AKS Networking Models: Kubenet, Azure CNI, and Overlay

PracticiumMarch 30, 2026

AKS Networking Overview

Azure Kubernetes Service (AKS) supports multiple networking models—Kubenet, Azure CNI, and Azure CNI Overlay. Each model balances IP management, scalability, and Azure service integration differently.

AKS Networking Models
AKS Networking Models

Kubenet

Pods receive IPs from a private range managed by the node. Outbound traffic is NATed through the node’s IP, and User Defined Routes (UDRs) connect pods across nodes.

ProsConsBest Use Case
Conserves IP addressesLimited Azure service integrationSmall dev/test clusters
Simple setupNAT adds latencyIP conservation scenarios

Azure CNI (Classic)

Each pod gets an IP directly from the Azure VNet subnet, making pods first-class citizens in the VNet with seamless Azure service integration.

ProsConsBest Use Case
Full Azure integrationHigh IP consumptionEnterprise workloads needing direct pod-to-service communication
Lower latency (no NAT)Requires careful subnet planningModerate scale clusters

Azure CNI Overlay (Dynamic)

Pods get IPs from a separate overlay CIDR, not directly from the VNet. Traffic is encapsulated and mapped to the VNet, avoiding subnet exhaustion.

ProsConsBest Use Case
Scales to thousands of podsEncapsulation overheadLarge-scale, multi-tenant workloads
Avoids subnet exhaustionNewer model (less mature)Dynamic scaling SaaS platforms
Supports dual-stack IPv4/IPv6Future-proof production clusters

Comparison Table

FeatureKubenetAzure CNIAzure CNI Overlay
IP ConsumptionLowHighLow (overlay CIDR)
Azure IntegrationLimitedFullFull
PerformanceNAT overheadDirect, fastSlight encapsulation overhead
ScalabilityLimitedSubnet-boundVery high
ComplexitySimpleMedium (subnet planning)Medium (overlay config)

Practical Guidance

  • Use Kubenet for dev/test clusters where IP conservation is critical.
  • Use Azure CNI for enterprise workloads needing direct Azure service integration.
  • Use Azure CNI Overlay for high-scale, multi-tenant workloads to avoid IP exhaustion and enable dynamic scaling.