Cloud comparison

Which service to use for each part of an ML system, on each cloud, and why it helps your model. Pick a cloud to see its advantages side by side, or open “Advantages” under any service.

Component	AWS	Google Cloud	Microsoft Azure	Open-source / self-hosted
Data storage	Amazon S3 Advantages Practically unlimited, highly durable storage for datasets and model files SageMaker, Athena and Glue read straight from S3, so you train without copying data Versioning keeps old dataset versions; lifecycle rules move cold data to cheaper tiers	Cloud Storage (GCS) Advantages One global namespace with strong consistency Vertex AI and BigQuery read directly from Cloud Storage Autoclass moves rarely used data to cheaper storage automatically	Azure Blob Storage / Data Lake Gen2 Advantages Data Lake Gen2 adds folders and fast analytics on big datasets Connects to Azure ML datastores, Synapse and Databricks Hot, cool and archive tiers to balance speed and cost	Local disk / MinIO (S3-compatible) / PostgreSQL Advantages No cloud bill, and data never leaves your hardware MinIO speaks the S3 API, so code moves to the cloud unchanged later PostgreSQL is a solid home for structured data
Notebooks	SageMaker Studio notebooks Advantages Managed JupyterLab with nothing to install Switch the instance from CPU to GPU when you need more power Built-in access to S3 data, experiments and the model registry	Vertex AI Workbench / Colab Enterprise Advantages Familiar Colab / Jupyter experience on managed machines Query BigQuery with SQL and analyse in Python in one place Idle shutdown keeps costs under control	Azure ML compute instance notebooks Advantages Managed notebooks that also open in VS Code Attach datastores and compute in a few clicks Auto-shutdown schedules stop forgotten machines	JupyterLab or VS Code Advantages Free, and runs on any laptop or server Full control over packages and extensions Works offline
GPU training	SageMaker Training on ml.g5.xlarge (A10G) / ml.p4d for large jobs Advantages Billed only while the training job runs; machines shut down automatically Managed Spot Training can cut training costs dramatically Scales from one GPU to distributed multi-GPU training	Vertex AI custom training with L4 (g2) / A100 (a2) GPUs Advantages Wide choice of accelerators: L4, A100, H100 GPUs and TPUs Spot VMs make long training runs much cheaper Billed only while the job runs	Azure ML compute clusters on NC-series (T4 / A100) GPUs Advantages NC / ND GPUs from T4 up to A100 and H100 Low-priority VMs reduce training costs Compute clusters scale to zero when idle	A local NVIDIA GPU, or rented GPUs (RunPod, Lambda, Vast.ai) Advantages An owned GPU has no hourly cost after purchase GPU rental marketplaces are often cheaper than the big clouds Choose the exact GPU model you need
ML platform & registry	Amazon SageMaker (experiments, model registry) Advantages One place for experiments, the model registry, approvals and deployment Lineage shows which data and code produced each model Access controlled with IAM roles	Vertex AI (experiments, model registry) Advantages Datasets, training, registry and endpoints in one product Experiments and TensorBoard built in AutoML for teams without ML specialists	Azure Machine Learning (MLflow-based registry) Advantages MLflow-native tracking and registry, portable to other platforms Responsible AI dashboard for fairness and explanations Designer and AutoML for low-code model building	MLflow (tracking + model registry) Advantages Open-source, vendor-neutral experiment tracking and registry The same MLflow API works locally and on Databricks or Azure ML Easy to self-host with Docker
Serverless serving	AWS Lambda (container image) or SageMaker Serverless Inference Advantages Scales to zero, so you pay nothing when nobody is using the model Handles spiky traffic automatically No servers to patch or manage	Cloud Run (scales to zero) Advantages Deploy any container and scale to zero A generous free tier for small projects Can attach an L4 GPU when you need one	Azure Functions (container) or Azure Container Apps (scale to zero) Advantages Container Apps scale to zero between requests Pay per request or per second of use Trigger predictions from queues or new files with Functions	Docker container on a small VM (or Fly.io / Render) Advantages Simple and cheap for low or steady traffic A container runs the same anywhere Predictable flat monthly price
GPU serving	SageMaker real-time endpoint on ml.g5 / Inferentia2 Advantages Low-latency real-time predictions with autoscaling Inferentia2 chips can lower the cost per prediction Production variants let you A/B test two models	Vertex AI Endpoint with GPU, or Cloud Run with L4 GPU Advantages Autoscaling endpoints with traffic splitting for safe rollouts Cloud Run GPUs scale to zero, so idle GPUs don't cost money Use prebuilt containers or your own	Azure ML managed online endpoint on GPU SKUs Advantages Managed online endpoints with blue/green deployments Autoscaling and built-in monitoring Secure with keys or Microsoft Entra ID	Triton Inference Server / vLLM / BentoML on a GPU box Advantages Request batching squeezes the most throughput out of a GPU vLLM is a leading engine for serving LLMs fast No per-request fees
Batch predictions	SageMaker Batch Transform, scheduled by EventBridge Advantages Scores millions of rows without keeping a server running Reads input from S3 and writes results back to S3 You pay only for the job's duration	Vertex AI Batch Prediction, scheduled by Cloud Scheduler Advantages Predict over files in Cloud Storage or whole BigQuery tables No always-on endpoint to pay for Results land in BigQuery for analysis and dashboards	Azure ML batch endpoints, scheduled by Azure ML schedules Advantages Process large datasets in parallel on a cluster Clusters shut down when the job finishes Results written to Blob Storage	Cron or Prefect/Airflow job running a Python script Advantages Just a scheduled Python script, easy to understand No vendor lock-in Runs on servers you already have
Pipelines	SageMaker Pipelines or Step Functions Advantages Repeatable train → evaluate → register workflows Conditional steps, e.g. only deploy if accuracy beats a threshold Retraining on a schedule or when new data arrives	Vertex AI Pipelines (Kubeflow) Advantages Managed Kubeflow / TFX pipelines Tracks the lineage of every dataset, model and metric Schedule retraining with Cloud Scheduler	Azure ML pipelines Advantages Reusable components shared across projects Schedules and triggers for automatic retraining Lineage between data, runs and registered models	Prefect, Dagster or Airflow Advantages Open source with large communities Workflows are plain Python Run locally or move to any cloud later
Vector database	Amazon OpenSearch Serverless or RDS PostgreSQL + pgvector Advantages Managed similarity search for RAG chatbots Hybrid keyword + vector search in OpenSearch pgvector keeps embeddings next to your relational data	AlloyDB / Cloud SQL + pgvector or Vertex AI Vector Search Advantages Vertex AI Vector Search handles billions of vectors at low latency AlloyDB / Cloud SQL with pgvector keeps vectors beside app data Managed backups and high availability	Azure AI Search or Azure Database for PostgreSQL + pgvector Advantages Azure AI Search combines keyword, vector and semantic ranking Plugs straight into Azure AI Foundry for RAG PostgreSQL + pgvector for relational and vector data together	Qdrant, Chroma or PostgreSQL + pgvector Advantages Free and open source Chroma is great for prototypes; Qdrant scales to production pgvector reuses an existing PostgreSQL database
LLM access	Amazon Bedrock (Claude and others) Advantages Claude and other models through one API, with no GPUs to manage Prompts and data stay in your AWS account and region Built-in guardrails, knowledge bases and agents	Vertex AI Model Garden (Claude, Gemini and others) Advantages Claude, Gemini and open models in one catalogue Enterprise data governance: your prompts aren't used to train the models Tuning and evaluation tools built in	Azure AI Foundry (Claude, OpenAI and others) Advantages Claude, OpenAI and open models under Azure governance Content safety filters built in Private networking and regional deployments	Ollama or vLLM serving an open-weights model (Llama, Qwen, Mistral) Advantages Prompts and documents never leave your network No per-token charges, only hardware costs Pick, and even fine-tune, any open-weights model
Monitoring	CloudWatch + SageMaker Model Monitor Advantages Alerts on latency, errors and traffic Model Monitor detects data drift against a training baseline Central logs make debugging predictions easier	Cloud Monitoring + Vertex AI Model Monitoring Advantages Model Monitoring alerts on feature skew and drift Dashboards for latency and errors in Cloud Monitoring Export logs to BigQuery to analyse predictions	Azure Monitor / Application Insights + Azure ML model monitoring Advantages Application Insights traces each request end to end Azure ML watches for data drift and prediction quality Alerts can go to email or Teams	Prometheus + Grafana + Evidently AI Advantages Industry-standard open-source monitoring Evidently produces ready-made drift and quality reports Build a dashboard for any metric
Privacy controls	VPC endpoints, KMS encryption, HIPAA-eligible services, keep data in one region Advantages VPC endpoints keep traffic off the public internet Encrypt data and models with your own KMS keys Many HIPAA-eligible services and compliance certifications	VPC Service Controls, CMEK encryption, data residency regions Advantages VPC Service Controls build a perimeter against data leaks Customer-managed encryption keys (CMEK) Choose the region where data is stored	Private endpoints, customer-managed keys, regional data residency Advantages Private endpoints keep traffic on Microsoft's network Customer-managed encryption keys Fine-grained access with Microsoft Entra ID	Everything stays on your hardware; encrypt disks and restrict network access Advantages Full control over where data lives Works in air-gapped environments You set the encryption and access rules (and own the security work)