cloud networking · field guide

load balancers, decoded
across aws, azure & gcp

Five functional categories, three clouds, a dozen-plus SKUs — and one architectural fork that splits the hyperscalers down the middle. Here's the map.

12 min reference updated jun 2026 terminal
part 1

the five categories

Strip the vendor names away and every hyperscaler's load balancing portfolio resolves to the same five functional patterns.

layer 4
network

TCP / UDP

layer 7
application

HTTP / HTTPS

global
anycast

EDGE ENTRY

internal
east-west

VPC / VNET

appliance
insertion

L3 · GENEVE

the global-vs-regional fork

The first dividing line between the three clouds isn't a feature — it's a philosophy.

aws + azure's default posture

regional-by-default

Each load balancer lives in one region. Global reach requires a separate edge layer — AWS Global Accelerator or Azure Front Door — stacked in front.

the appliance-insertion gap

The second fork: who has a purpose-built primitive for transparently steering traffic through third-party firewalls and IDS appliances.

gcp

no native equivalent

Closest pattern uses Private Service Connect producer/consumer architecture — fundamentally different mechanics, and considerably less standardized.

the mental model: if you remember nothing else — GCP optimizes for "one IP, anywhere." AWS and Azure optimize for "one region, then layer global on top." Appliance insertion is AWS and Azure's answer to a problem GCP hasn't formally solved yet.

part 2

deep dive — sku by sku

Every product name, every layer, every status flag. Including two Azure retirements that matter right now.

AWS

elastic load balancing

the flattest portfolio — four types, no internal tiering, consumption-based capacity units

~/aws/elb $ list-skus --output table
skulayermechanismstatus
ALB L7 HTTP/HTTPS/WebSocket/gRPC, path/host/header routing → EC2, IP, Lambda, ECS/EKS default choice
NLB L4 flow-hash TCP/UDP/TLS, static IP per AZ, source IP preserved, millions of conns/sec current
GWLB L3 GENEVE encapsulation (port 6081), GWLBe (PrivateLink-based), transparent appliance insertion current
CLB L4/L7 round-robin or least-outstanding-requests, no WAF, no WebSocket/HTTP2 legacy
global accelerator L4 static anycast IPs at the edge → NLB / ALB / EC2 / EIP. accelerates any TCP/UDP, not just HTTP separate product
pricing model

Hourly base (~$0.0225/hr) + capacity-unit consumption — LCU for ALB, NLCU for NLB, GLCU for GWLB.

notable differentiator

Global Accelerator is the only global-edge product across all three clouds that accelerates any TCP/UDP app — not just HTTP.

AZURE

load balancing & content delivery

the most tiered lineup — and the most recently re-tiered, with two retirements live right now

two retirements to check before you advise anyone: Basic Load Balancer retired 30 sep 2025. Application Gateway v1 (Standard/WAF) retired 28 apr 2026. Any customer running either needs a migration plan now.

~/azure/lb $ list-skus --output table
skulayermechanismstatus
standard lb L4 TCP/UDP, zone-redundant, outbound rules, HTTPS health probes current default
basic lb L4 regional only, no zones, no SLA retired 9/2025
gateway lb L3 GENEVE, chains to Standard LB frontend + Private Link consumer model — close port of AWS GWLB current
app gateway v2
standard_v2 / waf_v2
L7 HTTP(S)/WebSocket/HTTP2, path/host routing, autoscaling, dedicated subnet current
app gateway v1 L7 no autoscale, manual sizing retired 4/2026
front door
standard / premium
L7 global CDN + WAF + global HTTP routing, anycast. premium adds private link origins, bot protection current
traffic manager DNS DNS-level geo/priority/performance routing — not a data-plane proxy distinct category
most common confusion

Front Door (edge/global) vs. App Gateway (regional/VNet-injected). Teams often reach for App Gateway when they actually need Front Door for global HA.

why v1 retired

v2 added autoscaling — the single biggest functional gap that made v1 unsuitable for modern production workloads.

GCP

cloud load balancing

the most unified mental model — three primitives, multiple deployment scopes, global-first

~/gcp/lb $ list-skus --output table
skulayer / scopemechanismstatus
global ext. app lb L7 · global anycast IP, GFE + Envoy, native multi-region backends flagship
classic app lb L7 · global/regional GFE-based legacy — global in Premium Tier, regional-only in Standard Tier legacy
regional ext. app lb L7 · regional Envoy-based, single-region backends current
cross-region internal app lb L7 · multi-region internal Envoy-based, HA / failover across regions current
regional internal app lb L7 · regional internal Envoy + Andromeda stack current
proxy network lb
global / regional / internal
L4 (TCP/SSL proxy) terminates TCP/SSL, non-HTTP(S) traffic current
passthrough network lb
external / internal
L4 Direct Server Return, Maglev (ext) / Andromeda (int), broadest protocol range — TCP/UDP/SCTP/ESP/AH/GRE current
cross-cutting dimension

Network Service Tiers. Premium routes over Google's private backbone; Standard routes over public internet. Not every LB type supports Standard — a real design constraint, not just a cost lever.

on the radar

Post-quantum key exchange on SSL policies is rolling out now — default-on phased in oct 2026 → oct 2027. Worth tracking for PQC-readiness conversations.

the appliance-insertion gap, confirmed

GCP has no native GWLB equivalent. The closest pattern uses Private Service Connect producer/consumer architecture — fundamentally different from GENEVE-based steering, and considerably less standardized. Most GCP customers historically hand-rolled this with custom routes and an Internal LB, the way AWS customers did before GWLB existed. GCP increasingly pushes toward managed third-party integrations (e.g. Palo Alto Cloud NGFW for GCP) instead of a native primitive.

at a glance

aws
5
distinct products (incl. global accelerator)
  • flattest tiering model
  • consumption-based LCU pricing
  • GWLB = mature, first mover
azure
7
SKUs across 4 product lines
  • most internal tiering (v1/v2, std/waf)
  • 2 retirements in past 12 months
  • front door / app gateway overlap confuses
gcp
3
core primitives, many scopes
  • most unified mental model
  • global-by-default architecture
  • no native appliance-insertion sku
part 3

use cases by category

When to actually reach for each pattern, regardless of vendor.

layer 4 — network

NLB · Azure Standard LB · GCP Network LB
  • Ultra-low-latency TCP/UDP — gaming, trading, IoT telemetry ingestion
  • Static IP requirements for firewall allowlisting or partner peering
  • Preserving real client source IP for backend geo-routing or fraud detection
  • Non-HTTP protocols — MQTT, SMTP, DB connection pooling
  • TLS passthrough for compliance-driven end-to-end encryption
  • Absorbing extreme spikes without L7 processing overhead

layer 7 — application

ALB · App Gateway · GCP Application LB
  • Microservices routing by path or host to backend pools
  • WAF-integrated edge defense against OWASP Top 10
  • Blue/green and canary deploys via weighted target groups
  • Container-native routing — ECS/EKS, AKS (AGIC), GKE
  • Auth offload before traffic reaches backends
  • HTTP-triggered serverless backends (ALB → Lambda)

global / anycast

Global Accelerator · Front Door · GCP Global LB
  • Multi-region active-active behind a single IP or hostname
  • DDoS resilience absorbed at the provider edge, not your origin
  • DR with health-check-driven failover — faster than DNS TTL failover
  • CDN + load balancing convergence for content-heavy global apps
  • Non-HTTP global acceleration — Global Accelerator's standout

internal / east-west

Internal NLB/ALB · Azure Internal LB · GCP Internal LB
  • Service mesh ingress entirely within a VPC/VNet
  • Multi-tier apps — independently scaled web, app, data tiers
  • Compliance-driven segmentation — traffic never touches the public internet
  • Sitting behind PrivateLink, PSC, or Private Endpoint as the target

appliance insertion

GWLB (AWS) · Gateway LB (Azure) · no native GCP sku
  • Centralized firewall inspection in hub-and-spoke topologies — all spoke traffic transits a shared inspection VPC/VNet
  • Transparent IDS/IPS insertion (Palo Alto, Fortinet, Check Point) without re-architecting routing per network
  • Compliance-mandated deep packet inspection for regulated industries
  • Multi-tenant MSSP architectures offering security-as-a-service across customer environments
the bigger picture
Every category is converging conceptually. The implementations remain stubbornly vendor-specific.

Five categories, three clouds, a dozen-plus SKUs — and the one category without a true equivalent everywhere is appliance insertion. AWS and Azure both built dedicated GENEVE-based primitives because centralized inline inspection chaining was painful enough to warrant new infrastructure. GCP's gap is itself a signal: even the hyperscalers are acknowledging that bump-in-the-wire appliance chaining is a tax worth engineering around.