splithorizon // infographic
SPLITHORIZON // CLOUD NETWORKING Transit Gateway vs. Virtual WAN vs. NCC Three hub-and-spoke fabrics. One question: who actually transits regions? Verified against AWS / Microsoft / Google docs — June 2026 01 / HUB SCOPE & ARCHITECTURE AWS Transit Gateway Hub scope: REGIONAL One regional router per TGW. No global control plane — every region stands alone. INTER-REGION TRANSIT Manual peering — static routes only 3RD-PARTY APPLIANCE TGW Connect — GRE + BGP overlay 4 peers · 5 Gbps each · 20 Gbps total cap VPC → HUB ROUTES Internal API — not BGP Peering links per TGW pair: 1 (hard limit) AZURE Virtual WAN Hub scope: REGIONAL HUB Managed hub router per region; hubs in the same vWAN auto-mesh to each other. INTER-REGION TRANSIT Automatic — hub-to-hub BGP full mesh 3RD-PARTY APPLIANCE NVA-in-hub BGP peering Direct-connect VNet only · 16-bit ASN only VPC → HUB ROUTES BGP (hub router) Hub router throughput: 50 Gbps shared GOOGLE CLOUD Connectivity Center Hub scope: GLOBAL One hub spans every region. No separate inter-region object — it's just the hub. INTER-REGION TRANSIT Native — same global hub, no extra step 3RD-PARTY APPLIANCE Router Appliance spoke / NCC Gateway GCE VM + Cloud Router BGP · 8 per spoke VPC → HUB ROUTES BGP (dynamic route exchange) Topology locked at creation: mesh or star 02 / SCALE LIMITS THAT BITE LIMIT AWS TGW AZURE vWAN GCP NCC Hubs / accounts 5 per account global vWAN container quota-based per project Spokes / attachments per hub 5,000 ~500 minus hub count 250 active / 1,000 total Combined route table size 10,000 routes 10,000 routes quota-based dynamic routes Peering / cross-hub links 1 per TGW pair (hard) auto full-mesh n/a — single global hub Throughput ceiling ~100 Gbps / VPC attach / AZ 50 Gbps shared per hub router scales with Cloud Router / VMs VPN BGP route cap 100 (hard — session resets) per ExpressRoute circuit limits quota-based per hub 3rd-party appliance throughput 4 Connect peers, 20 Gbps cap NVA peers as branch attach 8 router appliance VMs/spoke ASN support 32-bit (standard BGP) 16-bit ASN only (hub router) 32-bit (Cloud Router standard) 03 / BGP ROUTE SHARING — THE REAL DIFFERENTIATOR CAPABILITY AWS TGW AZURE vWAN GCP NCC Hub-to-hub route propagation Static only — no BGP BGP, automatic BGP, automatic (1 hub) On-prem → hub propagation BGP (VPN/DX/Connect) BGP (VPN/ER/NVA) BGP (VPN/Interconnect/RA) Per-spoke route filtering route table assoc / prop route table + labels export filters, 16 CIDRs/spoke ECMP same attach type/path only implicit, hub router implicit, Cloud Router IPv6 dynamic route exchange Yes (on Connect) Not supported Not supported Static-route-only edge design Not possible — BGP mandatory discouraged, breaks selection not supported (preview adds-on) MTU (fabric / VPN) 8500 / 1500 bytes platform-managed platform-managed THE MENTAL MODEL THAT MATTERS TGW is a regional box you peer by hand. vWAN auto-meshes regional hubs over BGP. NCC skips the regional layer entirely — the hub is the global fabric. Pick based on how your org already thinks about regions and route ownership — not the spoke count on a slide. 04 / DESIGN CHECKLIST AWS-ANCHORED Budget for TGW-peering static-route overhead — it scales with region-pairs, not spoke count. One peering link max per TGW pair — no ECMP across it. → consider Cloud WAN to abstract this AZURE-ANCHORED Routing Intent changes NVA BGP rules — peering needs a secured hub. Size against the 600 address-space ceiling per hub, not raw VNet count. → check prefix count, not connection count GCP-ANCHORED Global hub means a routing mistake's blast radius is global by default. Mesh vs. star topology is locked at creation — get it right the first time. → use export filters for overlapping CIDRs splithorizon Decoding the modern network — from enterprise to AI. Source data verified against AWS / Microsoft / Google documentation, June 2026 · Quotas are adjustable — treat as sizing baselines, not hard ceilings. CLOUD NETWORKING · ROUTING · THE JOURNEY