For regulated data, the LLM question isn't cost — it's whether the tokens are allowed to leave. GDPR, the EU AI Act, and why on-prem inference jumped from 12% to 55%.
Open-weight models pass real production acceptance criteria in 2026. What Llama, Mistral and Qwen are each good at, and what 'good enough' actually means.
Self-hosting an LLM breaks even at roughly 2–5M tokens a day against frontier APIs — and 50M+ against open-model hosts. The honest maths, with charts.
The GPU is the cheap part. Engineering hours, the 2.5–3× infrastructure multiplier, and the honest cases where you should stay on the API — with a decision flowchart.
Your Lovable or Bolt app works, until real users arrive. What actually breaks in vibe-coded apps, what a rescue involves, and what it should cost.
When the LLM API bill becomes a board question, the answer isn't always self-hosting. How to think about renting vs owning inference — with the spend curves.