The Interconnect Empire: How Nvidia's Networking Pivot Redefines Infrastructure

For years, the industry viewed Nvidia through the lens of the graphics processing unit (GPU). But as The Register first reported, a closer look at the company's recent financial trajectory reveals a more aggressive evolution. Nvidia is no longer just a chipmaker; it has transformed into a networking hegemon, positioning itself as a system builder in a way that threatens the traditional order of data center infrastructure.

According to reporting from The Register, the scale of this shift is staggering. In the second quarter of 2026, Nvidia's networking revenues reached $17.5 billion. To provide a sense of scale, Cisco's total revenue across its entire product portfolio during that same period was $17.3 billion. This ascent was not an accident of the AI boom, but the result of a strategic pivot initiated in 2019 when CEO Jensen Huang executed a $6.9 billion acquisition of Mellanox.

At the time of the merger, Mellanox was primarily recognized for its ConnectX NICs and InfiniBand switches, though it lacked a dominant presence in Ethernet switching. InfiniBand was largely confined to supercomputing clusters and high-performance computing (HPC), with a notable exception being a large-scale deployment at Microsoft. However, the acquisition allowed Nvidia to move beyond selling discrete components. As The Register notes, Nvidia recognized that its primary customers were not looking for individual GPUs or servers, but rather entire clusters of systems. To deliver these, Nvidia required a robust interconnect solution, and Mellanox's InfiniBand provided the necessary foundation.

This strategy is rooted in what Huang calls "co-design." Rather than attempting to build the single best GPU or the single best switch in isolation, Nvidia designs components and software to function as a unified whole for specific workloads. Sameh Boujelbene, an analyst at Dell'Oro Group, told The Register that this integration is the primary driver of Nvidia's success, noting that customers choose the company specifically for a tightly integrated compute and networking architecture optimized for AI.

The practical implications of this consolidation are evident in the experience of Recursion, an AI and machine learning company founded in 2013. Ben Mabey, the CTO of Recursion, explained to The Register that the company's early efforts to apply machine learning to drug discovery involved massive datasets—now exceeding 50 petabytes—that far surpass the data requirements of typical Large Language Model (LLM) training.

Recursion's initial cluster, BioHive Zero, was a makeshift system utilizing GTX 1080 graphics cards intended for gaming, connected via Ethernet. Mabey reports that this setup quickly hit the limits of conventional networking. For its subsequent cluster—an A100-based system featuring approximately 300 accelerators—Recursion pivoted to Nvidia's ConnectX NICs and Quantum InfiniBand switches. Mabey noted that sourcing both compute and interconnects from a single vendor simplified troubleshooting and optimization. When Recursion sought help tuning the system, Nvidia treated the deployment as a single integrated system rather than a collection of discrete networking and compute resources. This level of holistic integration, Mabey suggests, was previously the sole domain of boutique supercomputer builders such as Atos or Cray.

While Nvidia's networking growth has been steady since 2019, The Register identifies two specific inflection points where revenue surged. The first occurred in the 2023 calendar year (FY24), triggered by the launch of ChatGPT and the subsequent AI gold rush. The second occurred across 2024 and 2025, coinciding with the shipment of Blackwell-generation GPUs in large quantities.

Boujelbene credits this second surge to the accelerated use of Spectrum-X Ethernet switches for AI back-end networks and the rollout of NVLink switches. This growth has been fueled by the increasing deployment of Nvidia GPU clusters and higher attach rates for Nvidia's cables, optics, NICs, and switches.

This shift toward Ethernet has been adopted by some of the world's largest AI projects. The Register reports that Elon Musk's Colossus supercomputer—a cluster based on 200,000 H100 and H200 chips—was one of the earliest and largest deployments to utilize Spectrum-X Ethernet switches when its first half came online in 2024. Similarly, Meta has utilized Nvidia's Spectrum switches extensively in its training clusters, including the Prometheus site in New Albany, Ohio, which features one million GPUs.

By controlling both the compute (the GPU) and the fabric that connects them (the networking), Nvidia has effectively rewritten the rules of infrastructure. The company is no longer selling parts; it is selling a proprietary, optimized ecosystem. For the next generation of gaming and AI infrastructure, the blueprint is clear: consolidation is the path to performance.