Stop 12 · Model Deployment & MLOpsTopics · 8/8

Cloud ML Platforms (SageMaker, Vertex AI Overview)

In short

Training on big GPUs, serving models to millions and tracking experiments takes a lot of infrastructure. Cloud ML platforms rent you all of it as managed services.

Managed platforms (AWS SageMaker, Google Vertex AI, and similar) that handle training, deployment, and monitoring infrastructure so you don't have to build it all from scratch.

The mascot holds up a key next to giant cooking pots on big burners in a catering kitchen.

Think of it like this

Renting the caterer's kitchen

For a 500-guest wedding, you don't build a kitchen at home. You rent a fully equipped catering kitchen with giant pots, burners and staff, pay for the days you use it, and hand back the keys. Cloud ML platforms rent you the kitchen for models.

Watch it happen

Two popular managed platforms

AWS SageMaker

Part of Amazon Web Services
Notebooks, training jobs, endpoints
Fits teams already on AWS

Google Vertex AI

Part of Google Cloud
Notebooks, training, endpoints, AutoML
Fits teams already on Google Cloud

They cover similar ground. Teams usually pick the one that matches the cloud their company already uses. Azure Machine Learning is the Microsoft equivalent.

The big ideas

What you need to know

Managed notebooks

Hosted Jupyter environments with chosen CPU or GPU sizes, so there's no local setup.

Training jobs

Send a script and data. The platform starts machines, trains, saves the model and shuts down.

Endpoints

Deploy a model behind a managed HTTPS endpoint that can scale with traffic.

Pay for what runs

You're billed while resources run. An idle endpoint or forgotten notebook keeps costing money.

How it works

A typical first project on any platform

1
Upload data
Put your dataset in the platform's object storage bucket.
2
Train
Run a training job with your script, choosing a machine size.
3
Register
Save the resulting model to the platform's model registry.
4
Deploy
Create an endpoint, send a test request, then delete it to stop billing.

Save yourself the pain

Common mistakes

Oops

Leaving a GPU notebook or endpoint running after practice.

Do this instead

Stop notebooks and delete endpoints when done. Set a budget alert before you start.

Oops

Learning every service in the console before shipping anything.

Do this instead

Start with one path: notebook, training job, endpoint. Add pipelines and monitoring later.

Your mission

Explore safely, without surprise bills

~30 min

Quick check

You deployed a model endpoint for practice and finished. What should you do next?

Go deeper

1Explore the free tier or a guided tutorial on one cloud ML platform.
2Learn what these platforms handle for you versus what you're still responsible for.
3Understand the basic cost model so you can reason about when self-hosting is cheaper.