A Vendor Shift IDC Calls Historic
IDC does not often describe a market change as one of the most significant it has ever tracked, but that is the language the firm used for Nvidia's ascent in data center Ethernet switching. In the first quarter of 2026, Nvidia took the number one spot by revenue, a position it simply did not occupy a year earlier. The company that built its empire selling accelerators has now become the largest seller of the switches that knit those accelerators together.
The numbers explain the superlatives. Nvidia's data center switching revenue jumped 192.7 percent year over year to 2.1 billion dollars in the quarter, lifting its share of the segment to 21.5 percent. Growing a near-standing-start business to market leadership in roughly twelve months is the kind of move that usually takes a decade, and it reflects how tightly AI infrastructure spending has compressed the normal cadence of enterprise networking.
Spectrum-X Is the Engine
The driver is Spectrum-X, Nvidia's networking platform built specifically for large AI systems. It is not a single switch but an end-to-end fabric that combines Spectrum Ethernet switches with BlueField DPUs and LinkX cables, engineered for the traffic patterns of GPU clusters running training jobs across thousands of accelerators. IDC cited significant traction from hyperscalers and cloud providers standing up AI factories, which is precisely the customer segment driving the broader build-out.
What makes Spectrum-X strategically potent is that it travels with the GPUs. When a hyperscaler commits to Nvidia Blackwell or Rubin systems, the path of least resistance is to buy the matching network from the same vendor, validated as a reference architecture and supported as one stack. We see this as the quiet mechanism behind the share gain: Nvidia is not winning competitive bake-offs switch by switch so much as it is bundling the network into the compute decision that was already made.
The Whole Market Is Inflating
Nvidia's rise is happening against a backdrop of extraordinary growth across the entire category. The worldwide Ethernet switch market reached 15.4 billion dollars in the quarter, up nearly 40 percent. The data center portion grew 61 percent year over year to 10 billion dollars, and 800 gigabit switches alone accounted for 35.8 percent of data center segment revenue. The fastest, most expensive gear is now the center of gravity, not the edge.
That matters because it tells us the AI build-out is not just buying more switches, it is buying dramatically faster ones. Every generation of GPU demands proportionally more network bandwidth to keep the cluster fed, so the switching market is being pulled upward by the same forces inflating GPU budgets. For finance leaders modeling AI capex, the network is no longer a rounding error against the silicon: it is a fast-growing, fast-obsolescing line of its own.
A Networking Business Bigger Than Most Pure-Plays
It is easy to lose sight of how large Nvidia's networking business has become because it sits in the shadow of the GPU franchise. On an annualized basis the networking division is now running at tens of billions of dollars, a scale that already rivals or exceeds the entire revenue of most standalone networking vendors. The company that the market still files under semiconductors now sells more data center Ethernet switching by revenue than anyone else, and it got there without buying an incumbent switch vendor, building the position organically off its Mellanox foundation.
That changes how we should think about the company's moat. For years the bear case on Nvidia centered on the eventual commoditization of accelerators as custom silicon from hyperscalers matured. The networking ascent complicates that thesis: even buyers who diversify their compute with in-house chips still need a high-performance fabric to connect it, and Nvidia has positioned Spectrum-X to capture that spend regardless of whose accelerators fill the racks. The fabric is becoming a second, durable revenue pillar that is less exposed to the custom-silicon threat than the GPU business itself.
The Power and Density Subtext
There is a physical story underneath the revenue numbers. The reason 800 gigabit switches already account for 35.8 percent of data center segment revenue is that AI clusters are packing accelerators ever more densely, and dense racks demand both faster interconnect and more power per square foot. The network is part of the same thermal and electrical budget as the GPUs, and as operators push toward higher-radix, higher-bandwidth fabrics, the switching gear itself becomes a meaningful contributor to a data center's power draw and cooling load.
For infrastructure leaders, that ties the networking decision directly to the grid constraints now dominating site planning. A faster, hotter fabric can shift a campus from air to liquid cooling and change its power profile, which in turn affects where it can feasibly be built. We would push teams to model the network not as an afterthought to the compute design but as part of the power and cooling envelope from the start, because in a build-out where megawatts are the binding constraint, every watt the switching layer consumes is a watt unavailable to the accelerators it serves.
Where Cisco and Arista Stand Now
The incumbents are not displaced, but they are recontextualized. Arista remained second in data center switching, the position it has long defended with hyperscale customers, and Cisco retained leadership in the overall Ethernet switch market when enterprise and campus segments are included. Broadcom looms as a competitor with its own merchant silicon ambitions. The point is that Nvidia did not erase these vendors, it leapfrogged them in the single segment where AI dollars are concentrated.
For Arista in particular this is an uncomfortable adjacency. Its differentiation has always been best-of-breed switching decoupled from any one compute vendor, and Nvidia's pitch is the opposite: a vertically integrated fabric tuned for its own GPUs. The competitive question for the next several quarters is whether large AI operators value the openness of a multi-vendor network enough to break the Nvidia bundle, or whether the convenience of a single validated stack wins out.
What This Means for the Buyer
For CIOs and infrastructure architects, the practical concern is concentration. When the same vendor supplies the accelerators, the network fabric, the DPUs, and increasingly the systems integration, negotiating leverage erodes and lock-in deepens. A reference architecture that ships as one stack is operationally easier, but it also makes it harder to second-source any layer when pricing or supply tightens. That trade-off deserves an explicit decision rather than a default.
We would advise treating the network as a deliberate architectural choice even inside an Nvidia-heavy estate, benchmarking Spectrum-X against open alternatives and the emerging Ultra Ethernet ecosystem before committing the entire fabric. The performance case for tightly coupled AI networking is real, but so is the strategic cost of handing one supplier the GPU, the switch, and the support contract. The vendors that survive this shift will be the ones that prove an open network can keep pace with a closed one.
