How to Build a Leaf-Spine Network Architecture — Complete Guide (2026)
Step-by-step guide to designing and building a leaf-spine network from scratch — hardware selection, cabling, BGP underlay routing, and EVPN/VXLAN overlay configuration.
What Is a Leaf-Spine Architecture?
Leaf-spine is a two-tier, non-blocking network fabric built for east-west traffic — server-to-server communication that dominates modern virtualized, containerized, and cloud-native environments.
The two layers:
- Leaf switches sit at the edge, connecting servers, storage, and endpoints
- Spine switches sit in the middle, connecting only to leaf switches — never to each other
The defining rule: every leaf connects to every spine. Any leaf can reach any other leaf in exactly two hops — leaf to spine to leaf. Traffic is load-balanced via ECMP across all spine links simultaneously.
This is why leaf-spine replaced the traditional three-tier model (access, distribution, core) for data center use. In a three-tier design, east-west traffic between two servers on different access switches had to travel up to the distribution tier and back down. In a leaf-spine fabric, it takes exactly two hops regardless of which racks the servers are in.
Step 1 — Size Your Fabric
Before selecting hardware, answer three questions:
How many servers? Each leaf typically has 48 downlink ports. Plan for 20 to 30 percent growth in your initial design.
What speed do your servers need?
- 10G NICs → 10G leaf downlinks (SFP+)
- 25G NICs → 25G leaf downlinks (SFP28)
- 100G NICs → 100G leaf downlinks (QSFP28)
What oversubscription ratio is acceptable? Formula: total downlink bandwidth divided by total uplink bandwidth to spine.
A 48x10G leaf with 6x40G uplinks: 480G downlink / 240G uplink = 2:1 oversubscription. For most enterprise workloads, 3:1 to 4:1 is acceptable. For storage replication and AI training, target 2:1 or below.
Step 2 — Select Your Spine Switches
Spine switches do one thing: forward traffic between leaves at line rate with minimal latency.
Key selection criteria:
- Port count must accommodate one port per leaf switch, times two (always deploy two spines minimum)
- Speed must match or exceed leaf uplink speed
- Latency: target sub-1 microsecond cut-through forwarding
- Buffer depth: shallow for HFT/HPC, deep for mixed storage and compute
Recommended spine platforms:
| Environment | Switch | Ports | Speed |
|---|---|---|---|
| SMB / mid-market | Arista 7050QX-32S | 32 | 40G |
| Enterprise | Cisco Nexus 9332PQ | 32 | 40G |
| Modern cloud-native | Arista 7050CX3-32S | 32 | 100G |
| AI/ML optimized | NVIDIA SN2700 | 32 | 100G |
Step 3 — Select Your Leaf Switches
Leaf switches connect directly to servers, storage, firewalls, and load balancers.
Standard leaf configuration:
- 48 downlink ports matching server NIC speed (SFP+, SFP28, or QSFP28)
- 6 to 8 uplink ports to spine at higher speed (QSFP+ or QSFP28)
- 1RU form factor
Recommended leaf platforms:
| Server Speed | Leaf Switch | Notes |
|---|---|---|
| 10G | Cisco Nexus 93180YC-FX | 48x10G SFP+ + 6x100G QSFP28 |
| 25G | Arista 7050CX3-48YC12 | 48x25G SFP28 + 12x100G QSFP28 |
| 25G | Juniper QFX5100-48S | 48x10G + 6x40G QSFP+ |
Step 4 — Select Transceivers and Cables
Leaf-to-spine uplinks:
- Same cabinet row → QSFP+ or QSFP28 DAC passive (cheapest, lowest latency)
- Cross-row under 30m → QSFP+ or QSFP28 AOC
- Longer runs → QSFP+ SR4 or QSFP28 SR4 over OM4 fiber
Leaf-to-server downlinks:
- Same cabinet → SFP+ or SFP28 DAC passive
- Cross-cabinet → SFP+ SR or SFP28 SR over OM4 fiber
Structured cabling:
- SFP+/SFP28 fiber connections → LC/LC OM4 duplex patch cables
- SR4 transceivers → MPO-12 OM4 trunk cables
Step 5 — Design IP Addressing
Leaf-spine uses Layer 3 routing at the leaf. Each leaf is a Layer 3 boundary:
- Spine-to-leaf links: /31 point-to-point subnets (e.g. 10.0.0.0/31 per link pair)
- Leaf loopbacks: /32 per device (e.g. 10.255.0.1/32 per leaf)
- Server subnets: /24 or /25 per leaf
- Spine loopbacks: /32 per spine
Step 6 — Configure BGP Underlay
BGP is the routing protocol of choice for leaf-spine. It is vendor-neutral, scales to thousands of nodes, and provides fine-grained path control.
eBGP model (recommended):
- Each leaf has a unique AS number
- All spines share one AS number (or each spine has its own)
- Each leaf peers eBGP with every spine over point-to-point /31 links
- Leaves advertise loopbacks and server subnets
- Spines advertise loopbacks only
- ECMP is automatic — every path to a destination has equal cost
Example AS scheme:
- Spine 1: AS 65000, Spine 2: AS 65001
- Leaf 1: AS 65100, Leaf 2: AS 65101, and so on
Step 7 — Deploy EVPN/VXLAN Overlay (Optional)
Pure Layer 3 fabric works for native-IP workloads. For environments needing Layer 2 extension (VMotion, clustered applications, legacy systems), add VXLAN with an EVPN control plane.
Key concepts:
- VTEP (VXLAN Tunnel Endpoint) — one per leaf, hosted on the loopback interface
- VNI (VXLAN Network Identifier) — maps to a VLAN (L2) or VRF (L3)
- Anycast Gateway — all leaves share the same virtual gateway IP/MAC, eliminating first-hop routing latency
EVPN/VXLAN configuration syntax differs between NX-OS, EOS, Junos, and AOS-CX — the RFC behavior is consistent but the CLI is not. Plan your NOS choice before starting configuration.
Step 8 — Physical Cabling Best Practices
- Label every cable on both ends before installing
- Use color coding by spine — blue for spine 1, green for spine 2
- Document the topology in a DCIM tool before you start cabling
- Install horizontal cable managers between every switch pair in the rack
- Verify with LLDP after cabling — show lldp neighbors on every switch
Common Mistakes to Avoid
Connecting spines to each other: Spines connect only to leaves. Spine-to-spine links violate the design and create forwarding loops.
Under-sizing spine port count: Every leaf must connect to every spine. Plan spine port count for your 3-year growth projection.
Mixing OM3 and OM4 in SR4 links: OM3 limits 40G SR4 to 100m. OM4 extends to 150m. Verify your installed fiber grade before ordering transceivers.
Ignoring ECMP hashing: Most platforms hash traffic flows across ECMP paths using source/destination IP, protocol, and port. Large single-flow transfers (backup jobs, single TCP streams) cannot be distributed. Design with this limitation in mind.
Key Takeaway: A well-built leaf-spine fabric with two spines, six leaves, and proper ECMP delivers 2.4Tbps of bisectional bandwidth at 40G — enough for 240 servers at full 10G utilization. The math scales linearly as you add spines.