Google Cloud, AWS, and Azure all offer managed machine learning platforms, hosted foundation models, custom AI chips, and GPU capacity. Their tools overlap a lot, and the names are confusing. The first question to answer is whether you want to build and train your own models or call a hosted model through an API. The answer decides which service you need, more than which cloud you already use.
Two kinds of AI service: build your own, or call a hosted model
Machine learning on a cloud falls into two broad groups:
- Build and host your own models. You prepare data, train a model, and deploy it on managed infrastructure. This suits teams with their own data and models, such as forecasting or fraud detection.
- Call a hosted foundation model. A provider runs large models, such as language or image models, and you send requests through an API. This suits teams that want generative AI features without training models from scratch.
Many products use both. A common pattern calls a hosted model for language tasks and trains a smaller model on company data for a specific prediction.
Managed model platforms
| Area | Google Cloud | AWS | Azure |
|---|---|---|---|
| Build, train, and deploy your own models | Vertex AI | SageMaker AI | Azure Machine Learning |
| Automated model building | Vertex AI AutoML | SageMaker Autopilot | Azure Machine Learning AutoML |
| Model catalog | Vertex AI Model Garden (more than 200 models) | SageMaker JumpStart and Bedrock | Azure AI model catalog in Microsoft Foundry |
| Data warehouse ML | BigQuery ML (train and run models with SQL) | Amazon Redshift ML | Azure Synapse and Fabric integrations |
Azure Machine Learning offers AutoML, a visual designer, MLflow-compatible workspaces, and managed endpoints for serving models. Vertex AI covers the full model lifecycle and includes Model Garden, which lists models from Google, open-source projects, and partners. SageMaker AI covers building, training, and deploying models, and it is the closest AWS match to the other two platforms.
Google’s BigQuery ML is worth a look for data teams. It lets you train and run some models with SQL, close to where the data already lives.
Hosted foundation models
Each provider now offers access to large models through a managed API:
- Google Cloud: Vertex AI gives access to Google’s Gemini models, along with models from partners and open-source projects. Claude models are also available through Vertex AI.
- AWS: Amazon Bedrock gives API access to models from several providers, including Claude, Llama, Mistral, Cohere, AI21, Amazon Nova, DeepSeek, and OpenAI’s gpt-oss. Bedrock became generally available in September 2023.
- Azure: Azure OpenAI Service, generally available since January 2023, provides OpenAI models inside Azure. Microsoft’s broader AI platform launched as Azure AI Foundry in November 2024, and it was renamed Microsoft Foundry in late 2025. Claude models were added to Microsoft Foundry in November 2025.
The model catalogs change often, so check each provider’s current list before choosing a model for a project.
Custom chips and GPUs
All three rent NVIDIA GPUs, including H100, H200, and Blackwell generations. Availability varies by region and by quota, so a large training job may need a quota request before it can run.
Each provider also designs its own accelerators:
- Google TPUs have been used internally since 2015 and were announced in 2016. The Trillium generation (v6e) became generally available in December 2024, and Ironwood (v7) was announced in April 2025. TPUs are available only on Google Cloud.
- AWS Trainium and Inferentia are custom chips for training and inference. Trainium2 became generally available in December 2024.
- Azure Maia 100 is Microsoft’s custom AI accelerator, announced in November 2023.
Custom chips can lower cost for suitable workloads, but they need framework support and testing. Most teams start on GPUs and move to custom chips only after measuring their workload.
Notebooks, pipelines, and frameworks
Each platform includes hosted notebooks for interactive work. SageMaker Studio, Azure Machine Learning studio, and Vertex AI Workbench play that role, and Google also offers Colab Enterprise. For repeatable workflows, each has a pipeline service: SageMaker Pipelines, Azure ML pipelines, and Vertex AI Pipelines, which is based on Kubeflow.
The main frameworks run on all three clouds. TensorFlow was released by Google in 2015, and JAX comes from Google. PyTorch was created at Meta and is now governed by the Linux Foundation. Choosing a cloud rarely means choosing a framework.
Data and ML together
Machine learning depends on data, so the platform that stores and processes your data often shapes the ML decision. BigQuery is Google’s data warehouse, and BigQuery ML lets analysts train models next to their data. Amazon Redshift and Amazon S3 play a similar role on AWS. On Azure, Microsoft Fabric and Azure Synapse fill the same role, with Azure Machine Learning connected to them.
If your data is already in one cloud, starting ML on the same cloud usually saves time and egress costs. Egress charges can grow quickly when training data moves between clouds, as the cloud pricing comparison explains.
Naming traps
The names are the biggest source of confusion. A few distinctions matter:
- Bedrock is not SageMaker AI. Bedrock calls hosted models through an API. SageMaker AI builds, trains, and deploys your own models.
- Claude is not limited to one cloud. It is available through Bedrock, Vertex AI, and Microsoft Foundry.
- Azure OpenAI is not the same as OpenAI’s own API. The two have different terms, regions, and quotas, so check which one a project requires.
- Model and GPU availability depends on region and quota. A model that works in one region may not be available in another.
Which fits which team
- Google Cloud fits teams whose work centers on data analytics, custom model development, and Google’s Gemini or TPU capacity.
- AWS fits teams that want the widest range of managed services around model hosting, and that want to choose among several model providers through one API.
- Azure fits teams already on Microsoft tools that want OpenAI models inside Azure’s security and compliance boundary.
These are general patterns, not rules. Test the specific model and service you plan to use with a realistic dataset before committing.
Certifications for ML roles
AWS offers the Machine Learning Engineer Associate and AI Practitioner certifications. The older ML Specialty exam is being retired, so check its status before planning. Microsoft offers AI-102 (Azure AI Engineer) and DP-100 (Azure Data Scientist). Google offers the Professional Machine Learning Engineer certification. For how these fit into each provider’s wider certification path, see the cloud certifications guide.
For the broader platform context, read the AWS vs Azure vs Google Cloud comparison. The Google Cloud overview covers Vertex AI and BigQuery in more detail, and the Azure overview covers the Azure AI services. For a smaller provider with its own GPU capacity, see the Oracle Cloud Infrastructure explainer.