Arista Prepares Networking Infrastructure for 10 Gigawatts of AI Systems

Adapting networking technology to a new computing architecture takes time. The AI-related expansion at Ethernet switch vendors, including rivals Cisco Systems and Arista Networks, has developed over several years, and the market is still in its early stages.

The challenge involves more than producing higher-bandwidth Ethernet switches and reducing the cost per bit. Those improvements supported the cloud networking era from 2010 through 2020, when increasing aggregate bandwidth per ASIC enabled larger networks and clusters.

AI workloads impose different requirements. They need high-radix devices and flatter networks because they are highly sensitive to latency. GPUs and other XPUs are expensive, so leaving them idle while they wait for data can have a significant economic cost. Unlike less expensive compute resources, AI accelerators cannot easily absorb inefficiencies through broad workload interleaving.

AI infrastructure also requires extensive automation. Balancing workloads across tens of thousands or hundreds of thousands of accelerators within a datacenter or region, and across millions of accelerators between regions, requires dynamic routing, switching, and large-scale telemetry processing. Management tools therefore need AI capabilities, including Arista’s Autonomous Virtual Assistant, because the volume and speed of the data exceed what human operators can analyze manually. Security presents another use case, as attackers also use AI to discover and exploit vulnerabilities.

Ethernet’s Expanding Role

Hyperscalers, cloud builders, AI model developers, national HPC laboratories, sovereign computing projects, and AMD generally prefer open networking technologies rather than proprietary networks throughout the AI cluster interconnect hierarchy. Nvidia is the major exception. It moved early in building tightly coupled, shared-memory GPU clusters and has gained first-mover advantages in market share, revenue, and profits.

Nvidia’s NVSwitch and NVLink serve scale-up and scale-in networking roles, but they face potential competition from more open and presumably less expensive Ethernet-based alternatives. These could include ESUN and UALink protocols, native UALink for connecting GPU memories, and UCI-Express for connecting components within an accelerator package.

The transition will take time. Arista currently earns revenue from Ethernet switches used in conventional cloud infrastructure and AI front-end networks, which connect AI clusters to external systems and storage. It also generates substantial revenue from its R series Ethernet switches, which add routing capabilities for scale-across connections between datacenters and regions.

Distributed AI infrastructure is necessary because power availability limits the size of individual datacenters and regions. Hyperscalers and cloud builders therefore need to spread large clusters across multiple locations.

Arista has not yet generated significant revenue from Ethernet switches for scale-up networks, sometimes called AI back-end networks. However, hyperscalers and cloud builders have been testing ESUN and UALink over Ethernet, or UALoE, on Ethernet fabrics.

During a call with Wall Street analysts covering Arista’s third-quarter financial results, chief executive officer Jayshree Ullal said the company had worked with AMD, Arm, Meta Platforms, Microsoft, and Qualcomm on rack-scale switching based on its Etherlink switching and routing product lines. She said scale-up trials using the ESUN protocol were expected to begin in 2027, initially with rack-scale systems containing 128 or 144 accelerators per rack and as many as 1,000 accelerators across a row. The figure was likely intended to represent 1,024 accelerators rather than exactly 1,000. Production deployments of ESUN scale-up networks are not expected until 2028.

Arista has not announced support for UALink switching, which would compete more directly with Nvidia’s NVSwitch stack. That could change because third-party UALink ASICs are available from Astera Labs, Marvell, Upscale AI, and possibly Panmnesia.

Arista does not have a significant scale-in networking offering for connections within AI accelerator packages. That area is addressed by companies such as AMD, Nvidia, Broadcom, Marvell, Alchip, and potentially Intel and Qualcomm.

Until scale-up networking becomes a larger business, Arista has substantial opportunities in scale-across and scale-out networking. The company’s third-quarter results illustrate the current strength of those segments.

Third-Quarter Results

Arista’s product revenue began to show the steep growth associated with generative AI. In the September quarter, product sales reached $3.19 billion, an increase of 66.9 percent from the same period a year earlier and 22.4 percent sequentially. Services revenue increased 17.9 percent to $467.6 million.

Total third-quarter revenue was $3.66 billion, up 58.5 percent year over year and 20.5 percent sequentially. The sequential increase was roughly two to three times higher than the company’s normal rate and represented its highest level since the early years following its founding.

Operating income was $1.64 billion, up 67.9 percent, while net income reached $1.56 billion, an increase of 82.9 percent. Net income represented 42.7 percent of revenue, the highest profit level in the company’s history. Arista ended the quarter with $14.36 billion in cash and equivalents, $11.4 billion in customer purchase commitments, and $6.84 billion in deferred revenue.

Ullal also provided figures used to update Arista’s annual product-segment model for 2026 and 2027. The company expects to sell approximately $3.5 billion in AI front-end and back-end networking equipment and services in 2026, with that figure rising to about $7 billion in 2027.

Campus switching, an adjacent market in which Arista competes with Cisco and other vendors, is expected to reach $1.25 billion this year and $2 billion next year. At the midpoint of Arista’s growth targets, overall revenue would increase 52 percent in 2026 to $13.3 billion, followed by 39.1 percent growth in 2027 to $18.5 billion.

The company’s non-AI and campus businesses are also expected to grow in 2026, largely because of scale-across equipment in the routing segment. Core cloud and datacenter switching and the edge segment are growing more slowly. Traditional workloads have been expanding more rapidly in servers than in switching, although that pattern could change.

Customers and Accelerator Coverage

Arista’s customers that each represent at least 10 percent of annual revenue, Microsoft and Meta Platforms, are working with the company on ESUN scale-up networks for their Maia and MTIA XPUs. Ullal said both customers remained on track to be major spenders in 2026 and 2027. She also said Arista could add one or two more major customers during the coming year, potentially supporting growth in 2027.

The market has changed substantially. In 2024 and 2025, Arista was discussing four or five customers with AI clusters containing between 50,000 and 100,000 Nvidia GPUs, while InfiniBand remained common in those systems.

By 2026, more than 100 customers were using Arista’s AI-focused Etherlink switches. Ullal forecast that by the end of 2027, Arista would provide networking for 10 gigawatts of AI cluster capacity and connect ten or more different types of accelerators.