Why SagaMaker Serverless Inference expensive than Lambda?

While Lambda and SageMaker Serverless Endpoints share similarities in both infrastructure and pricing structure, they're ultimately optimized for different use-cases and provide different managed feature sets: AWS Lambda provides a general serverless compute service whereas Amazon SageMaker provides end-to-end capabilities for building, deploying, and managing AI/ML models in particular.

As with other AWS architecture decisions (for e.g. comparing a serverless Lambda + Step Functions architecture to containerized options on Amazon ECS/EKS/Fargate, or fully self-managed instances with Amazon EC2), when making a choice it's important to consider the "Total Cost of Ownership" (TCO) - including the benefits and effort savings that more fully-managed services could provide for your team and use case.

Of course the same choice isn't right for everybody, but some key reasons I've seen customers prefer SageMaker Serverless deployment over plain AWS Lambda include:

1. SageMaker supports many model deployment options from serverless to real-time, asynchronous, or batch inference - usually with no code change required on the model side for porting between them. For example you could migrate to a GPU-capable, Real-Time Inference endpoint later with no code change required in the clients requesting inferences; or run your model against a large batch of data on a cluster of instances using SageMaker Batch Transform.
2. SageMaker provides a range of model governance features and broader monitoring and MLOps capabilities. Organizations working across a portfolio of multiple projects/use-cases (where some might be batch, some need GPU, etc) can significantly improve their agility and governance with centralised tooling like SageMaker Model Registry, serverless multi-step model building pipelines, and the centralised dashboards of models & endpoints in the SageMaker Model Dashboard and the Home > Deployments > Endpoints dashboard in SageMaker Studio.

SageMaker Serverless endpoints don't quite support all the same features as real-time (instance-backed) SageMaker endpoints today, but using them already provides useful integrations that streamline overall ML deployment and management.