Leaf‑Spine Architecture in Large‑Scale Data Centers: Structure, Equipment Breakdown, and CAPEX Analysis

Facebook proposed the leaf‑spine architecture in 2014 to meet the rapid growth of east‑west traffic in ultra‑large data centers, and it has become the dominant topology for cloud providers worldwide.

The topology consists of three hierarchical layers: the cabinet layer (servers, Top‑of‑Rack switches, optical modules), the Leaf layer (Leaf switches, optical modules), and the Spine layer (Spine switches, optical modules). Each cabinet typically houses 24 servers and a ToR switch; each Leaf switch connects to 48 ToR switches, and each Spine switch interconnects all Leaf switches within a plane.

Adopting leaf‑spine drives a surge in high‑speed optical modules, especially 100 Gbps modules for inter‑leaf and leaf‑spine links, while server‑to‑ToR links commonly use 25 Gbps modules. The architecture shifts most traffic to short‑distance, high‑bandwidth east‑west flows, reducing north‑south load on spine switches.

Based on typical port ratios (1:6 up/down) and assuming a fully loaded cabinet (24 servers, 1 ToR), the equipment count per cabinet is estimated as: 24 servers, 1 ToR switch, 0.08 Leaf switches, 0.01 Spine switches, 48 × 25 Gbps optical modules, and roughly 28 × 100 Gbps optical modules. Detailed calculations for each layer are provided, showing how the numbers scale with the number of cabinets (N).

The financial breakdown shows that, for a typical IDC, the CAPEX ratio of servers : network equipment : optical modules is approximately 15 : 72 : 3.3 : 2.6 (or 16 % : 77 % : 4 % : 3 %). A fully loaded cabinet’s total investment ranges from ¥120,000 to ¥180,000, with the majority spent on servers and networking hardware.

Overall, the leaf‑spine design enables data centers to scale to hundreds of thousands of servers, improves bandwidth efficiency, and dictates the demand for next‑generation optical modules and high‑port‑density switches.