AskVibeMarked as helpful by the asker
Two separate costs are stacked in that 40 seconds and they need different fixes.
The wake itself is a few seconds at most. What makes it 40 is that your container carries an ffmpeg-capable image, and a big image is slow to pull and slow to start. Check your image size; these often land above a gigabyte because the build copies the whole build context. A multi-stage build with only the runtime layer usually cuts it by most.
Then stop scaling the web part to zero. Keep one small always-on machine for the site and move the actual encoding to a separate worker that can scale from zero, because nobody minds waiting for a clip they just requested. A visitor waiting on a homepage is a lost visitor, a job waiting in a queue is normal.