Networking & Optical Interconnects: The Constraint After Compute
AI clusters are only as fast as the connections between their accelerators. Once compute and memory bottlenecks are addressed, networking becomes the next gating constraint — and the build-out of next-generation interconnects, both electrical and optical, becomes a structural investment theme. NVIDIA’s networking stack (acquired via Mellanox), Broadcom’s switching silicon, Marvell’s interconnect IP, and the optical transceiver ecosystem (Lumentum, Coherent, Fabrinet) all benefit from the same physical reality: at cluster scale, data movement is more expensive than data processing.
Why Networking Becomes the Bottleneck
Training a frontier model requires synchronizing weights across tens of thousands of GPUs. All-reduce operations dominate training time as cluster size grows. A 1-millisecond latency in cross-cluster communication compounds across millions of operations per training step. Bandwidth, not compute, increasingly limits effective cluster utilization. Inference at scale faces the same dynamic — agentic workloads with multi-step reasoning chains generate continuous inter-node traffic.
The Stack
- Switching silicon: Broadcom (Tomahawk, Jericho), Marvell (Teralynx). The chips that route packets between racks.
- Optical transceivers: Lumentum, Coherent, Fabrinet, InnoLight, Eoptolink. Convert electrical signals to optical for fiber transmission.
- Networking systems: Arista, Cisco, NVIDIA Spectrum, Juniper. Assembled switches at rack and cluster level.
- Cables and connectors: TE Connectivity, Amphenol. The physical layer often overlooked until shortages emerge.
The Optical Transition
Copper interconnects work at short distances and lower bandwidth. As GPU clusters scale and rack power densities climb, optical takes over — first inter-rack, eventually intra-rack via co-packaged optics. The transition opens revenue lines for pure-play optical names (LITE, COHR, FN) that did not exist at scale before AI. The dollar content per port in optical is multiples higher than in copper, and the ramp is just beginning.
Co-Packaged Optics (CPO)
The next architectural shift brings the optical engine onto the same package as the switch chip — eliminating long electrical traces and dramatically improving bandwidth and power efficiency. Broadcom and Marvell are pushing CPO designs into production for 2026–27 deployment. CPO is the pivot that determines which players win the next networking cycle. It also extends the optical TAM into territory copper cannot reach.
Risk Factors
Customer concentration: NVIDIA, Google, Meta, Microsoft account for the majority of demand. Technology transitions can leapfrog incumbents (a CPO failure is a Broadcom opportunity, and vice versa). Inventory cycles in optical have historically been brutal — 2022–23 saw a sharp pullback before the AI ramp restarted the cycle. The names with the cleanest secular setup are also the names with the highest historical drawdown volatility.