The conversation surrounding AI infrastructure is often dominated by GPUs. While GPUs are crucial, it’s the underlying interconnect fabrics that enable their full potential, facilitating the training and execution of models with trillions of parameters. These interconnects operate on multiple levels, from communication within the chip, across systems, and between massive data centers scaling to accommodate thousands of accelerators.

Nvidia’s dominance in this arena is evident, but Broadcom has been quietly yet steadily crafting technologies that address the entire spectrum of interconnect needs—from expansive Ethernet networks to intra-package communications. Broadcom’s approach is unique among major players, utilizing a merchant silicon model that licenses its IP and sells its chips across the industry, allowing other companies to leverage its technology without overtly advertising Broadcom’s involvement. This includes collaborations with industry giants like Google and potentially Apple.

Broadcom is also gaining traction in the hyperscale sector, providing valuable expertise and solutions that relieve companies from developing these complex interconnects internally. Its roots in Ethernet switching are serving it well. Despite potential negative perceptions stemming from the VMware acquisition, Broadcom’s leadership in Ethernet technology is undeniable. High-performance computing clusters require immense switching capability, and Broadcom’s products are at the forefront, offering innovative solutions like the Tomahawk 6 with a staggering switch capacity of up to 102.4 Tbps.

In terms of scale, a cluster equipped with 128,000 GPUs could necessitate over 5,000 switches dedicated solely to compute processes, with additional networking for storage and management. Broadcom’s latest offerings reduce these scaling challenges significantly. Furthermore, Broadcom is looking to establish itself in the domain of scale-up Ethernet to enhance chip-to-chip connectivity for dense computing environments, positioning itself against Nvidia’s NVLink system.

Broadcom’s foresight extends into the realm of co-packaged optics (CPO), a concept introduced with the Humboldt switches: integrating lasers and digital processors directly on the switch package reduces energy consumption by eliminating separate connectors. The next iterations of the Tomahawk line aim to exploit this by embedding optics for direct fiber connections, which significantly enhances efficiency and speed.

Finally, Broadcom’s innovations extend to chiplet architecture—building on Moore’s Law by enabling multi-die systems. With cutting-edge 3.5D packaging technologies such as hybrid copper bonding, Broadcom is paving the way for ultra-dense, efficient chip designs that promise unprecedented interconnect speeds and reduced signal pathways. The first products employing these advanced techniques are expected to reach production as early as 2026, promising to reshape the AI hardware landscape.