Well, that didn’t take long. In April 2020, Nvidia completed its $6.9 billion acquisition of Mellanox Technologies for its InfiniBand and Ethernet switching, and a little more than five years and a GenAI boom later Nvidia has been crowned the leading revenue generator for Ethernet switching in the datacenter by IDC.
A rising tide lifts all boats, but one boat is riding a slightly higher wave and that is Nvidia’s Spectrum-X Ethernet switch and DPU combo, which has become a preferred network for back-end networks for interconnecting AI clusters that scale beyond the rackscale domain of the higher-performance NVSwitch memory fabric also sold by Nvidia.
The magnitude of the Nvidia Ethernet leap is dramatic, just as has been the jump in compute for Nvidia over the past two years.
Cisco Systems is still the dominant supplier of Ethernet switches across the datacenter, the edge, and the campus, but perhaps not for long given the expected demand of AI clusters in the coming years. Nvidia’s Ethernet business is now larger than that of Arista Networks, which has been dogging Cisco in the datacenter for the past decade and a half and which aimed to break $10 billion in sales next year, two years earlier than its top brass had planned. And even with the just completed Juniper acquisition, the embiggened HPE Networking will have trouble catching Nvidia in the datacenter even moreso than both Cisco and Arista.
Let’s go over the numbers from IDC about the Ethernet switching market for the second quarter of 2025, which as usual we have enhanced with our own estimations, and then drill down into the datacenter that we focus on here at The Next Platform. (Our estimates are shown in bold red italics in the tables below, as always.)
In the June quarter, IDC reckons that companies around the world bought $14.51 billion of Ethernet switching year, an increase of 42.1 percent from the year ago period and a whopping 23.9 percent sequentially from Q1 2025.
The router market bumbles along, as always. We are tangentially interested in routers because it is a source of revenue for Cisco, Huawei Technologies, and Juniper that helps fuel their switching ambitions and because it shows that despite the addition of routing functions to network ASICs over the past decade and a half, dedicated routers are still used by telcos, service providers, and quite a number of cloud builders. In Q2, router revenues grew by 12.5 percent to $3.52 billion. The ratio of switch revenue to routing revenue for Ethernet gear averaged about 1.8 to 1 before the GenAI boom, but since then has been creeping up and in the most recent quarter, given the massive AI buildout within datacenters, stood at 4.1 to 1.
Switching is exploding, and there is a chance that routing may, too, as datacenters are interconnected to make larger AI cluster footprints.
IDC said in statements accompanying its numbers that Cisco had 32.9 percent of router sales, which works out to $1.16 billion, and increase of 17.7 percent year on year. This means Cisco is growing faster than the router market at large. With only 13.3 percent growth, Huawei, at $1.08 billion in router sales, is growing slower than the market. We do not know Juniper’s slice of routing, but in the wake of the HPE acquisition it will no doubt be mentioned in the next report, which will be for the first full quarter that Juniper has been part of HPE.
The New Switcheroo
Here is the historical trend of Ethernet switch revenues by vendor that IDC puts together each quarter, with the Q2 numbers tacked on:
Q1 looked to be a little bit soft, as we can now see, but was nothing out of the ordinary. The first quarter of any year tends to be weaker than the fourth quarter of any prior year. It is just during boom times that this doesn’t always happen.
A stacked bar chart is, by definition, hard to convey trendline information at a granular level, which is why companies use them. It gives you a feel for the data, but not as good as a longer set of actual trendlines that overlap by category. So here is what we think is a better visual, with a longer runway of data, to help you see the trends:
And here is some of the data we have reconstituted from quarterly statements by IDC and guesstimates that we have to make to fill in the gaps:
Cisco is still the biggest peddler of Ethernet switch gear, at $3.96 billion, but it is growing about a quarter as fast as the market overall because it is not really getting much action in AI networking. At least not compared to upstart Nvidia. Moreover, Arista is giving Cisco some grief in the campus switching market where it has had hegemony for decades – although not as much grief as Ullal & Co had been hoping, we think. (These things take time.)
Nvidia, which is only sells datacenter Ethernet switching gear, is now not just a little bit ahead of Arista, but a lot bit. In the June quarter, we reckon that based on IDC statements that Nvidia had $2.26 billion in Spectrum switch sales, up by a factor of 7.5X compared to the year ago quarter. (BOOM! echo echo echo echo. . . . )
Arista, interestingly did not grow as fast as the average that was significantly lifted by the massive AI buildout that Nvidia has engineered – and engineered over the past decade brilliantly, we might add – had Ethernet switch sales of $1.83 billion, up “only” 33.5 percent year on year and up “only” 12.4 percent sequentially.
Huawei had $1.2 billion in Ethernet switch sales, up 2 percent from the year ago period but up 77.6 percent sequentially from a pretty bad Q1. Other vendors, including the original design manufacturers who make custom gear for hyperscalers, cloud builders, telcos, and large service providers, accounted for $4.58 billion, up 42.5 percent and keeping pace with the overall market.
Things get more interesting when you separate datacenter switching from campus and edge switching:
Our model filling in the gaps of the public data from IDC says that datacenter Ethernet switch sales were $8.72 billion in the quarter, up 71.6 percent year on year and up 28.1 percent sequentially. That represented 60.1 percent of overall Ethernet switch sales, flipping and flopping the share that datacenter switching had compared non-datacenter switching (40 percent to 60 percent) prior to the GenAI boom.
That non-datacenter portion accounted for $5.79 billion in Ethernet switch sales, up a mere 12.5 percent.
In the chart above, the port counts to the right of the red line are estimates made by us. IDC stopped giving out port counts after the data for Q2 2022 was released. So take our estimates of the cost per port, outlined below, with a whole lot of salt. We are filling in gaps, but don’t treat that like the gospel truth.
Now, let’s break the datacenter segment of the Ethernet switch market down by vendors:
As you can see, Nvidia has pulled far ahead of both Cisco and Arista. Cisco had $1.26 billion in datacenter Ethernet switch sales, up 9.1 percent, and Arista had $1.66 billion, up 34.2 percent. We have no idea what Huawei’s Ethernet switch sales were in the datacenter, but it has to be a sizable chunk of its $1.2 billion in overall Ethernet switch sales. Put a gun to our heads and we would use the 80-20 rule: 80 percent datacenter, 20 percent campus and edge. Call it around $900 million or so.
The ODMs as a group had $1.71 billion in Ethernet switch sales into the datacenter, up 76.9 percent year on year and accounting for 19.6 percent revenue share. ODMs have been around a fifth of the market for the past six quarters, and it looks to us like Nvidia’s success in AI clusters has stunted the growth of ODMs a bit. HPE, with only $103 million in sales, was up 20.3 percent but this was noise in the data.
Every quarter based on the IDC data, we try to aggregate all of the port counts and port speeds and port revenues to get a sense of how much Ethernet bandwidth the world is consuming and give a sense of the cost per bit trends. Let’s start with the trends, which are not necessarily smooth:
Again, any data to the right of the red line is based on estimated port counts. Anything to the left is using the port counts formerly disclosed by IDC with some estimates here and there where they were vague.
The cost per bit trends down as time goes by, but you can see instantly that there are three different bands of networking being used and they have very different price profiles, and a fourth one is emerging with the advent of 800 Gb/sec switches.
You can get a cheap 1 Gb/sec Ethernet switch, but the cost per bit it terrible and you could never aggregate these to build a datacenter-scale network. There would be far too many hops in any network topology that the latency would stall any massively distributed computing job unless it was embarrassingly parallel (like Web serving or EDA processing is).
The 10 Gb band, which led the data analytics revolution in the late 2000s, has been great for the datacenter but is only now being used in the campus and edge in high volume because the price per bit has come down by an order of magnitude. It will be harder to make it come down much lower than it is today.
The 100 Gb/sec tier brought a factor of 4X improvement in cost per bit at 10X the bits twiddled with, but it has been a relatively flat curve in terms of bringing down the cost per bit over time.
The initial 200 Gb/sec and 400 Gb/sec switches were very pricey and datacenters used them as sparingly as possible. We only have two datapoints for 800 Gb/sec switches as a group, and they are in the same ballpark on a cost per bit basis as 100 Gb/sec switches, which is about twice as expensive as the average of 200 Gb/sec and 400 Gb/sec gear. (IDC lumped together 200 Gb/sec and 400 Gb/sec, we didn’t.)
And finally, this table brings it all back together:
That $14.51 billion dollars in Ethernet switch spending worldwide in Q2 happened through the consumption of 353.8 million ports, most of them not in the datacenter. A total of 12,374 exabits per second of capacity was installed in Q2 2025, and 20.6 percent of that came from 800 Gb/sec switches, 47.8 percent came from 200 Gb/sec and 400 Gb/sec switches, and 24.4 percent came from 100 Gb/sec switches.