NVIDIA has moved from selling the chips inside AI data centers to owning more of the fabric around them, and Cisco and Arista now have a harder problem than a lost quarter of share.
For decades, if you were wiring up a serious data center, you called Cisco or Arista. That is no longer the opening assumption. According to IDC figures reported by Business Insider this week, NVIDIA became the top vendor by revenue in data center Ethernet switching in Q1 2026, with $2.1 billion in revenue and 21.5% of the market. Arista was close behind at 20.7%. Cisco, still the name most people associate with enterprise networking, was behind both in this specific data center segment.
That is a sharp move, not a gentle climb. IDC senior research manager Brandon Butler told Business Insider that NVIDIA had less than 4% of the data center Ethernet switch market in Q1 2024. Two years later, it is sitting at the top of a market IDC put at $10 billion for the quarter, up 61% from a year earlier. When a company moves that far that fast, you should assume something structural has changed.
The change is AI infrastructure. Big GPU clusters don't behave like ordinary enterprise networks, and NVIDIA knows exactly where the pain sits because it sells the GPUs creating it. Spectrum-X is not just another switch line. NVIDIA sells it as a network fabric for AI systems, combining Spectrum Ethernet switches, BlueField DPUs, ConnectX networking, software and cables into one stack. The company says Spectrum-X can deliver about 1.6 times higher AI performance than standard Ethernet setups for certain workloads.
That claim is doing real commercial work. If you're building a large AI cluster, the cost of the switch is not the main question. The expensive failure is buying thousands of GPUs and then leaving them waiting because the network can't feed them properly. NVIDIA's pitch is brutally simple: buy the fabric from the same company that built the compute layer, and you reduce the number of places where the system can break.
Frankly, that is why Cisco and Arista should worry. They are not suddenly bad networking companies. Arista still has deep hyperscaler relationships, and Cisco still owns vast parts of enterprise and campus networking. But the fastest-growing corner of data center spending right now is not ordinary corporate switching. It is AI cluster buildout, and NVIDIA has turned that buildout into a bundled architecture conversation.
NVIDIA is selling the whole machine
Jensen Huang has been open about this strategy. At NVIDIA's shareholder meeting this week, Business Insider reported that Huang said Spectrum-X is now larger than all other Ethernet networking peers combined. NVIDIA finance chief Colette Kress also said on the company's May earnings call that broader data center networking revenue had tripled to $15 billion from the previous year. Those are not side-business numbers. They are what happens when networking becomes part of the AI system, not an accessory to it.
You can trace some of this back to Mellanox, the networking company NVIDIA bought in 2019. At the time, it looked like a smart data center deal. In hindsight, it gave NVIDIA the missing layer it needed before the AI boom turned every serious cloud buildout into a networking problem. GPUs made NVIDIA powerful. Mellanox helped make that power harder to route around.
Arista is not standing still. It has pushed 800G networking for AI workloads and has worked with NVIDIA on validated designs, which tells you where customer demand is going. Customers want high-speed Ethernet, but they also want it to behave properly around NVIDIA GPUs. That puts Arista in an awkward but familiar place: strong enough to matter, yet forced to build around another company's center of gravity.
Cisco has a different issue. It still has the broadest networking franchise of the three, but breadth can look less useful when the hot market is narrow and specific. If the buyer is trying to stand up AI capacity quickly, a general-purpose networking story carries less weight than a system already tuned for the chips, DPUs and software the buyer has selected.
The moat is moving outward
The important point is not that NVIDIA won one IDC table. Market share can move. Cloud providers can diversify suppliers, and IDC's own commentary flagged that risk. Buyers don't like being trapped by one vendor, especially when that vendor already controls the most expensive part of the build.
Still, the direction is clear. NVIDIA's moat used to be discussed mostly in terms of GPUs and CUDA. Now it includes the network fabric that keeps those GPUs busy. If a hyperscaler buys NVIDIA GPUs, connects them with NVIDIA networking, runs NVIDIA DPUs and optimizes around NVIDIA software, it is not running a neutral data center with NVIDIA parts inside. It is running an NVIDIA-shaped data center.
That distinction matters for anyone selling into AI infrastructure. The old fight was about who had the best switch. The new fight is about who controls the performance of the whole cluster. Cisco and Arista can still win plenty of business outside that tight AI stack, but in the spending pool everyone wants most, NVIDIA has made the switch part of the machine.
Nobody should treat this as the end of Cisco or Arista. That would be lazy. But you should treat it as a warning. The AI boom is not just moving money from CPUs to GPUs. It is pulling servers, networking, storage and software into tighter systems, and NVIDIA is the company pulling hardest.
Also read: Oracle cut 21,000 jobs in a year and put the blame directly on AI • The United Nations is building its own cloud and you should be paying attention • The AI boom is keeping inflation alive and interest rates higher than founders want
