Data center networking in 2026: switches, NICs, and the optical-transceiver bottleneck behind every AI cluster

Data center networking moved from a procurement footnote to a primary constraint in 2024 and 2025. AI training clusters at sixteen thousand GPUs and beyond run more network bandwidth between GPUs than they consume on storage and external traffic combined. 800G optical transceivers, NVIDIA Spectrum-X switching, and the next-generation NIC layer all became allocation-driven supply categories. Lean SupplAI was built to track this layer with the same depth as compute and memory, because in modern AI infrastructure, the network bottleneck arrives before the compute one.

For procurement teams scoping AI infrastructure or large data center builds in 2026, network sourcing is no longer a downstream decision. The transceiver allocation question often determines the overall build timeline.

Top-of-rack and spine switches

For Ethernet-based AI clusters: Arista Networks (7050X4 leaf, 7800R3 modular spine, 7060X6 800G), Cisco Nexus 9000 series, NVIDIA Spectrum-X (SN5600 800G platform with Spectrum-4 ASIC), and Juniper QFX/PTX. Below the brand layer, the underlying merchant silicon is concentrated at Broadcom (Tomahawk 5, Jericho 3-AI), NVIDIA (Spectrum-4), and Marvell (Teralynx 10). Most non-NVIDIA switch vendors ship Tomahawk-based systems, which makes Broadcom silicon a key sub-tier in the networking supply chain.

InfiniBand versus Ethernet

InfiniBand (NVIDIA Quantum-2 NDR 400G, Quantum-3 XDR 800G) remains the default for the largest AI training clusters because of latency and lossless behavior. Ethernet (with RoCEv2 and DCQCN, plus emerging Ultra Ethernet Consortium standards) is closing the gap and offers cost and ecosystem advantages. The procurement implication is that NVIDIA Spectrum-X versus InfiniBand is now a real sourcing decision, with cost and timing trade-offs that depend on cluster scale and workload mix.

NICs and DPUs

NVIDIA ConnectX-7 (400G) and ConnectX-8 (800G) dominate AI server NIC choice, with BlueField-3 DPUs adding offload for storage and networking. Intel and Marvell ship competing 400G/800G NICs (Intel E810, Marvell OCTEON 10). AMD Pensando holds the DPU position outside NVIDIA. For OCP-form-factor designs, OCP-NIC 3.0 is the dominant socket, with multiple vendors qualified.

Optical transceivers: the binding constraint

The 800G optical-transceiver market has been the most allocation-constrained networking category since 2024. The qualified vendors include Coherent (formed from II-VI / Finisar), Lumentum, Innolight (China), Eoptolink (China), and Hisense Broadband. Silicon photonics-based transceivers from Coherent and Lumentum have eased some of the constraint, but capacity remains tight for both QSFP-DD and OSFP form factors. For AI cluster procurement, transceiver allocation often determines build timing more than switch or GPU availability.

DAC and AOC cabling

For short-reach interconnect (under three meters), Direct Attach Copper (DAC) cables from Amphenol, Molex, TE Connectivity, and Coherent dominate. For mid-reach (three to thirty meters), Active Optical Cables (AOC) from the same vendors plus Foxconn ICT cover the volume. Programs running 800G should plan cable inventory alongside transceiver inventory, since cable assemblies are increasingly the long-lead item.

How Lean SupplAI tracks data center networking supply

Lean SupplAI indexes data center networking suppliers across switches, NICs, DPUs, optical transceivers, and cable assemblies, with merchant-silicon sub-tier mapping for the switch layer. For procurement teams scoping AI infrastructure or general data center builds, Lean SupplAI surfaces the full networking stack with allocation status visible inline.

What sets Lean SupplAI apart