The data centre fiber infrastructure is the physical foundation of every digital service — and it is designed to last 15–20 years while the active equipment it connects is replaced every 3–5 years. The consequence of specifying the wrong fiber type, the wrong connector density, or the wrong cable pathway capacity today is a full infrastructure replacement when the next transceiver generation arrives — a project measured in months and millions, not weeks and thousands.
AI workloads have forced this infrastructure decision to an earlier inflection point than expected. The east-west traffic demands of GPU clusters running LLM training — hundreds of thousands of gradient synchronisation messages per second between hundreds of GPUs — make 10G spine connections appear as inadequate as dial-up modems. The fiber infrastructure specified for the next data centre build must be designed for 400G today, 800G in two equipment cycles, and 1.6T beyond that.
Fiber Type Selection for Data Centre
| Fiber Type | Bandwidth-Distance | Max Reach at 100G | Max Reach at 400G SR8 | SWDM Support | Best Use Case |
|---|---|---|---|---|---|
| OM3 | 2,000 MHz·km | 100 m | 50 m | No | Legacy — not recommended for new builds |
| OM4 | 4,700 MHz·km | 150 m | 100 m | No | Acceptable for 400G intra-hall short reach |
| OM5 | 28,000 MHz·km (SWDM) | 150 m | 150 m | Yes — SWDM4, SWDM8 | Recommended — 400G/800G future-proof |
| OS2 | Unlimited (SMF) | 40 km+ | 500 m (DR4) / 2 km (LR4) | DWDM/CWDM | Inter-building, DCI, long campus links |
Spine-Leaf Design Principles
- Equal-cost multipath (ECMP): Every leaf connects to every spine — any-to-any communication between two servers traverses exactly two hops (leaf → spine → leaf). ECMP routing load-balances traffic across all spine-to-leaf links simultaneously
- Fiber symmetry requirement: ECMP requires equal latency across all paths between any leaf pair — fiber length from each leaf to each spine must be equal (within the transceiver's latency sensitivity). Pre-measured, equal-length trunk cables from a central overhead tray achieve this in practice
- Breakout strategies for 400G: MPO-16 trunk cables break out to individual 400GBASE-SR8 links at each end via cassette or direct breakout. Single MPO-96 trunk provides 6× 400G links per cable — making 400G port density achievable within standard cable tray fill ratios
- TIA-942-B compliance: Tier classification drives redundancy requirements — Tier 3 (N+1 power/cooling, concurrent maintainability) and Tier 4 (2N power/cooling, fault tolerance) require dual fiber paths between all switch pairs to separate physical pathways
- Physical layer testing: ISO/IEC 14763-3 / TIA-526-14B optical power testing and OTDR trace for every installed fiber link — providing baseline measurements for future fault diagnosis and warranty evidence
- Cable management in AI GPU halls: AI GPU cluster fabric (NVLink, InfiniBand) adds cabling density beyond Ethernet — cable tray fill calculations must account for both Ethernet fabric AND GPU interconnect cabling in the same overhead pathway
800G and 1.6T: Data Centres Being Designed Today Must Plan for Three Transceiver Generations
800G transceiver ports are entering production hyperscale data centres in 2025–2026, with 1.6T co-packaged optics following in 2027–2028. Both require OM5 or OS2 fiber — OM3 and OM4 are already at their distance limits for 400G parallel optic modules and cannot support 800G SWDM8. Data centres specified today with OM5 fiber and MPO-96 pre-terminated trunk infrastructure will accommodate 400G → 800G → 1.6T transceiver upgrades without recabling, while facilities that deployed OM3 in 2018–2020 face full fiber replacement ahead of 800G deployment — an infrastructure cost that OM5 specification at build time would have avoided entirely at a premium of less than 20% on the original fiber cost.