Deploying an AI model takes an afternoon. Keeping it affordable is the hard part. Learn how cold starts, GPU utilization, and cost per request decide the bill.