Nvidia Quietly Builds Networking Behemoth To Rival Chips

Nvidia’s fastest-rising business isn’t its famed GPUs. It’s the vast, intricate web of hardware and software that moves bits between them. In just a few years, Nvidia’s data center networking unit has become a multibillion-dollar engine, reporting roughly $11 billion in quarterly revenue and more than $31 billion for the full year, according to the company’s latest results. That growth, up 267% year over year last quarter, is transforming networking from a supporting act into a strategic pillar that could rival the chip business itself.

A Second Engine Inside Nvidia: Networking as a Growth Driver

AI compute hogs attention, but the interconnect defines the ceiling for performance. Training and serving large models at scale require predictable bandwidth, low latency, and near-lossless fabrics. Nvidia’s lineup—NVLink for GPU-to-GPU communication, InfiniBand switches and adapters with in-network compute, Spectrum-X for Ethernet-based AI fabrics, and co-packaged optics to push signaling closer to silicon—has cohered into a full-stack platform purpose-built for what the company calls “AI factories.”

The payoff is tangible: networking is now the company’s second-largest revenue contributor behind compute. As clusters grow from thousands to tens of thousands of accelerators, fabric performance determines how efficiently those accelerators run. Even single-digit improvements in tail latency or congestion management translate into meaningful cost savings at hyperscale.

From Mellanox to AI Factories: Building the Foundation

The foundation was laid with Nvidia’s $7 billion acquisition of Mellanox in 2020. Mellanox pioneered high-performance InfiniBand and RDMA technologies widely used in supercomputing. Nvidia integrated those assets with its GPUs, systems, and software, extending into data processing units (BlueField DPUs) and the DOCA software framework to offload networking, storage, and security from CPUs. The result: a tightly coupled compute-and-network stack tuned for AI workloads rather than general-purpose data center traffic.

Crucially, Nvidia sells this as a system. Instead of piecemeal components, it offers end-to-end designs through partners like HPE, Dell, Lenovo, Inspur, and Supermicro. That approach lets customers procure a validated fabric that snaps into DGX, HGX, and cloud-scale deployments, reducing integration risk and accelerating time to production.

InfiniBand And Ethernet Both Accelerating

Nvidia is pushing on two fronts. InfiniBand remains the flagship for large-scale AI training, prized for RDMA performance, adaptive routing, and in-network computing that can reduce collective communication overhead. Major cloud and research operators have adopted InfiniBand for AI supercomputers, citing determinism and efficiency at scale.

At the same time, Spectrum-X targets enterprises and cloud providers standardizing on Ethernet. By combining Ethernet switches, smart NICs, traffic shaping, and telemetry tuned for AI workloads, Nvidia aims to deliver near-InfiniBand behavior on widely deployed Ethernet fabrics. Industry analysts at firms such as Dell’Oro Group and 650 Group project rapid growth in AI-optimized fabrics across both InfiniBand and Ethernet, with Ethernet capturing a rising share as AI inference spreads beyond hyperscalers.

GTC Signals a New Networking Phase for AI Systems

At its annual GTC conference, Nvidia expanded the roadmap again. The Rubin platform introduces six new chips designed to drive the next wave of AI supercomputers, alongside an Inference Context Memory Storage platform aimed at reducing memory thrash and improving service-level efficiency. New Spectrum-X Ethernet Photonics switches underscore a broader shift to optical integration, where co-packaged optics can cut power per bit and extend lane speeds as electrical signaling nears practical limits.

The intent is clear: address the hardest bottlenecks in multi-tenant, multi-rack AI clusters—tail latency, congestion collapse, and unpredictable flows—through silicon, optics, and software co-design. That mirrors the strategy Nvidia applied to GPUs over the last decade, where performance gains increasingly came from platform-level engineering rather than a single chip breakthrough.

Why This Could Rival the Chip Business in Scale and Spend

Networking scales with every accelerator sold, every model checkpoint synchronized, and every inference request routed. As clusters add 800G and move toward 1.6T lanes, the bill of materials for switches, optics, adapters, and network processors balloons. Bernstein and other research firms have noted that networking is taking a larger slice of AI system spend, sometimes approaching the cost of compute itself in the largest builds. For providers operating at cloud scale, shaving even a few watts per port or a few microseconds per hop compounds into millions saved.

Customers are reinforcing this trend. Hyperscalers building GPU-rich clusters need high-radix, high-throughput fabrics; enterprises deploying retrieval-augmented generation or vector search want predictable latency on Ethernet. Both buyer groups prioritize validated architectures over best-effort integration, which plays directly to Nvidia’s full-stack stance.

A Crowded but Advantageous Arena for AI Networking

Competition is intensifying. Broadcom, Marvell, and Cisco are advancing 51.2T and 102.4T Ethernet switch silicon and AI-centric features; Arista and other system vendors are crafting Ethernet architectures for AI fabrics; AMD’s Pensando and Intel’s IPU initiatives target offload and data plane acceleration. Nvidia’s edge is architectural control across GPUs, interconnects, DPUs, switches, optics, and software—plus a partner-led channel that can deliver complete, supported systems.

The risk is execution at extreme scale: supply of optics and substrates, power and cooling constraints, and the complexity of managing lossless fabrics across thousands of nodes. But if Nvidia repeats its GPU playbook—tight integration, aggressive roadmaps, and developer tooling—its networking arm is positioned to capture a growing share of AI infrastructure spending.

The Bottom Line: Networking Now Anchors Nvidia’s AI Push

Nvidia’s networking business has quietly become core to its AI story. With explosive revenue growth, a differentiated stack spanning InfiniBand and Ethernet, and fresh silicon and photonics unveiled at GTC, the company is building a second franchise that doesn’t just complement its chips—it amplifies their value. For the next generation of AI factories, the network is the backplane, and Nvidia intends to own it.