Part of $NVIDIA (NVDA.US)$ 's moat & value-add is its contro...

Part of $NVIDIA(NVDA.US)$ 's moat & value-add is its control of the networking for AI data center.

And $Broadcom(AVGO.US)$ remains the only alternative chip vendor to it, while $Arista Networks(ANET.US)$ served as the end OEM.

$Meta Platforms(META.US)$ 's latest two H100 DCs use two different networking infra, one from $NVIDIA(NVDA.US)$ 's full stack including Mellanox switch while another uses $Arista Networks(ANET.US)$ switch powered by $Broadcom(AVGO.US)$ chip, and $Cisco(CSCO.US)$ switch with its own Silicon One chip

Interesting that Meta said the two have the same performance. This is surprising given $NVIDIA(NVDA.US)$ always prefer and fine tune the system for its full stack DC.

The incoming Blackwell GPU is going to be another challenge for $Broadcom(AVGO.US)$ as it needs to be brought in 880G Ethernet RoCE in mass deployment too. And $NVIDIA(NVDA.US)$ 's latest chip seems to be prioritizing NVlink + Infiniband NVDA's proprietary interconnect bandwidth vs. PCI-E.

So potentially by limiting the speed of PCI-E + RoCE on the chip I/O interface level, Nvidia can make it less appealing for players to go for an open DC architecture. This remains a threat to Broadcom.

Obviously Meta will continue to push the open architecture but so far it seems to be the minority.