After you train an ML model and need to deploy it to production, you have a number of options. If your model runs rarely (1-50x / day), you can set it up as a batch job through various services. In this case it will run to completion, then take itself offline. If your model needs to always be available, via a customer-facing product with constant usage, then you'll deploy it as an endpoint through various services.

Batch models

• AWS Batch. Lets you run a model deployed as a Docker container (eg via ECR) to completion, using price-saving features like spot instances. Much cheaper than Sagemaker, but at cost of spin-up time.

Endpoint models

• AWS SageMaker lets you deploy trained models to a REST endpoint. Also lets you train models & view analytics and various training insights.
• GCP Cloud ML. GCP's equivalent to SageMaker.
• Cortex is similar to SageMaker, with many added benefits. It's free and open source, using your AWS stack to deploy services (like SageMaker) but allowing cost-savings via spot instances, better than SageMaker's 40% EC2 added cost. Soon they'll support scale-to-0 instances, for when your ML server doesn't have traffic; a huge cost saving.
• Other competitors include PaperSpace Gradient, FloydHub, and more.