open-weight families and which sizes are practical.
- Runners: Ollama, llama.cpp, vLLM and others.
- Front-ends: chat UIs, coding assistants, image tools.
- Agents and automations: see self-hosted AI agents.
Where to run it
| Option | Good for | Watch out for |
|---|---|---|
| Your own PC or Mac | Learning, privacy, small models | VRAM limits, power use |
| A CPU VPS | Front-ends and agents that call an API | Too slow for local models |
| A rented cloud GPU | Larger models, bursts of heavy use | Hourly cost adds up if left on |
Hosting picks: GPU cloud and VPS & hosting.
Cost: self-hosted vs API
When an API is cheaper and when your own GPU is. Try the AI token cost calculator.
Privacy and security basics
What stays private when you self-host, and how to lock down a public server.
Getting started
A first project for each path: local, VPS and rented GPU.