A Comparative Analysis of Amazon SageMaker AI and Google Agent Platform Notebooks

Amazon SageMaker AI and Google Cloud both offer fully managed Jupyter notebooks for designing and developing machine learning and deep learning models on managed cloud compute. The two offerings have a lot in common, and they also have key differences. This post compares them side by side.

Both products have changed since 2018. Google deprecated Cloud Datalab and recommends Vertex AI Workbench in its place. In April 2026, Google made Vertex AI part of Gemini Enterprise Agent Platform. Vertex AI Workbench is now Agent Platform Workbench, and Colab Enterprise is part of the same platform. On AWS, the original Amazon SageMaker service was renamed Amazon SageMaker AI in December 2024. AWS now uses the name Amazon SageMaker for a wider data, analytics and AI platform built around SageMaker Unified Studio. This post covers SageMaker AI and its notebook environment, Amazon SageMaker Studio.

The table below maps the services we compare.

Role	AWS	Google Cloud
Machine learning platform	Amazon SageMaker AI	Gemini Enterprise Agent Platform (formerly Vertex AI)
Managed JupyterLab environment	JupyterLab spaces in SageMaker Studio	Agent Platform Workbench instances
Other notebook options	SageMaker notebook instances	Colab Enterprise
Managed training	SageMaker AI training jobs	Agent Platform Serverless Training
Hyperparameter tuning	Automatic model tuning (AMT)	Hyperparameter tuning jobs
Online model serving	Real-time inference endpoints	Agent Platform Endpoints

Notebook Deployment

Amazon SageMaker AI and Google Cloud both offer simple notebook setup. In SageMaker AI, you first set up a domain. The Set up for single user (Quick setup) option in the SageMaker AI console creates the domain, a user profile, an IAM execution role and an Amazon S3 bucket for you. You then open SageMaker Studio, choose JupyterLab and create a JupyterLab space. You can pick the Amazon EC2 instance type that runs the space. SageMaker AI also still offers standalone notebook instances, which run the Jupyter Notebook application on a single ML compute instance.

On Google Cloud, you create an Agent Platform Workbench instance from the Instances page in the Google Cloud console or with the Google Cloud CLI. This example from the gcloud reference creates an instance from the latest Workbench image:

bashCopy
gcloud workbench instances create example-instance \
    --vm-image-project=cloud-notebooks-managed \
    --vm-image-family=workbench-instances \
    --machine-type=n1-standard-4 \
    --location=us-central1-b

When the instance is ready, the console shows an Open JupyterLab link. Colab Enterprise needs even less setup. You create a notebook, and Colab Enterprise provisions a runtime for it when you need one.

Pre-loaded Notebook Libraries

JupyterLab in SageMaker Studio uses the SageMaker Distribution image by default. The image includes PyTorch, TensorFlow, Keras, NumPy, pandas and scikit-learn.

Agent Platform Workbench instances come with a set of pre-installed deep learning libraries, including support for TensorFlow and PyTorch. Their kernels are based on conda environments, and you can add your own.

On both platforms, you can install more packages from a notebook cell with the %pip magic command. For example:

pythonCopy
%pip install --upgrade xgboost

Managed Compute Infrastructure

A JupyterLab space in SageMaker Studio runs on a single Amazon EC2 instance with a single Amazon EBS volume for storage. You can switch to a faster instance or increase the EBS volume size when you need to. SageMaker AI can shut down idle JupyterLab and Code Editor applications to control costs. An administrator sets this idle shutdown at the domain or user profile level. SageMaker AI sends metrics to Amazon CloudWatch and sends logs from training jobs, endpoints and notebook instances to Amazon CloudWatch Logs. Studio workloads are subject to service quotas, which you can review in the Service Quotas console.

An Agent Platform Workbench instance runs on a virtual machine that you size when you create it. You can choose a CPU-only or GPU-enabled configuration and edit the machine later. By default, Workbench instances shut down after a period of inactivity. You can change the time limit or turn this off. Colab Enterprise starts runtimes when you need them and shuts them down when you no longer need them. You can create runtime templates to choose the machine type, accelerators and disk size.

Optimized Machine Learning Algorithms

Amazon SageMaker AI provides built-in algorithms that you train by pointing them at your data. They cover supervised learning (for example XGBoost, Linear Learner, LightGBM, CatBoost and AutoGluon-Tabular), unsupervised learning (for example k-means clustering and principal component analysis) and deep learning tasks such as forecasting with DeepAR and machine translation with Sequence-to-Sequence. SageMaker JumpStart adds pre-trained models and solution templates.

On Google Cloud, the closest options are Agent Platform AutoML, which trains models on your own data, and Model Garden, which offers pre-trained models. Agent Platform training also has integrated support for PyTorch, TensorFlow, scikit-learn and XGBoost code that you write yourself.

Distributed Training

Amazon SageMaker AI runs training jobs on instances that it provisions for the job. The instances are released when training completes. A job can run on one instance or on many, and SageMaker AI provides distributed training libraries for large PyTorch workloads.

Code that you run in an Agent Platform Workbench notebook runs on that single instance. To train across several machines, you submit your code to Agent Platform Serverless Training, a managed service that runs training applications on Google Cloud infrastructure. It supports multi-node distributed training on GPUs. Workbench also has an executor that runs a notebook file once or on a schedule as an Agent Platform training job. SageMaker Studio offers a similar feature that schedules a notebook to run as a job.

Automated Hyper-parameter Tuning

Amazon SageMaker AI automatic model tuning (AMT) runs many training jobs over the hyperparameter ranges you specify. It then picks the hyperparameter values that give the best model, measured by a metric that you choose. You can use it with built-in algorithms, custom algorithms and the SageMaker AI framework containers.

Agent Platform offers the same capability through hyperparameter tuning jobs. A tuning job runs several trials of your training application with different hyperparameter values and finds the best values within the ranges you set.

Model Deployment

Trained models can be deployed from Amazon SageMaker AI to real-time inference endpoints. The endpoints are fully managed and support automatic scaling. SageMaker AI also supports other hosting options, such as batch transform.

On Google Cloud, you import a trained model into Agent Platform Model Registry and deploy it to an Agent Platform endpoint for low-latency online inference. Agent Platform also supports batch inference, which sends a request directly to a model that is not deployed to an endpoint.

Pricing

Both platforms bill for the compute and storage you use. Idle shutdown on both platforms helps you avoid paying for notebooks that you are not using. The details depend on your instance types, Region and workload. See the Amazon SageMaker AI pricing page and the Gemini Enterprise Agent Platform pricing page for current rates.

Conclusion

Amazon SageMaker AI is a complete machine learning service that links building, training, tuning and deploying models in one place, with SageMaker Studio as its main workspace. In 2018, Google Cloud Datalab was a standalone notebook that relied on other services such as Cloud ML Engine for large-scale training and deployment. Today, Agent Platform Workbench and Colab Enterprise are part of Gemini Enterprise Agent Platform, so Google's notebooks also connect directly to managed training, hyperparameter tuning and model deployment.