Self-hosted AI: what it actually costs and when it makes sense

Every week someone asks us to set up "our own ChatGPT". Sometimes it is exactly right. Sometimes it costs three times more than the subscription it replaces and does less. The difference is knowing which situation you are in.

When self-hosting genuinely wins

When the data cannot leave. Legal firms, clinics, anything under contractual confidentiality — if sending client data to a third party is not allowed, the decision is already made.

When volume is high and steady. Per-token pricing is cheap until it is not. Past a certain constant load, a machine you rent monthly becomes cheaper and, more importantly, predictable.

When you need it to know your documents. A model connected to your own knowledge base — contracts, manuals, history — answers questions no public model can.

When it does not

If usage is occasional and light, the subscription wins by a wide margin. And the frontier models are still better than anything you can run on a reasonable server. If you need maximum reasoning quality, self-hosting is a downgrade.

# Una instalacion basica de Open WebUI con Ollama
docker run -d --name openwebui \
  -p 3000:8080 \
  -v openwebui:/app/backend/data \
  --add-host=host.docker.internal:host-gateway \
  ghcr.io/open-webui/open-webui:main

The honest summary

Self-hosted AI buys you control, privacy and a predictable bill. It costs you some quality and adds a system to maintain. If those trade-offs match your situation, it is excellent. If they do not, we will tell you so.