# As Goes AI Compute, So Goes Ethernet Networking > Source: > Published: 2026-07-06 15:04:26+00:00 # As Goes AI Compute, So Goes Ethernet Networking InfiniBand may still be the preferred interconnect for scale out networks for plenty of HPC labs and more than a few AI centers, but the broadening out of GenAI from model training to production inference means that bog standard Ethernet, which is evolving to have many of the capabilities of InfiniBand as well as some features in which it is lacking, means that the market for Ethernet switches, network interface cards, and DPUs is going to grow a lot faster than the historical trends. Ethernet switching for AI can grow just like the severely constrained market for AI compute is doing. In fact, in Q1 2026, the Ethernet switch market grew faster than [the overall server market did](https://www.nextplatform.com/compute/2026/06/17/the-server-boom-balances-price-increases-against-chip-shortages/5258066) according to the latest statistics from IDC. To be specific, the overall server market grew by 30.4 percent to $122.62 billion in Q1, while Ethernet switch sales – this does not include DPUs, NICs, or cables, but just the switches – rose by 39.8 percent to $15.36 billion. If you drill down further, isolating datacenter Ethernet switch sales from edge and campus devices, Ethernet in the glass house (more of a metal warehouse these days) rose by 61 percent to just a hair over $10 billion. Clearly, at least some customers are spending a lot more on bigger networks and faster ones, which is driving Ethernet switch sales growth much higher than for edges and campuses, which rose by 12.3 percent to $5.35 billion in the March quarter. It is hard to say for sure, but a fairly good portion of the revenue increases we see in the Ethernet space are due to the hikes in switch component prices, particularly those that have DRAM memory embedded in them for auxiliary storage. The router market keeps plugging along even though modern switch ASICs from Broadcom, Marvell, and Cisco Systems include routing functions as well as those for switching. Router revenues in Q1 were up 11.3 percent to $3.75 billion, with service providers and midrange and large enterprises driving sales because they are not able to turn switches into routers with their own network operating system. [DriveNets has a router operating system called Network Cloud](https://drivenets.com/solutions/service-provider-routing/) that can be deployed on commodity hardware if you really want to disaggregate your NOS from your switch like a hyperscaler and cloud builder. But even these companies sometimes buy commercial routers from the likes of Cisco, Juniper Networks, and Huawei Technologies. In fact, service providers and what IDC calls “cloud SPs,” which could mean the big cloud builders or not, accounted for $2.9 billion in router revenues in Q1, which was 77.2 percent of router sales during the period and which grew by 12.9 percent. Enterprises tend to stretch out their router upgrade cycles, and sales among “normal companies” hit $867 million, only rising 6.1 percent. Router sales in the Americas rose by 19.6 percent, EMEA was up 9.8 percent, and Asia/Pacific was up 2.8 percent. If IT infrastructure is an arms race, the United States is clearly and emphatically winning according to all of the data I have seen in the past several years. IDC normally puts out a pretty chart with the vendor shares, but this time around the Ethernet data was in a blog post, not a press release, and the chart was not included. Because IDC had moved the quarterly Ethernet data to a blog, we never did see the data for Q4 2025, which we have included in our tables and charts, and the blog posted back in March for Q4 2025 did include that chart. It is shown below: We like our line chart better than IDC’s stacked bar chart because you can see the vendor trends a lot more easily. The blogs do not talk about sales by original design manufacturers (ODMs) as the old press releases about the Ethernet switch and router market did, so we have done our best to estimate ODM sales for Q4 2025 and Q1 2026; we assumed a more or less steady state market share. As a reminder, the data above is for all Ethernet switch sales, including edge, campus, and datacenter gear. Here is the same data in a table: The big news this time around was that Nvidia was the number one supplier of Ethernet switches *in the datacenter*, surpassing both Cisco and Arista Networks. But as far as we know, that happened back in Q2 2025. And in Q4 2025, Nvidia beat Cisco but not Arista in the datacenter. And by the way, we checked again with IDC and Nvidia’s sales in this dataset are for SpectrumX Ethernet switches only, and does not include sales of cables, NICs, and DPUs. These contribute pretty mightily to Nvidia’s overall switching revenues, which also includes Nvidia’s InfiniBand gear. Nvidia only ships datacenter gear, which Arista, Cisco, and others sell into edge and campus spaces. In Q1, Nvidia’s datacenter Ethernet switch revenues were $2.1 billion, up by a factor of 2.9X year on year, while Cisco’s datacenter revenues were up “only” 43.2 percent to $1.78 billion. Arista’s datacenter Ethernet sales, according to the IDC data, was up 40.1 percent to $2.07 billion – very nearly what Nvidia made in datacenter sales. Here is a table that lays datacenter Ethernet switch sales all out since 2022: We have had to make a lot of estimates in this table based on incomplete data because IDC does not release data on all vendors all the time. The market researchers at IDC only talk publicly about the top five vendors plus ODMs, and now they have dropped ODMs from the discussion. We fill in gaps where we can. Ditto for the cost per bit chart, which has the cost per bit based on Ethernet port speeds, shown below: Many years ago, IDC gave out data on revenues by port speeds and changes in port counts so you could reckon cost per bit by generation. It was very interesting data, and we miss it. We have made our guessed about what port counts must look like assuming ever-decreasing costs per bit by ASIC generation. It is admittedly a guess, but in the absence of data, that is what we all have to do. Here is what is important as far as I am concerned. Sales of switches with 800 Gb/sec ports rose by a very dramatic 10.3X to $3.58 billion in Q1 2026, and it won’t be long now before we see 1.6 Tb/sec and even 3.2 Tb/sec ports appearing in switches. As far as we can tell, 800 Gb/sec switch sales were down 14.3 percent sequentially. Q1 sales for many technologies are generally lower than Q4 of the previous year, so this is not surprising. I think the port count has risen by a factor of 22X, and that means the cost per port and therefore the cost per bit has been cut by more than half since a year ago. This is what happens when there is competitive pressure as well as technological advancement. Switches with 800 Gb/sec ports (this is native, not with cable splitters) represented 35.8 percent of datacenter Ethernet sales in Q1 2026, and switches with native 200 Gb/sec or 400 Gb/sec ports represented another 34.1 percent of the market, with revenues of $3.41 billion, off 10.4 percent year on year and off 52 percent sequentially. Clearly the shift is on to 800 Gb/sec gear, and clearly except for enterprises stuck at 10 Gb/sec ports on their Ethernet networks, 100 Gb/sec switches just don’t cut it anymore for anyone build big and serious networks. It costs 61 cents per Gb/sec for a native 800 Gb/sec switch port, on average based on our estimates on port counts, compared to 53 cents per Gb/sec on a blended average of 200 Gb/sec and 400 Gb/sec ports. Ethernet switches with native 100 Gb/sec ports cost twice that, those with native 10 Gb/sec ports more than double that. Switches with native 1 Gb/sec ports cost between $10 and $11 per Gb/sec. Yes, the lower speed ports have lower capital outlay, which is why they still are selling, but you sure do pay a premium when it comes to bits pushed through ports.