What an AI Scale-Out Fabric Actually Carries
Typical AI clusters use a leaf-and-spine Ethernet architecture. Top-of-rack switches connect to spine switches; every GPU-to-GPU conversation crosses this fabric. At 400G, the workhorse is the QSFP-DD transceiver — eight 50 Gb/s lanes of high-speed traffic bridging compute and network. Scale a cluster from hundreds to tens of thousands of GPUs, and you scale transceivers the same way.
Leaf-and-Spine, Simplified
GPU ↔ GPU · East-WestEast-West, Not North-South
Distributed training is dominated by GPUs talking to GPUs, not clients to servers. Every training step ends with a collective operation (all-reduce) that fans gradients across the whole cluster.
The Fabric Carries Synchronization
It's not just data — the fabric carries the synchronization that keeps thousands of accelerators marching together. A fabric that can't keep up doesn't slow one link; it stalls the entire job.
Tens of Thousands of Optics
Every port needs a module. Hundreds of GPUs become thousands of optics quickly, and every module choice is multiplied across the cluster.
Why the Physical Layer Decides the Outcome
At 400G, the QSFP-DD form factor is where this plays out. Each module carries eight lanes running up to 50 Gb/s each — doubling the density of QSFP28 — and the MSA spec packs up to 36 ports per switch slot for 14.4 Tb/s of fabric capacity in a single slot. The right module choice is a reach-and-media decision as much as a speed decision.
Bandwidth Is the Training Bottleneck
If the fabric can't move data as fast as GPUs compute, utilization drops and training time stretches. Optics help determine fabric speed.
Reliability Compounds
One bad link in a large fabric can stall a job across thousands of GPUs. Transceiver quality and interoperability are uptime decisions.
Cost Scales Linearly
Every port needs an optic. Optical strategy is a direct cost-per-port decision that multiplies across the cluster.
Choose the module by how the fabric is physically built: SR8 for short multimode runs inside the rack, DR4 for 500 m single-mode, LR4 for up to 10 km between buildings. The right optic is not the fastest one — it's the one aligned to rack distance, medium, and cluster design.
Cheap Optics Do Not Fail Cheaply.
Here's the part you can't see: network instability is one of the primary causes of low GPU utilization in large-scale AI training. A link that flaps — drops, re-negotiates, drops again — interrupts distributed communication and forces portions of the training process to be recalculated. In an all-reduce, one flapping link means the whole collective stalls until every GPU has acknowledged.
Failure Signatures at Scale
- Link flaps that interrupt collective communication and trigger re-computation across thousands of GPUs.
- Marginal transmit power or temperature drift that surfaces only under sustained load.
- FEC error storms that turn a healthy fabric into a troubleshooting exercise.
- Interoperability failures across switch, optic, and server vendors that take weeks of finger-pointing to isolate.
- Idle accelerators sitting on the training critical path while the job recovers link by link.
The economics are brutal. A "savings" at the procurement desk becomes a very public training stall at the worst possible time. On our validated fabric, we've logged over 1 million cumulative port-hours with zero link flaps observed — and our testing has since passed 4 million link flap-free hours. That's not a marketing difference. That's the difference between a cluster that trains predictably and one where the network is a roulette wheel.
What a Validated Optics Strategy Actually Buys You
An open Ethernet scale-out fabric gives customers what a proprietary stack can't: freedom to choose their optics. In the Open AI Factory, the fabric is open, switching is open, and transceivers are multi-vendor validated — no single supplier controls the bandwidth roadmap.
Preserve Freedom of Choice
Open Ethernet fabrics let customers choose optics without handing the bandwidth roadmap to a single proprietary stack.
Compress Time to Deployment
Validated optics shorten the integration phase that otherwise keeps expensive accelerators waiting in the rack.
Control Quality at the Source
Factory ownership matters because testing, yield, and delivery timelines shape how confidently large clusters can scale.
400G Is a Reach-and-Media Decision
The right module choice matches the fabric's physical build. Each 400G QSFP-DD variant is optimized for a different span, medium, and cost profile — pick the one aligned to your topology and you avoid both overspending and stranded ports.
The Open AI Factory Owns Its Stack.
Openness isn't just about procurement freedom — it's about operational control. Micas co-owns and operates its own factory in Malaysia, one of two manufacturing sites in Asia, running clean-room facilities with advanced automation. That means we control testing, quality, and delivery — we're not at the mercy of someone else's allocation schedule when your cluster scales.
Open Switching
SONiC-based switching keeps the fabric adaptable and prevents customers from getting boxed into one supplier's roadmap.
Vendor Validation
Modules are qualified against compute and switching layers so deployment doesn't become a field-integration project.
Factory Control
Manufacturing ownership supports testing discipline, quality assurance, and delivery confidence when scale-out demand spikes.
AI-Factory Logic
The network is productive infrastructure, not plumbing — because the fabric directly shapes how much GPU capital gets utilized.
Optech's 400G QSFP-DD transceivers are built for exactly this environment: high-performance optical modules for AI scale-out, data centers, and telecom. In the Open AI Factory they're validated against the fabric and compute layers so they work the day they arrive.
The Fabric Is the Connective Tissue
of an AI Factory.
Open, validated optics at the scale-out layer mean more performance, more uptime, and more control over how your AI infrastructure scales. When the choice of optic is open — and validated before it lands in your rack — the cluster scales the way it should: predictably.
Meet us at booth #316.
Come see the Open AI Factory fabric live, or book time with us at the show.