Not every AI infrastructure story is a hyperscaler story. Most enterprises, cloud providers, and regional operators are building something more practical: high-bandwidth networks with real performance — without hyperscaler-scale budgets.
That’s the world of the spine–leaf buildout — one of the most important optical-transceiver use cases in the AI infrastructure market. The links between spine, leaf, and top-of-rack (ToR) layers are where optics live, and for most builds 400G short-reach optics are the sweet spot: high bandwidth, high port density, and cost that scales.
|
25.6T
|
500 m
|
4×100G
|
51.2T
|
| Fabric capacity of one 64×400G spine switch | DR4 single-mode reach — the spine–leaf sweet spot | Breakout per QSFP-DD port | 128×400G spine capacity / CPO switch |
01 — The Bottleneck
East–West Traffic Is the New Bottleneck
AI training and inference move data sideways. Every accelerator talks to every other accelerator — gradient syncs, all-reduce, embedding lookups — and that east–west traffic now dominates the fabric. The network must keep up with compute, or GPU utilization pays the price.
In large training clusters, the #1 cause of wasted GPU compute isn’t the model — it’s network instability and link flaps that interrupt distributed communication and force costly re-computation.
Legacy three-tier and heavily oversubscribed designs were built for north–south, client-to-server traffic. They choke on AI’s east–west patterns. That’s why the spine–leaf Clos topology — every leaf connected to every spine, no single point of congestion — has become the standard for AI-ready data centers.
For most organizations, the sweet spot on that fabric is 400G short-reach optics: enough bandwidth to keep accelerators fed, dense enough to fit the rack, and priced for real budgets.
02 — For AI-Infra Teams
What 400G Spine–Leaf Actually Changes
Cost per Bit
Short-reach 400G delivers the bandwidth you need without paying for long-reach capability you’ll never use. DR4 runs 500 m on single-mode fiber — more than enough for spine-to-leaf and leaf-to-ToR links inside a building — at a fraction of the cost of 2 km FR4 or 10 km LR4 modules. You buy the reach you use.
Port Density
400G QSFP-DD packs eight 50G PAM4 lanes into one hot-pluggable module. That density rewrites rack economics: a 64×400G spine carries 25.6T in 4 RU, and a 128×400G spine reaches 51.2T. More bandwidth per port means fewer switches, less power, less cabling.
Breakout Flexibility
One 400G port breaks out to 4×100G or 2×200G, so a single switch generation serves servers, storage, and fabric uplinks. Migrate rack by rack, not all at once — the same QSFP-DD port carries 400G today and sits on the roadmap to 800G tomorrow.
Predictability
The last thing a buildout needs is an optic that doesn’t interoperate with the switch it’s plugged into. Validated transceivers mean no surprises at turn-up — no link flaps, no vendor finger-pointing, no re-cabling in the middle of a deployment window.
The proof: our industry-first 51.2T co-packaged optics switch — now in volume production. Built with Broadcom’s Tomahawk 5 and manufactured in our co-owned Malaysia factory, it shows where the 400G curve bends toward 800G:
- ~38% lower switch power vs. pluggable designs (validated Supermicro + AMD build)
- 4M+ cumulative link flap-free hours
- ~3% scale-out overhead vs. native throughput
03 — The Method
The Open AI Factory Approach
We treat optics as a strategic part of the architecture, not an afterthought. It’s why the “factory” in our name is literal:
Open Switching Layer
Broadcom-powered Ethernet fabric that works with open NOS environments — ONIE-loaded switches running community or Enterprise SONiC, so you own the software stack.
Validated Before It Ships
Transceivers are validated against the switching and compute layers before shipping. Optech 400G SR8 and DR4 QSFP-DD modules are qualified in that loop — tested against the fabric so they deploy without surprises.
Multi-Vendor Optics
Customers keep choice: multi-vendor optics mean no single supplier owns the speed roadmap.
Ready Racks
Lenovo infrastructure designed to carry dense, high-bandwidth builds — compute, storage, and fabric that arrive racked, cabled, and ready to light up.
The 400G Optics Reference
The reach you need determines the module you buy. For a spine–leaf buildout, the action is almost all in the first three rows:
| Module | Reach | Fiber | Connector | Best for |
|---|---|---|---|---|
| 400G SR8 | 100 m (OM4) | Multimode | MPO-16 | ToR–leaf, intra-rack |
| 400G SR4 | 50 m (OM4) | Multimode | MPO-12 | Leaf–ToR short hops |
| 400G DR4 | 500 m | Single-mode | MPO-12 | Spine–leaf, leaf–ToR |
| 400G FR4 | 2 km | Single-mode | Duplex LC | Building interconnect |
We’re at booth #316 with Broadcom showcasing our Open AI Factory story. Planning a 400G spine–leaf build? Have it IRL.
Density Is the New Economics
You don’t need a billion-dollar cluster to benefit from modern AI networking. Spine–leaf buildouts with 400G optics deliver the bandwidth, density, and cost efficiency most organizations need — deployable on your terms, on an open fabric, with validated optics that just work.