AWS vs. GCP vs. Azure: Core Differences
How AWS, Google Cloud and Azure differ underneath: account structure, identity, networking, zones and pricing, plus a service-by-service map.
On the surface, AWS, Google Cloud and Azure sell the same things: virtual machines, object storage, managed databases, Kubernetes, serverless functions and AI services. A service map that pairs EC2 with Compute Engine and Azure Virtual Machines is useful, and there is one at the end of this article. But the differences that trip people up when they move between clouds are rarely about which service exists. They are about how each provider organizes accounts, grants permissions, scopes networks and charges for usage.
This article covers those structural differences first, because they shape every design decision, and then maps the common services across the three providers. If you are deciding which provider to use rather than learning how they differ, read Choosing a Cloud Provider: One Cloud or Several alongside this one.
How each cloud organizes resources
Every resource you create lives inside a container that determines who can access it, where it is billed and which policies apply. The containers have different names and, more importantly, different boundaries.
flowchart TB
subgraph AZ["Azure"]
Z0["Entra ID tenant"] --> Z1["Management groups"]
Z1 --> Z2["Subscriptions"]
Z2 --> Z3["Resource groups"]
Z3 --> Z4["Resources"]
end
subgraph GCP["Google Cloud"]
G1["Organization"] --> G2["Folders"]
G2 --> G3["Projects"]
G3 --> G4["Resources"]
end
subgraph AWS
A1["Organization"] --> A2["Organizational units"]
A2 --> A3["Accounts"]
A3 --> A4["Resources"]
end
| AWS | Google Cloud | Azure | |
|---|---|---|---|
| Main isolation boundary | Account | Project | Subscription |
| Grouping above it | Organizational units in AWS Organizations | Folders | Management groups |
| Grouping below it | None built in; tags | None built in; labels | Resource groups |
| Where billing attaches | Each account, consolidated in the organization | A billing account linked to each project | Each subscription |
| Guardrails across the hierarchy | Service control policies | Organization policies | Azure Policy |
The practical consequences:
- On AWS, the account is the unit of isolation. Teams typically get separate accounts for each environment and workload, and cross-account access is done with roles. Creating many accounts is normal and expected.
- On Google Cloud, the project plays that role, and projects are lighter to create. Folders let you apply permissions and policies to many projects at once, and those settings are inherited downward.
- On Azure, there are two levels. Subscriptions are the billing and quota boundary; resource groups inside them hold resources that share a lifecycle. Deleting a resource group deletes everything in it, which makes it a convenient unit for test environments.
How identity and permissions work
All three use role-based access, but the models differ enough that a habit from one cloud can create a security hole in another.
- AWS IAM attaches JSON policies to identities (users, groups, roles) or directly to resources such as S3 buckets. Applications and people from outside the account usually assume a role to get temporary credentials. The evaluation logic combines identity policies, resource policies, permission boundaries and organization-level policies, and an explicit deny always wins.
- Google Cloud IAM binds a principal to a role on a resource: an organization, folder, project or individual resource. Bindings are inherited downward, so a role granted on a folder applies to every project inside it. Workloads run as service accounts.
- Azure splits identity from authorization. Users, groups and workload identities live in Microsoft Entra ID; Azure RBAC role assignments then grant them roles at a scope (management group, subscription, resource group or resource), inherited downward. Applications use managed identities rather than stored secrets.
Here is the same task in each cloud: letting an application’s identity read objects from one storage bucket.
# AWS: attach a managed read-only policy to the application's role.
# (For one bucket only, you'd write a custom policy scoped to its ARN.)
aws iam attach-role-policy \
--role-name app-role \
--policy-arn arn:aws:iam::aws:policy/AmazonS3ReadOnlyAccess
# Google Cloud: grant the service account a role on one bucket
gcloud storage buckets add-iam-policy-binding gs://example-app-data \
--member="serviceAccount:app@example-project.iam.gserviceaccount.com" \
--role="roles/storage.objectViewer"
# Azure: assign a data-plane role to the managed identity, scoped to one storage account
az role assignment create \
--assignee "<managed-identity-principal-id>" \
--role "Storage Blob Data Reader" \
--scope "/subscriptions/<subscription-id>/resourceGroups/app-rg/providers/Microsoft.Storage/storageAccounts/exampleappdata"
Notice that the AWS example, as written, grants read access to every bucket in the account. That’s the easy path in AWS, and the reason least-privilege policies there take deliberate work. The three IAM articles go deeper: AWS IAM, Google Cloud IAM and Azure identity with Microsoft Entra ID.
In Google Cloud, you grant a user the Viewer role on a folder. What can they see?
Correct! Google Cloud IAM bindings are inherited down the hierarchy, so a role on a folder applies to all projects and resources beneath it.
How networks are scoped
This is the difference that most often surprises engineers moving to or from Google Cloud.
flowchart LR
subgraph AZN["Azure VNet: one region"]
ZS1["Subnet: spans the region's zones"]
end
subgraph GCPN["Google Cloud VPC: global"]
GS1["Subnet: region 1"]
GS2["Subnet: region 2"]
end
subgraph AWSN["AWS VPC: one region"]
AS1["Subnet: zone a"]
AS2["Subnet: zone b"]
end
| AWS | Google Cloud | Azure | |
|---|---|---|---|
| Virtual network scope | Regional (VPC) | Global (VPC) | Regional (VNet) |
| Subnet scope | One availability zone | One region | Whole region |
| Connecting networks across regions | Peering or Transit Gateway | Already one network | Peering or Virtual WAN |
| Default firewall model | Security groups on instances, network ACLs on subnets | VPC firewall rules and policies, targeted by tags or service accounts | Network security groups on subnets or NICs |
| Private connection to on-premises | Direct Connect | Cloud Interconnect | ExpressRoute |
On AWS, a multi-zone design needs a subnet in each zone. On Google Cloud, one subnet already covers every zone in its region, and a single VPC can have subnets in many regions that reach each other without peering. Azure sits in between: a VNet is regional, but its subnets span the region’s zones.
Load balancing follows the same pattern. Google Cloud’s external Application Load Balancer can be global, with one anycast IP address serving users worldwide. On AWS and Azure, regional load balancers are the default, and global entry points come from separate services: AWS Global Accelerator or CloudFront, and Azure Front Door.
Regions, zones and where you can deploy
All three providers divide the world into regions, and most regions into availability zones: separate data centers with independent power, cooling and networking, close enough for low-latency replication. A few differences matter in practice:
- AWS builds regions with multiple availability zones as standard, and identifies zones per account (
us-east-1ain one account may be a different physical zone than in another; use zone IDs to compare). - Google Cloud regions typically have three zones, named like
us-central1-a. - Azure offers availability zones in most but not all regions, and pairs many regions for disaster recovery and staggered platform updates.
Region counts change every year, so check the providers’ own maps rather than a number in an article. The infrastructure articles cover each in detail: AWS, Google Cloud and Azure.
You need VMs in three availability zones of one region. On which provider can a single subnet hold all of them?
Correct! Google Cloud subnets are regional and Azure subnets span the region's zones, while an AWS subnet lives in exactly one availability zone.
How pricing and discounts differ
The headline model is the same everywhere: pay for what you use, with discounts for commitment and for interruptible capacity. The mechanics differ.
| Discount type | AWS | Google Cloud | Azure |
|---|---|---|---|
| Commit to spend or usage | Savings Plans, Reserved Instances | Committed use discounts (resource-based or spend-based) | Savings plan for compute, Reservations |
| Automatic discount for steady use | None | Sustained use discounts on some older machine series | None |
| Interruptible capacity | Spot Instances | Spot VMs | Azure Spot Virtual Machines |
| Bring existing licenses | License Mobility, Dedicated Hosts | Sole-tenant nodes for some licenses | Azure Hybrid Benefit for Windows Server and SQL Server |
Two patterns hold across all three. Data coming in is generally free, and data going out to the internet is charged. And the cost of managed services, support plans and data transfer usually adds up to more than small differences in VM prices. The AWS billing article walks through the cost controls worth setting up on any of them.
Service map
The table pairs the services you are most likely to need. “Equivalent” means they fill the same role; features, limits and pricing still differ, so check the documentation before assuming a design ports directly.
| Category | AWS | Google Cloud | Azure |
|---|---|---|---|
| Virtual machines | EC2 | Compute Engine | Virtual Machines |
| Managed Kubernetes | EKS | GKE | AKS |
| Serverless containers | ECS with Fargate, ECS Express Mode | Cloud Run | Container Apps |
| Functions | Lambda | Cloud Run functions | Azure Functions |
| Platform as a service | Elastic Beanstalk | App Engine | App Service |
| Object storage | S3 | Cloud Storage | Blob Storage |
| Block storage | EBS | Persistent Disk, Hyperdisk | Managed Disks |
| File storage | EFS, FSx | Filestore | Azure Files |
| Managed relational | RDS, Aurora | Cloud SQL, AlloyDB | Azure SQL Database, Azure Database for PostgreSQL and MySQL |
| Key-value and document | DynamoDB | Firestore, Bigtable | Azure Cosmos DB |
| In-memory cache | ElastiCache | Memorystore | Azure Managed Redis |
| Data warehouse | Redshift | BigQuery | Microsoft Fabric, Azure Synapse Analytics |
| Messaging and events | SQS, SNS, EventBridge | Pub/Sub, Eventarc | Service Bus, Event Grid |
| Streaming | Kinesis, Amazon MSK | Pub/Sub, Managed Service for Apache Kafka | Event Hubs |
| Content delivery | CloudFront | Cloud CDN | Azure Front Door |
| DNS | Route 53 | Cloud DNS | Azure DNS |
| Web application firewall | AWS WAF | Cloud Armor | Azure Web Application Firewall |
| Keys and secrets | KMS, Secrets Manager | Cloud KMS, Secret Manager | Key Vault |
| Security posture | Security Hub, GuardDuty | Security Command Center | Microsoft Defender for Cloud |
| Monitoring and logs | CloudWatch | Cloud Monitoring, Cloud Logging | Azure Monitor |
| Audit trail | CloudTrail | Cloud Audit Logs | Activity log |
| Infrastructure as code | CloudFormation, CDK | Infrastructure Manager (Terraform) | ARM templates, Bicep |
| Container registry | ECR | Artifact Registry | Azure Container Registry |
| CI/CD | CodePipeline, CodeBuild | Cloud Build, Cloud Deploy | Azure Pipelines |
| ML platform | SageMaker AI | Vertex AI | Azure Machine Learning |
| Foundation models | Amazon Bedrock | Vertex AI (Gemini and partner models) | Microsoft Foundry |
Google maintains a longer AWS and Azure service comparison, and Microsoft publishes a similar AWS to Azure comparison. Both are useful when you need a service that isn’t in the table above.
Where the map is misleading
Some pairs in the table are close; others share a category but behave quite differently:
- BigQuery and Redshift. BigQuery is serverless and billed by data scanned or by reserved slots, with no cluster to size. Redshift offers both provisioned clusters and Redshift Serverless; the operational model depends on which you choose.
- DynamoDB, Firestore and Cosmos DB. All three are NoSQL databases, but their data models, consistency options and pricing units differ. Schema and access patterns designed for one rarely move unchanged.
- Cloud Run and its counterparts. Cloud Run runs containers and functions on the same platform, with request-based scaling to zero. On AWS, the same job is split between Lambda and ECS on Fargate, whose Express Mode sets up the load balancer and scaling for you.
- Azure Front Door. It combines a CDN, global load balancing and a WAF in one service, roughly what AWS does with CloudFront, Global Accelerator and AWS WAF together.
- Identity. Microsoft Entra ID is a full corporate directory that also serves Microsoft 365. AWS IAM and Google Cloud IAM handle cloud permissions; workforce identity usually comes from a separate directory, federated in.
What is the main operational difference between BigQuery and a provisioned Redshift cluster?
Correct! BigQuery has no cluster to size; you pay for queries or reserved slots. A provisioned Redshift cluster has nodes you choose and manage, though Redshift Serverless also exists.
Further reading
- Google Cloud: compare AWS and Azure services to Google Cloud
- Azure for AWS professionals
- AWS Organizations, Google Cloud resource hierarchy and Azure management groups
- For the virtual machine services specifically: Deep Dive into AWS EC2 Instances, Exploring GCP Compute Engine Instances and Understanding Azure Virtual Machines
Key takeaways
- The isolation boundary is the account on AWS, the project on Google Cloud and the subscription on Azure, with organizational units, folders and management groups above them.
- AWS IAM attaches JSON policies to identities and resources; Google Cloud binds principals to roles on resources with inheritance; Azure keeps identities in Entra ID and grants roles through Azure RBAC at a scope.
- Google Cloud VPCs are global with regional subnets; AWS VPCs are regional with zonal subnets; Azure VNets are regional with subnets that span zones.
- All three discount commitments and interruptible capacity, but the mechanics differ, and only Google Cloud applies automatic sustained use discounts, on some machine series.
- Most services have a counterpart on each cloud, but pairs like BigQuery and Redshift, or the three NoSQL databases, differ enough that designs rarely move unchanged.
Part 1 of 2 · Comparing the Cloud Providers