Before you configure a single VLAN, you pick a shape. The three campus/DC topologies solve different problems at different scales — and choosing the wrong one means either over-building a two-switch office into a three-tier campus, or cramming a data centre into a collapsed core that melts under east-west traffic. This is the decision, made in plain terms.
All three exist to move frames between hosts; they differ in how they trade device count, failure domain, and bandwidth predictability. Collapsed-core merges the core and distribution into one pair of boxes — cheapest, simplest, right for a single building. Three-tier (access → distribution → core) separates those roles so a large campus can grow modularly and contain faults per-building. Spine-leaf abandons the hierarchy entirely: every leaf connects to every spine, giving deterministic any-to-any bandwidth — the shape a data centre or a big virtualised campus needs when the traffic is east-west (server-to-server) rather than north-south (user-to-internet).
The mistake is treating this as a prestige ladder where "more tiers = better." It's a fit question. A 40-room hotel on a spine-leaf fabric is absurd; a 5,000-user hospital on a collapsed core is negligent. Get the shape right and everything downstream — STP, redundancy, segmentation — falls into place. Get it wrong and you fight the topology forever.
01
Core and distribution roles live on the same pair of Layer-3 switches. Access switches hang off them. It's the honest answer for the vast majority of businesses.
A single high-street office, a hotel, a clinic, a small factory — one wiring closet or two, a few hundred hosts at most. A pair of L3 switches does inter-VLAN routing, runs the first-hop redundancy, terminates the WAN, and aggregates the access switches. It's cheap, it's simple to reason about, and one competent engineer holds the whole thing in their head. The cost is a larger failure domain — the two core switches are a lot of eggs in one basket — which you mitigate with a redundant pair, aligned STP roots, and LACP uplinks, not with more tiers.
For an SMB, collapsed-core isn't the compromise — it's the correct answer, and anything more elaborate needs a reason. When a vendor proposes a three-tier design for a 60-person office, ask what specific problem the extra tier solves; usually it's a bigger invoice, not a real requirement. The times to graduate: multiple buildings/closets that need independent fault domains, or a host count and switch count large enough that one distribution pair can't aggregate it cleanly.
02
Access, distribution, and core become separate layers. Each building or zone gets its own distribution pair; the core just moves traffic between distribution blocks fast. This is how you scale a campus without one giant flat failure domain.
When you have several buildings, hundreds of switches, and thousands of users, the distribution layer becomes the modular unit: each building's access switches home to that building's distribution pair, which connects to a lean, fast core. A fault, a maintenance window, or a broadcast storm is contained to one distribution block instead of taking down the campus. The distribution layer is also the natural routing and policy boundary — summarise routes there, enforce ACLs there, terminate VLANs there. The core stays deliberately dumb and fast: no policy, just forwarding.
03
Every leaf connects to every spine; hosts connect to leaves; spines connect only to leaves. Any host is exactly two hops from any other, with equal bandwidth on every path. The data-centre answer.
Hierarchical designs assume traffic is mostly north-south — users reaching the internet or a few central servers. Virtualisation and modern apps flipped that: the heavy traffic is east-west, server-to-server, and in a three-tier design that traffic hairpins up to the core and back down, oversubscribed and unpredictable. Spine-leaf (a Clos fabric) fixes it structurally — every leaf is one hop from every spine, so any-to-any bandwidth is deterministic and scales by adding spines. It runs an ECMP-routed underlay (OSPF/BGP) with a VXLAN/EVPN overlay carrying the tenant networks — the subject of its own post. You reach for it in the data centre, a large virtualised campus, or anywhere east-west bandwidth must be guaranteed — and almost nowhere else.
04
Match the shape to the situation, not to fashion. This is the whole post compressed.
| If… | Use |
|---|---|
| One building, ≤ few hundred hosts, SMB | Collapsed-core (redundant pair) |
| Multiple buildings/closets, independent fault domains | Three-tier |
| Thousands of users, modular growth, per-block policy | Three-tier |
| Data centre / heavy east-west / VM mobility | Spine-leaf + VXLAN/EVPN |
| Guaranteed any-to-any bandwidth at scale | Spine-leaf |
| "A vendor said we need three tiers" for a 60-person office | Ask what problem it solves — usually collapsed-core |
Takeaways
NOCTIS designs right-sized network architectures across Crete — from a clean collapsed-core for a hotel to modular campus and fabric designs — matched to your actual scale and growth, not a vendor's invoice.
Book a Discovery Call →