A production record, not a tutorial.
This book documents a GPU inference service running on Azure Container Apps serverless GPU (NVIDIA T4) with min-replicas=0 — where idle time costs nothing. Over 48 days of real operation, GPU charges appeared on only 12. Total cost, including the month spent building it: about $13/month.
It covers the full deployment, and every trap that cost me time getting there: the torch/transformers version conflict, a missing file that passes every startup check and fails only at inference, a startup ordering bug that produces 504s, and — on Windows — an az CLI that crashes mid-build while the build itself keeps running in the cloud.
The example workload is text-to-speech, but almost nothing here is specific to TTS. Image generation, Whisper, a small language model: swap the model, keep the structure.
The tradeoff is a five-minute cold start. This book is honest about when that disqualifies the architecture.
