Broadcom is challenging Nvidia's dominance in AI networking with custom switching silicon, competing for market share in interconnect and Tomahawk-class infrastructure.

Broadcom's chief European architect claimed the firm stands ahead of Nvidia and others in addressing AI bottlenecks in scale-out networks, citing Ethernet as its competitive advantage.

In a presentation at the Xcelerated Compute show, Kamran Naqvi, Broadcom's chief network architect for EMEA, touted the company's Thor Ultra network interface card (NIC) as "the world's first 800 gigabit Ethernet NIC." He acknowledged Nvidia's rival offering, the ConnectX-8 SuperNIC, noting that it "only supports 800G on InfiniBand. For Ethernet, it has to be four times 400G."

"I'm sure Nvidia will catch up. They're a fantastic company. At the moment, we have a lead. Why not brag about it?" Naqvi said.

Broadcom has emerged as a strong backer of Ethernet alongside AMD and OpenAI in the AI infrastructure market. OpenAI architects designed the Multipath Reliable Connection (MRC) protocol as an Ethernet-based transport standard for AI networks, with Broadcom contributing to its development. Naqvi positioned MRC and the Ultra Ethernet Transport (UET) specification as industry standards for AI scale-up networking. He emphasized Broadcom's role as a founding member, citing the company's belief "that innovation happens when industry aligns on interoperable open standards."

Both MRC and UET offer packet multipath capabilities that enable efficient load balancing across AI workloads through packet spray workflows.

On the product side, Naqvi highlighted the Tomahawk Ultra Ethernet switch chip as "very suited" to AI scale-up environments, designed from the ground up with ultra-low latency. He noted that the Tomahawk 6 chip is already being deployed by "many clients" for scale-up purposes. He also referenced an unnamed German customer currently running a proof of concept with a Broadcom partner to build infrastructure supporting 28,000 GPUs across four clusters as part of its scale-out efforts.

Naqvi stressed that optics will play an integral role in future AI networks. While advocating for optical solutions, he recommended copper where practical, echoing Nvidia CEO Jensen Huang's prediction that copper will continue to dominate for "several years." "It's reliable. It's cheap. It has a negative power," Naqvi said.

Broadcom has invested heavily in co-packaged optics (CPO) technology. Meta's deployment, which achieved one million hours of flap-free operation in 2025, has now reached 50 million hours of flap-free operation. "So it's very reliable. It saves you up to 65% to 70% of power utilization," Naqvi noted. This advancement addresses Meta's critical AI infrastructure challenge: GPU idle time while waiting for network completion.

Naqvi illustrated the stakes: "You build this billion-dollar infrastructure, and networking is only the 10% of the cost of building an AI infrastructure. But if your network is not done right, your network becomes the bottleneck, and it costs you a lot more."