Training a frontier model doesn’t happen on one chip; it happens on thousands, lashed together so tightly they act as a single machine. The network that does the lashing is now a frontier of its own — and it scales with every accelerator added.
Scale-up versus scale-out
Two fabrics do the work. Scale-up connects the handful of accelerators inside a single server or rack at enormous bandwidth; scale-out connects thousands of those racks across a data center. Both have to grow in lockstep with compute, and each is its own supply chain with its own bottlenecks.
The optics frontier
Once traffic leaves the rack, copper runs out of reach and the signal moves to light. Optical transceivers step from 400G to 800G to 1.6T as clusters grow, and the industry is pushing toward co-packaged and linear optics and silicon photonics to keep power and cost in check. This is one of the fastest-scaling parts of the entire build-out.
Switch silicon and SerDes
The bits are steered by switch ASICs and moved on and off chips by SerDes — the high-speed lanes that decide how fast a chip can talk to the outside world. As lane rates climb, this silicon becomes as much a constraint as the accelerator itself.
The investable insight: optics scale with GPU count
Here is why networking matters to an investor: every accelerator needs multiple optical ports, so optics demand tracks the number of chips deployed rather than any single vendor’s roadmap. When cluster sizes jump, the optics and photonics suppliers ride the volume regardless of whose accelerator wins.
Who sits on it
The exposure runs through the switch-silicon and networking names, the transceiver makers and their laser, DSP and silicon-photonics suppliers. THE ENTITY groups them under the networking theme and maps each to the exact part of the fabric it supplies. Research, not advice.
Go deeper in the model
See it live on the map
Plans from $19/mo — cancel anytime. Two minutes to your first read.