How does scale-to-zero work, and how do I avoid slow cold starts?

Containers scale automatically with traffic and can drop to zero when idle, so you only pay for GPU time when a request comes in. Scaling back up from zero causes a cold start (pulling the image, loading weights). To reduce it: set a minimum of 1 replica to keep one warm, keep your image small, and cache model weights on a Shared Filesystem or the persistent scratch disk that's attached to your deployment instead of downloading them each start. HuggingFace models are cached automatically.