Adverto is evaluating SageMaker Serverless Inference versus an always-on real-time endpoint for a batch of models with highly spiky traffic (many minutes of zero requests, occasional bursts). Which statement about the Serverless Inference cost model versus an always-on real-time endpoint is most accurate for minimizing cost in this workload?

Serverless Inference is billed based on the number of invocations and the compute duration (GB-seconds), meaning you only pay when your model is actively processing requests. This makes it highly cost-effective for spiky, low-duty-cycle workloads where there are long periods of inactivity. In contrast, an always-on real-time endpoint is billed per instance-hour, regardless of whether it's receiving traffic or sitting idle. For a workload with many minutes of zero requests, the idle time on an always-on endpoint would incur significant unnecessary costs. The other options are incorrect: Serverless Inference does not have a fixed monthly fee or unlimited free invocations. Always-on real-time endpoints do incur instance-hour charges even when idle. While SageMaker Serverless Inference shares some conceptual similarities with Lambda, its pricing model is distinct and does not include a 1M free request tier like Lambda.