East-West Traffic and Why the Fabric Is Different

The interconnect that matters most for AI training lives not under the ocean but inside the building, wiring accelerators to one another. This intra-cluster fabric obeys different economics from the wide-area network, and understanding why is essential to seeing where scarcity forms inside the compute layer itself.

Start with the traffic pattern. Traditional data center workloads are dominated by north-south traffic, data flowing between users outside the building and servers inside it. AI training inverts this. During a large training run, the dominant traffic is east-west, flowing between accelerators inside the cluster as they exchange intermediate results at every step of the computation. A model spread across thousands of accelerators must, many times per second, synchronize state across all of them. The volume of this internal traffic dwarfs the external traffic, and its timing is unforgiving: because the accelerators compute in lockstep, the slowest link sets the pace for the entire cluster. A single congested path stalls thousands of expensive chips while they wait for data.

This lockstep dependency is what makes the fabric a hard constraint rather than a soft one. In a conventional network, congestion degrades performance gracefully; some requests are slower, but the system keeps working. In a training cluster, congestion on the fabric idles the accelerators, and idle accelerators are the most expensive stranded capital in the entire stack. The fabric therefore cannot be an afterthought sized to average demand. It must be provisioned for the peak synchronized burst, which means the ratio of network capacity to compute capacity inside an AI cluster is far higher than in any previous generation of data center.

Now the superlinear scaling. As a cluster grows, the challenge of keeping every accelerator able to communicate with every other one grows faster than the accelerator count. The number of possible pairwise connections rises with the square of the number of endpoints, and while real fabrics do not wire every pair directly, the amount of switching and optical interconnect needed to preserve high bandwidth between any two points scales in a way that outpaces linear growth. Double the cluster and you more than double the switches and transceivers required to keep it non-blocking. This scaling is the central fact of fabric economics, and it has a direct consequence: each generation of larger training systems demands a disproportionate increase in networking components, straining their supply chains ahead of, and harder than, the compute components themselves.

The physical components that carry this load are worth naming precisely, because each is a potential bottleneck. Switch silicon is the chip at the heart of each network switch, and its capacity, measured in total throughput, has become a headline specification that gates how large a cluster can be built with how many tiers of switching. Optical transceivers are the modules that convert electrical signals to light and back at each end of a fiber link; a large cluster needs enormous numbers of them, and their manufacture involves precision optics and assembly that cannot be scaled overnight. The fiber and cabling inside the building, though less glamorous, must be installed in vast quantities with exacting tolerances. A shortage in any one of these idles the accelerators just as surely as a shortage of the accelerators themselves.

There is a power dimension that ties the fabric back to the rest of the stack. Moving data between chips costs energy, and at cluster scale the power spent shuttling bits across the fabric is a meaningful and rising share of total cluster draw. A fabric that moves the same data for fewer joules directly relieves the power constraint that gates the whole build. This is why the fabric is not purely a bandwidth story; it is a power story in disguise, and the components that reduce energy per bit sit at the intersection of the two scarcest layers in the stack.

The investment implication is that the fabric supply chain is a distinct scarcity point that moves faster than the compute cycle and is less visible. Accelerator shortages are widely tracked and heavily reported. The transceiver and switch-silicon supply chains that must scale in superlinear proportion to those accelerators are followed by far fewer investors, yet they can gate cluster construction just as hard. My read: because the fabric scales superlinearly while attention scales linearly, the intra-cluster interconnect supply chain is structurally under-watched relative to its constraining power, and that gap between attention and impact is exactly where an informed analyst finds mispriced exposure. The next lesson traces this supply chain component by component and identifies where the durable choke points sit.