Content channels rarely die from bad content. They die from unit economics: when every video costs time or money, the pipeline stops the first week you are busy or broke. So the design constraint for my YouTube pipeline was absolute: zero marginal cost per video, zero human input.
The cost table that matters
- Voice: self-hosted Kokoro TTS (82M parameters) — replaces a paid voice API. This single decision is most of the "$0/month".
- Compute: Oracle Cloud ARM64 free tier runs the FFmpeg microservice and TTS side by side.
- Thumbnails: Pollinations AI. Scripts: LLM APIs within free/cheap tiers at this volume.
- Upload: YouTube Data API v3 — free, but quota-managed (the real constraint at 8 uploads/day).
Is small TTS good enough?
Kokoro at 82M parameters lands remarkably close to commercial API quality for narration workloads. For short-form content — where the audio sits under music and motion — the difference is inaudible to the audience and irrelevant to the algorithm. Save premium voices for products where voice IS the product.
The pipeline shape
n8n orchestrates: script generation → Kokoro voice synthesis → FFmpeg assembly (Ken Burns, karaoke subtitles, 9:16 reframe) → thumbnail → upload. Six Shorts and two long-form videos land daily, and the queue refills itself. It has run since 2024. The interesting engineering is in the FFmpeg microservice split — covered in its own note.
The best unit economics in content is a pipeline that runs itself for free.from the system this note is based onYouTube Automation Pipeline — full case study →