Arista 10 gigawatt AI networking is the central claim behind Arista Networks’ strong growth outlook, but the forecast spans several layers of AI infrastructure rather than proving that Ethernet has already displaced Nvidia inside the rack. CEO Jayshree Ullal said the company expects networking deployments of more than 10 gigawatts across 10 or more types of compute accelerators by the end of 2027. That figure includes scale-up, scale-out and scale-across networks, and some scale-across capacity may already be installed. Arista reported third quarter 2026 revenue of 3.66 billion dollars on November 3, 2026, up 58.5 percent from a year earlier and 20.5 percent from the second quarter.
Arista 10 gigawatt AI networking: Growth reaches a record pace
Ullal called the third quarter result the highest year-over-year revenue growth since Arista’s initial public offering. Product sales climbed to 3.19 billion dollars, up 66.9 percent, while services revenue added 467.6 million dollars, up 17.9 percent. The mix shows that network hardware is leading the expansion as demand for AI infrastructure accelerates.
Arista expects fourth quarter revenue of 3.9 billion to 4.0 billion dollars and full-year 2026 revenue of about 13.3 billion dollars, up about 48 percent. For 2027, the company targets growth of 38 to 40 percent, or roughly 18.5 billion dollars. AI fabrics revenue is expected to double from about 3.5 billion to about 7 billion dollars, while campus networking is forecast to rise from about 1.25 billion to about 2 billion dollars.
On those targets, AI fabrics would account for about 26 percent of 2026 revenue and about 38 percent of 2027 revenue. That is a meaningful shift for a company whose networking products historically served cloud and data center customers, with AI infrastructure becoming a larger part of the growth mix. The forecast also depends on hyperscalers continuing to expand AI training and inference deployments and on Arista winning additional large customers.
What the gigawatt forecast includes
The distinction between network layers matters for investors, engineers and buyers evaluating the competitive threat to Nvidia. Scale-up links accelerators inside one tightly coupled rack-scale domain. Scale-out links racks across a facility. Scale-across links separate data centers or regions into one logical cluster.
Ullal was direct about the near-term limitation of scale-up revenue. She said she does not think there is a material number associated with scale-up in 2027. Scale-up trials are expected to begin in 2027, with production in 2028. As a result, the more than 10 gigawatt forecast should not be read as evidence that Arista is already taking meaningful scale-up business from Nvidia’s proprietary systems.
Arista said it has passed 100 Etherlink customers. The customer comparison requires care. In 2024 and 2025, Arista described four or five customers running AI clusters of 50,000 to 100,000 Nvidia GPUs, with InfiniBand still common. The 2026 count of more than 100 covers customers of AI-focused Etherlink switches, which is a broader group. The headline customer expansion therefore reflects both broader adoption and a changed definition of the customer base.
Microsoft and Meta are each identified as accounting for 10 percent or more of Arista’s revenue. Ullal said the company expects to add at least one, maybe two more customers at that scale. Those relationships could provide a foundation for larger deployments, although each also creates concentration risk for a supplier whose growth is increasingly tied to AI spending.
Ethernet alternatives gain structure
Arista has worked with AMD, Arm, Meta, Microsoft and Qualcomm on rack-scale switching based on Etherlink switching and routing lines. Its ESUN work is aimed at making Ethernet more competitive for tightly coupled accelerator networks. ESUN, short for Ethernet for Scale-Up Networking, is an Open Compute Project workstream launched on October 13, 2025 with 12 founding members, including AMD, Arista, Arm, Broadcom, Cisco, HPE Networking, Marvell, Meta, Microsoft, Nvidia, OpenAI and Oracle. More than 175 companies now participate.
The ESUN 1.0 specification was announced on March 10, 2026. It replaces the IP and UDP header stack, typically 28 to 48 bytes, with a compact 4-byte ESUN header for scale-up traffic. The design also uses Priority Flow Control, Credit Based Flow Control and Link Level Retry to support lossless delivery. These changes could reduce the burden of small messages, although independent measurements of the claimed efficiency and latency benefits have not been reported.
There are security boundaries. Because the ESUN header replaces the IP header, it is not compatible with IPsec. MACsec is proposed for point-to-point link encryption, while end-to-end encryption is out of scope. Buyers evaluating ESUN systems will need to account for that distinction as they compare open Ethernet architectures with established proprietary approaches.
Scale-up remains a technical contest
Arista’s Etherlink SU-144 rack-scale architecture supports up to 144 accelerators in one domain and extends to 1,024 accelerators across racks. Its orthogonal-chassis design supports 100 to 400 kilowatt liquid cooling. The Next Platform also reported that Ullal discussed initial systems with a radix of 128 or 144 accelerators within a rack and roughly 1,000 accelerators across a row.
Competitors are not following one common Ethernet path. The open market is divided between UALink and ESUN with SUE-T. Broadcom has walked away from UALink, while AMD’s Helios racks run UALink over Broadcom Ethernet switches. Arista has not committed to UALink switching, although third-party UALink chips are available from Astera Labs, Marvell and Upscale AI.
Nvidia retains a major technology lead in tightly coupled systems. Nvidia NVSwitch connects 72 accelerators in one domain and reaches 576, which Timothy Prickett Morgan of The Next Platform believes requires two switch tiers. UALink 1.0 can reach 1,024 in a single switching level. InfiniBand port hops run roughly 100 to 120 nanoseconds, about one-third to one-half the latency of Ethernet switches.
Prickett Morgan estimates Nvidia holds about 95 percent of GPU revenue and about 75 percent of combined GPU and custom accelerator revenue. His view is that NVSwitch could remain dominant for years before Ethernet alternatives become more established. For Arista investors, the implication is that scale-out and scale-across offer a nearer route to growth, while scale-up remains an option whose timing and economics are still unproven.
Financial strength supports the ramp
Arista reported GAAP operating margin of 44.9 percent and non-GAAP operating margin of 49.1 percent. GAAP net income was 1.56 billion dollars. Non-GAAP earnings per share were 1.20 dollars and GAAP earnings per share 1.22 dollars. Gross margin was 64 percent, down from 65.2 percent a year earlier.
Cash and marketable securities rose to 14.4 billion dollars, and Arista carries no debt. The balance sheet gives the company room to invest in liquid cooling, high-radix switching, AI software and hyperscale customer work. It also reduces the financial pressure that might otherwise accompany a sharp increase in inventory and supplier commitments during an AI infrastructure boom.
Management is building ahead of demand. Purchase commitments rose to 11.4 billion dollars from 9.7 billion dollars, while inventory reached 3.2 billion dollars from 2.5 billion dollars. Chief Financial Officer Chantelle Breithaupt said those commitments include chip-related spending and a multiyear memory agreement. Such purchasing supports capacity for customer demand, but it also raises the risk that forecasts will be revised if accelerator deployment schedules change.
Arista’s fourth quarter forecast calls for gross margin of 63 to 64 percent and operating margin of 48 to 49 percent. For 2027, management projected gross margin of 62 to 64 percent and operating margin of approximately 49 percent. These ranges indicate that the company is preparing for substantial growth while recognizing pressure from investment and the changing mix of AI hardware.
Investors reward the growth outlook
Arista shares closed up 3.52 percent at 169.91 dollars in the regular session and rose 12.1 percent after hours to 190.46 dollars. The reaction reflected the strength of the revenue outlook and the scale of the AI networking opportunity. The forecast for more than 10 gigawatts gives investors a new measure of potential deployment exposure, even though it combines new and already installed capacity.
The next phase will be determined by execution rather than the headline alone. Investors should watch whether scale-up trials begin in 2027, whether Arista commits to UALink switching and whether independent tests validate ESUN performance claims. Scale-out and scale-across demand can carry the business through 2027, but sustained upside may depend on whether Ethernet can overcome the latency and adoption barriers facing scale-up.
Arista 10 gigawatt AI networking is best understood as a broad infrastructure forecast, not a near-term displacement claim. The company’s results show that customers are buying Ethernet switches for AI front ends, back ends and the links that span facilities today. Its growth targets require continued hyperscale investment, new large customers and progress toward production scale-up systems in 2028. The opportunity is substantial, but the route to realizing it will be measured one network layer and one customer deployment at a time.