Arista는 10기가와트 규모의 AI 시스템을 위한 네트워킹 구축에 나서고 있다.

It takes a while to tame a new computing architecture, so it is understandable that the AI-related ramp at all Ethernet switch makers, and particularly at arch rivals Cisco Systems and Arista Networks, has taken many years to come to fruition. To be frank, they are all just getting started.

It is not just a matter of making higher bandwidth Ethernet switches and riding down the cost-per-bit curves as the aggregate bandwidth per ASIC doubles and the cost of moving bits falls a little, allowing for elastic consumption and ever-larger networks to balance ever-larger clusters. That would be relatively easy, and it pretty much defines the cloud networking era from 2010 through 2020.

AI is different. It demands high-radix devices and flatter networks because the workload is inherently latency-sensitive, and given the extremely high cost of a GPU or XPU relative to other kinds of compute, keeping these compute engines waiting is economically devastating. If GPUs and XPUs were "cheap as chips," an English phrase referring to the fried potatoes that accompany fried fish in pub food, we would tolerate some inefficiencies, interleave workloads across more devices, and make it up in volume. That is not an option for AI training or AI inference. It hasn't been for years and won't ever be again. The tokenomics will kill the goose that lays the golden trillions of tokens otherwise.

Moreover, AI requires AI. Balancing workloads across tens of thousands to hundreds of thousands of accelerators in a datacenter or region, and across millions of accelerators across regions, requires all kinds of dynamic routing and switching and the processing of untold amounts of telemetry. Keeping these accelerators busy means embedding AI into management tools, like Arista's Autonomous Virtual Assistant, because human beings cannot correlate and analyze data at the pace of computers. Securing this infrastructure also requires AI, because bad actors are also using AI to find vulnerabilities and exploit them. All of this development has also taken time.

**The Inevitability Of Ethernet**

No one, not hyperscalers, not cloud builders, not AI model builders, not national HPC labs, not sovereigns, not AMD, except Nvidia, wants proprietary networks anywhere in the interconnect hierarchy for AI clusters. Nvidia got out well ahead of everyone else in creating tightly coupled, shared-memory GPU clusters and deserves its first-mover advantages when it comes to market share, revenue, and profits. But, as I have pointed out, NVSwitch and NVLink are the InfiniBand of scale-up and scale-in networking, and they will have to compete against more open, and presumably cheaper, alternatives based on Ethernet running ESUN or UALink protocols, and perhaps native UALink for gluing GPU memories together, and UCI-Express for gluing components together within an accelerator package.

This transition is going to take time. At the moment, Arista Networks is making money selling Ethernet switches for regular cloud infrastructure and for AI front ends, the networks that connect AI clusters to the outside world and to storage. The company is also making serious money selling its R series Ethernet switches, which are augmented with routing functions to perform scale-across datacenter linkages. These are used to make AI clusters span multiple datacenters and regions. This is necessary because you can only build a datacenter or region so big, since power is often limited in a given geographic area. The hyperscalers and cloud builders have to spread it around.

Arista has yet to make any serious money selling Ethernet switchery for scale-up networks (also called the AI back end by some), but some trials have been under way at the hyperscalers and cloud builders to run the ESUN and UALink over Ethernet (UALoE) protocols atop Ethernet fabrics.

On a call with Wall Street analysts going over the company's third-quarter financial results, Jayshree Ullal, Arista's chief executive officer, said that the company has worked with AMD, Arm, Meta Platforms, Microsoft, and Qualcomm on rackscale switching based on its Etherlink switching and routing lines. She added that she expected scale-up trials based on the ESUN protocol to start in 2027, initially with rackscale systems with a radix of 128 or 144 accelerators within a rack and with as many as 1,000 accelerators across a row. (The number is almost certainly 1,024 accelerators, not 1,000, and she was speaking generally.) Production rollouts for ESUN scale-up networks are not expected until 2028.

This seems like a long time away, and the wonder is what is taking so long.

Arista has not said it will embrace UALink switching, which would compete more directly against Nvidia's NVSwitch stack, but that is always a possibility given the availability of third-party UALink ASICs from Astera Labs, Marvell, Upscale AI, and perhaps Panmnesia.

Arista does not really have a play for scale-in networking down on AI accelerator packages. That is something for AMD, Nvidia, Broadcom, Marvell, Alchip, and perhaps Intel and Qualcomm to do.

Before scale-up ramps, there is plenty of money to be made peddling scale-across and scale-out networking, as Arista's third-quarter results show.

You can finally see the beginning of the proverbial hockey stick curve driven by GenAI in the product revenue curve shown above. Arista has been waiting years for that to happen.

In the September quarter, Arista posted product sales of $3.19 billion, up 66.9 percent year on year and up 22.4 percent sequentially. Services revenues plugged along behind, as always, growing 17.9 percent to $467.6 million in the quarter.

Added together, total revenues for Q3 came to $3.66 billion, up 58.5 percent, with sequential growth of 20.5 percent overall. This is two to three times the normal sequential growth and is the highest level set by the company since its founding years. (Those early years barely count, because there were so few deals and so little money back then.)

Arista's operating income was $1.64 billion, up 67.9 percent, and its net income was $1.56 billion, up 82.9 percent. That net income is a very healthy 42.7 percent of revenues and is the highest level of profit the company has ever enjoyed. The company exited Q3 with $14.36 billion in cash and equivalents, $11.4 billion in customer purchase commitments, and $6.84 billion in deferred revenues, also very healthy relative to the size of the business.

Ullal gave us some numbers that allow us to update our annual Arista product segment model, which forecasts 2026 and 2027 and is shown below:

And here is the same data shown graphically so you can feel the curves.

On the call, Ullal said Arista was on track to sell about $3.5 billion in AI front-end and back-end networking gear and services in 2026, and that this would double to $7 billion or so in 2027. Campus switching, an easy adjacency that helps take money away from Cisco and others, would hit $1.25 billion this year and grow to $2 billion next year. At the midpoint of growth targets, Arista will grow overall revenues by 52 percent in 2026 to hit $13.3 billion, and will grow by another 39.1 percent in 2027 to reach $18.5 billion.

We have to guess how the non-AI and campus portions of the Arista business will do, and we gave it our best shot in the table and chart above. The portions of the business not related to AI front-end and back-end networks, which include scale-across networking linking datacenters and regions, are growing well in 2026, mostly thanks to the scale-across gear in the routing product segment. The core cloud and datacenter switching business and the edge segment are growing, but not by much. I see this dichotomy in servers, too. Traditional workloads are growing faster in servers, in fact, than they are in switching. This could yet change.

Looking ahead, Ullal said that current customers accounting for 10 percent or more of annual revenue, which would be Microsoft and Meta Platforms, are working with Arista on ESUN scale-up networks for their respective Maia and MTIA XPUs and remain on track to be big spenders in 2026 and 2027. She added that the company might add one or perhaps two more big spenders in the coming year, which presumably is fueling its growth in 2027.

The market is inflecting, hard. In 2024 and 2025, said Ullal, Arista was talking about having four or five customers with AI clusters in the range of 50,000 to 100,000 Nvidia GPUs, and still seeing a lot of InfiniBand going into machines. Here in 2026, Arista already has more than a hundred customers using its AI-centric Etherlink switches, and by the end of 2027, she is forecasting that the company will provide networking for 10 gigawatts of AI cluster capacity and will be l