Nvidia's networking business moves centerstage

Nvidia’s latest numbers put a hard figure on a part of the AI buildout that has often sat behind the GPU story. On May 20, the company said data-center networking revenue reached a record $14.8 billion in the quarter ended April 26, up 199% from a year earlier and 35% from the prior quarter. (nvidianews.nvidia.com) Nvidia reported total data-center revenue of $75.2 billion and said it would begin using a new reporting structure that separates hyperscale customers from a second bucket covering AI clouds, industrial customers and enterprises. That matters because AI clusters are no longer described by Nvidia as collections of chips alone. Jensen Huang, Nvidia’s chief executive, said on the earnings release that the “buildout of AI factories” was accelerating, while the company’s March product announcements described Rubin-era systems as combinations of GPUs, CPUs, NVLink switches, SuperNICs, DPUs and Ethernet switches operating as one machine. (nvidianews.nvidia.com) ### Why is networking suddenly getting this much attention? Nvidia’s own architecture announcements now place networking in the middle of the system, not at the edge of it. In March, the company said the Vera Rubin platform combines the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch for workloads ranging from pretraining to real-time agentic inference. (nvidianews.nvidia.com) The company also tied that hardware stack to a broader “AI factory” pitch. In a separate March release, Nvidia said its Vera Rubin DSX reference design spans compute, Spectrum-X Ethernet networking and storage, and is meant to guide the design, buildout and operation of large AI factories. (nvidianews.nvidia.com) ### How big is this business already? The $14.8 billion quarterly networking figure is large enough to stand on its own. Annualized, that run rate would be roughly $59.2 billion, based on Nvidia’s reported April-quarter number. That compares with Broadcom’s disclosure in March that its AI revenue rose 106% year over year to $8.4 billion in its fiscal first quarter, driven by “custom AI accelerators and AI networking,” according to Chief Executive Hock Tan. (nvidianews.nvidia.com) Broadcom’s figures are not a direct apples-to-apples comparison, because the company’s AI revenue includes accelerators as well as networking. But the comparison helps explain why Nvidia’s interconnect business is drawing more scrutiny from investors and suppliers: the networking line is no longer a rounding error inside the data-center segment. (nvidianews.nvidia.com) ### Where does Vera Rubin fit into the economics argument? Nvidia’s March announcements framed Rubin as a full-system platform aimed at lowering the cost of producing AI output. The company said the Vera Rubin DSX reference design is built to deliver “maximum token per watt” and faster time to production, while the Rubin platform release said the system was optimized for every phase of AI, including agentic inference. (nvidianews.nvidia.com) CNBC, citing Nvidia’s product disclosures, reported on May 20 that the Vera Rubin system delivers 10 times more performance per watt than its predecessor, Grace Blackwell. Nvidia’s own March materials also said advances in these integrated systems were improving cost efficiency and energy efficiency across training and inference workloads. (nvidianews.nvidia.com) ### What does this change for startups and the rest of the stack? Nvidia’s product language points toward opportunity in the layers around the chip. The DSX reference design names power, cooling, networking, storage, software and control systems as parts of the same buildout, and lists partners including Cadence, Siemens, Schneider Electric, Vertiv and Procore. (nvidianews.nvidia.com) That supplier list suggests the spending wave is spreading across infrastructure categories that determine utilization, latency and operating cost. For startups, that shifts some of the contest toward interconnects, inference efficiency, orchestration and AI-factory operations rather than only raw model or chip performance, an inference supported by Nvidia’s own emphasis on “codesigned” infrastructure and pod-scale deployment. (cnbc.com) ### What should readers watch next? Nvidia said on May 20 that it will report Data Center results under new Hyperscale and ACIE sub-markets going forward, a change that could give investors a clearer read on where networking-heavy AI factory demand is landing. Rubin systems are already in what Nvidia called full production, and the company said the Omniverse DSX Blueprint is generally available alongside the Vera Rubin DSX reference design. (nvidianews.nvidia.com 1) (nvidianews.nvidia.com 2) (nvidianews.nvidia.com 3)