Executive summary
Multi‑cloud architecture is not a single console that magically governs Azure, Amazon Web Services and Alibaba Cloud. It is a set of four decisions—where a workload runs, which identity system controls it, how network traffic reaches it, and where the system of record lives—made separately for each provider. By documenting those decisions per workload, teams can avoid the common slogan that multi‑cloud is simply “using more than one cloud”. This deep dive explains the technical fundamentals of each provider, outlines an architecture that respects the boundaries, and gives a concrete implementation approach with diagrams, tables and code snippets.
Introduction
Enterprises increasingly spread workloads across Azure, AWS and Alibaba Cloud to meet latency, regulatory and cost objectives. The promise of a unified multi‑cloud control plane often masks the reality that each cloud has its own governance model, identity service and networking limits. When does a multi‑cloud strategy become a requirement, and when is it just marketing speak? This article answers that by examining the four core decisions that remain distinct regardless of how many clouds are involved.
Technical fundamentals
Azure landing zones
Microsoft’s Cloud Adoption Framework defines a landing zone as a flexible architecture for governing, securing and scaling a multi‑subscription Azure environment. A platform landing zone provides a centralized foundation—management‑group hierarchy, policy inheritance, shared connectivity (hub‑and‑spoke or Virtual WAN) and security monitoring. Workload landing zones are the subscription spaces where application teams deploy resources. A workload is a collection of Azure resources, code, models, data and infrastructure that together produce a business outcome. The platform landing zone does not extend into AWS or Alibaba Cloud accounts.
AWS Identity and Access Management (IAM)
IAM is the web service that controls authentication and authorization inside an AWS account. Each account starts with a root user that has full permissions; best practice is to avoid using the root user for daily tasks. Administrators create users, groups, roles and attach policies that grant the least privilege needed. Authorization is evaluated per request against the principal’s identity, attached policies and any resource‑based policies. IAM changes are highly available but eventually consistent, so a permission change may not be visible instantly.
Alibaba Cloud VPC and RAM
Alibaba Cloud’s network planning separates physical and logical layers. A Virtual Private Cloud (VPC) is confined to a single region; a VPC cannot span regions. Within a VPC, a vSwitch is a zonal subnet. Cross‑region connectivity requires a VPC peering connection, Cloud Enterprise Network, or VPN gateway. Default quotas are ten VPCs per region and 150 vSwitches per VPC, both raiseable. After a CIDR block is assigned to a vSwitch it cannot be changed, and overlapping CIDR blocks between VPCs prevent communication. Resource isolation is achieved with separate accounts or Resource Access Management (RAM) users for delegation.
Architecture and components
The architecture separates each cloud into its own column with four swim lanes representing the four decisions. Only the inventory projection from Azure Arc’s multicloud connector crosses columns, and a deliberately engineered network path connects a specific VPC to an Azure hub for data extraction.
+-------------------+-------------------+-------------------+
| Azure | AWS | Alibaba Cloud |
+-------------------+-------------------+-------------------+
| Workload placement| Workload placement| Workload placement|
| Identity (Entra) | Identity (IAM) | Identity (RAM) |
| Network (VWAN) | Network (SG/VPC) | Network (VPC) |
| Data home (SQL) | Data home (S3) | Data home (OSS) |
+-------------------+-------------------+-------------------+
^ ^
| Inventory projection (Azure Arc) |
+-----------------------------------+
Table of core objects per decision
| Decision | Azure | AWS | Alibaba Cloud |
|---|
| Placement | Subscription in a landing zone | Account | Account or RAM user |
| Identity | Entra ID tenant & role assignment | IAM user / role / policy | RAM user or separate account |
| Network | Hub‑and‑spoke VNet or Virtual WAN | VPC, security groups, route tables | VPC, vSwitch, peering or VPN |
| Data home | Azure SQL, Cosmos DB, storage account | Amazon S3, RDS, DynamoDB | OSS, ApsaraDB, PolarDB |
Detailed implementation approach
- Define primary cloud and region – Choose based on latency, data‑localization rules, service availability or existing team expertise. Document the justification.
- Create isolated account boundaries – For Azure, provision a platform landing zone and separate workload subscriptions under a management‑group hierarchy. For AWS, create a dedicated account for the workload; assign IAM roles and policies. For Alibaba Cloud, create a new account or RAM user and a VPC in the chosen region.
- Apply governance – Use Azure Policy to enforce naming, tagging and security standards in Azure. Use AWS Service Control Policies (SCP) at the organization level if applicable, noting that they are not covered by the source notes but are a best‑practice extension. In Alibaba Cloud, enforce quota requests and CIDR planning manually.
- Configure identity – Map each human operator to three identities: Entra ID, IAM principal and RAM user. Automate provisioning via SCIM or custom scripts; treat joiner and leaver processes as three separate workflows.
- Establish network paths – Deploy a hub VNet in Azure with a Virtual WAN gateway. In AWS, create a VPC with a Transit Gateway if needed. In Alibaba Cloud, set up a VPC and a VPN gateway that terminates to the Azure hub. Use non‑overlapping CIDR blocks across all VPCs and VNets.
- Determine data gravity – Identify the system of record. If the record resides in Azure, replicate data to AWS S3 using Azure Data Factory or a third‑party tool, treating the copy as a distinct store with its own IAM permissions. If the record is in Alibaba Cloud, keep compute close to the VPC and only extract needed data to other clouds.
- Enable Azure Arc multicloud connector – Activate the connector to inventory AWS EC2 instances and Google Cloud VMs (note: Alibaba Cloud is not listed). The connector provides read‑only resource metadata in Azure Resource Graph; it does not create VPCs or translate IAM permissions.
- Validate consistency – After any IAM or RAM change, wait for eventual consistency before relying on the permission in production. Verify propagation by testing a low‑impact operation.
- Document hand‑off – Record the four decisions for each workload in a shared repository. Include CIDR plans, identity mappings and data‑home location.
Architecture and data‑flow
The diagram below illustrates a typical data‑flow where user requests hit an Azure Front Door, are routed to an Azure App Service (system of record), and a background job streams anonymized analytics to an AWS Lambda function via an encrypted VPN tunnel. The analytics job writes to an Amazon S3 bucket, which is then processed by an Alibaba Cloud Function Compute instance after a VPC peering connection transports the data across the VPN.
[User] --> Azure Front Door --> Azure App Service --> Azure SQL (system of record)
|
| VPN (Azure ↔ AWS)
v
AWS Lambda --> S3
|
| VPC Peering (AWS ↔ Alibaba)
v
Alibaba Function Compute --> OSS
Security and governance
- Identity segregation – Entra ID groups do not appear in IAM policies, and IAM roles are invisible to Azure management groups. Each cloud requires its own provisioning pipeline and termination process.
- Network isolation – Overlapping CIDR blocks cause communication failure in Alibaba Cloud and Azure. Use a central address‑plan spreadsheet and assign distinct /16 blocks per cloud.
- Data protection – Encryption at rest is native in each provider (Azure Storage Service Encryption, AWS SSE, Alibaba OSS encryption). Cross‑cloud transfers should use TLS‑based VPNs or dedicated encryption gateways.
- Policy enforcement – Azure Policy can enforce tagging and security standards. AWS SCPs can enforce similar guardrails, but they are outside the source material; mention them as an optional extension. Alibaba Cloud lacks a native policy engine, so manual review and quota enforcement are required.
Enterprise use cases
- Global e‑commerce – Front‑end hosted in Azure for low latency to North America, inventory data replicated to AWS for analytics, and payment processing kept in Alibaba Cloud to satisfy Chinese regulatory requirements.
- Media streaming – Video origin stored in Azure Blob, edge cache nodes deployed in Alibaba Cloud regions close to Southeast Asian users, with transcoding workloads in AWS because of specialized GPU instances.
- Financial services – Core transaction database remains in Azure for compliance, risk models run in AWS for access to specific ML services, and disaster‑recovery snapshots are stored in Alibaba Cloud OSS to diversify risk.
Benefits and limitations
Benefits
- Ability to locate compute where latency or compliance matters.
- Leverage best‑of‑breed services that may exist only in one provider.
- Reduce vendor lock‑in by spreading risk.
Limitations
- Increased operational overhead – three separate identity lifecycles, three networking stacks, three data‑gravity decisions.
- No unified failover; each cloud’s SLA applies only to its own resources.
- Cross‑cloud network paths add latency and complexity; overlapping CIDR blocks cause failures.
- Azure Arc multicloud connector provides inventory only; it does not create a shared control plane.
Operational considerations
- Monitoring – Deploy native monitoring agents (Azure Monitor, CloudWatch, Alibaba Cloud CloudMonitor) and aggregate logs into a central SIEM via syslog or API ingestion.
- Cost management – Tag resources consistently in each cloud; use Azure Cost Management, AWS Cost Explorer and Alibaba Cloud Billing to consolidate reports.
- Change management – Because IAM changes are eventually consistent, schedule permission updates during low‑traffic windows and verify propagation before release.
- Disaster recovery – Treat each cloud copy as a separate system of record; define RPO and RTO per copy rather than assuming a single recovery point.
Best practices
- Document the four decisions per workload – placement, identity, network, data home.
- Use non‑overlapping CIDR blocks across all VPCs and VNets.
- Assign dedicated accounts or subscriptions to isolate production from test and to limit blast radius.
- Automate identity provisioning across Entra ID, IAM and RAM to keep joiner/leaver processes in sync.
- Leverage Azure Arc only for inventory – do not rely on it for cross‑cloud networking or IAM translation.
- Validate network paths with end‑to‑end ping and traceroute after any VPC peering or VPN change.
- Treat copies as separate stores – implement separate access controls and monitor replication lag.
Analysis
The source material makes clear that multi‑cloud is fundamentally four independent decisions. Azure landing zones provide a governance framework only for Azure; they do not extend to AWS accounts or Alibaba Cloud VPCs. IAM operates inside an AWS account and is eventually consistent, meaning permission revocation is not instantaneous. Alibaba Cloud’s VPC limits and immutable CIDR blocks reinforce the need for careful address planning. The Azure Arc multicloud connector offers a read‑only inventory view of AWS and Google Cloud resources but does not create a unified management plane, and it does not support Alibaba Cloud. Consequently, any design that assumes a single console can provision networking, identity or data‑home across all three clouds is unsupported.
What the sources do not settle
- The exact cost impact of maintaining three separate governance stacks.
- How to implement organization‑wide policy enforcement in AWS without a native Azure‑like Policy engine.
- The best method for synchronizing user lifecycle events across Entra ID, IAM and RAM beyond custom scripts.
- Performance characteristics of cross‑cloud VPN tunnels versus dedicated interconnects.
- Strategies for global disaster recovery that span all three clouds with a single RTO.
Conclusion
Multi‑cloud architecture is viable when each workload’s placement, identity, network and data‑home decisions are made deliberately and documented. Azure landing zones, AWS IAM and Alibaba Cloud VPCs each solve their part of the puzzle, but none provides a shared control plane. By respecting the boundaries, using non‑overlapping address spaces, automating identity provisioning and leveraging Azure Arc for inventory only, enterprises can achieve the benefits of latency optimization, regulatory compliance and risk diversification without falling into the slogan‑driven trap of “just use three clouds”.
References
What is an Azure landing zone? | Microsoft Learn
What is Multicloud connector enabled by Azure Arc? | Microsoft Learn
What is IAM? | Amazon Web Services
Plan networks for VPC | Alibaba Cloud
These published CloudExchangers articles are further desk reading. They are not sources for the technical claims above.