I run a production AI app for €60 a month. Here is the actual bill
Описание и применение
В небольшом AI‑приложении для ухода за питомцами инференс стоит лишь несколько центов в месяц, а основная часть расходов приходится на хранение и базу данных. Умные решения: один вызов модели для аудио, кэширование рекомендаций, ежедневные счётчики пользователей и простая VPS‑инфраструктура позволяют держать ежемесячный счёт в пределах €60–200.
Задача
How to keep a production AI pet‑health app running on a single VPS for €60–200/month while serving AI features
Результат
Inference costs only a few cents per heavy user per month; storage and database are the main cost drivers. A simple VPS stack with caching and per‑user limits keeps the bill between €60–200/month for a small user base.
Рабочий процесс
- 1. Send audio once to a multimodal model for transcription and structured output
- 2. Apply rule prefilter, vector rank, then one model pass for recommendations
- 3. Cache recommendations for seven days
- 4. Use per‑user daily counters instead of global rate limits
- 5. Run NestJS, Postgres, pm2, nginx, Cloudflare on one VPS
- 6. Prune original images, re‑encode to WebP
- 7. Store confidence scores and require user confirmation before writing to record
Доказательства
A very heavy user — thirty assistant messages a day plus a daily health check — costs a few cents a month in inference. Their storage and database footprint costs more than their AI does.
Inference is a tiny fraction of the bill; storage, database, and caching dominate.
We spent months worrying about the wrong line item.
- Дата записи: 2026-09-22T20:00:31.761Z