
- Arista launches lower-speed routers for front-end AI networks, where inference shares the network with storage, WAN and legacy apps
- The vendor claims up to 60% lower power than rivals, a comparison against unnamed platforms built on older 7-nm and 16-nm chips
- Data sovereignty is driving enterprise demand for local inference, especially in Europe and Asia-Pacific, Arista said
Arista Networks is taking its AI networking push to the front end of the network with new R4 routers built for distributed AI inferencing. The launch includes compact and hardened access routers, a coherent transport router and a 25.6-Tbps pluggable optical line system.
The vendor is responding to the requirements of agentic AI, where distance is expensive and moving data is often restricted by regulation. Requests served from a faraway cluster come back slowly, and some data is not allowed to leave its home country. Distributed inferencing requires running trained AI models at many sites close to users instead of in a few centralized clusters.
Arista’s recent AI launches went after the back end of the network, designed for training clusters, scale-up racks and scale-across links between sites. The front end is a different problem, said Brendan Gibbs, area VP of product line management at Arista, in an interview. There, inference traffic shares the network with storage, WAN links and legacy applications, and that mix calls for routing, deep packet buffers and built-in encryption, Gibbs said. Back-end networks, by contrast, connect AI accelerators to one another for a single job, training, and are built for the highest speeds and densities.
“A lot of our prior portfolio announcements have been the superlatives, like highest density, highest speeds,” Gibbs said. This launch goes the other way, with 25-gig and 100-gig aggregation for AI servers.
The new 7282R4 router is an example of that transition. It is a fixed router, a self-contained box with a set number of ports, as opposed to a chassis that takes swappable cards. It carries two redundant supervisor modules, the processors that run the control plane. Large data centers get that redundancy by spreading traffic across many switches, but an edge site or a peering point may depend on a single router, Gibbs said.
“The redundancy goes into the device instead of the network,” Gibbs said.
Who is Arista targeting?
Arista has two target customers groups and their needs are splitting apart, Gibbs said. Hyperscalers and neoclouds commonly run clusters of 100,000 accelerators and are heading toward 1 million across multiple data centers. Those customers require 1.6-terabit links and liquid cooling. Enterprises are building AI clusters of roughly 500 to 1,000 accelerators, air-cooled, at lower speeds.
Arista’s bet is that more inference running close to users means more sites to connect across metro and edge networks, said analyst Sameh Boujelbene, VP of data center switch and AI networks market research at Dell’Oro Group.
“Whether edge inference is delivered by telcos, cloud providers or specialist edge operators, Arista wants to sell the routing infrastructure that connects those sites,” Boujelbene said.
Data sovereignty is driving enterprise interest, particularly in Europe and Asia-Pacific, where gigawatt-scale data centers are scarce and linking smaller sites is the default, Gibbs said.
“Maintaining privacy of data and maintaining privacy of inferencing requests inside national sovereignty is really one of the first discussions we have with enterprises,” Gibbs said.
How does Arista compare with Cisco, HPE and Nokia?
For customers smaller than hyperscalers, Arista is already the No. 1 data center switch vendor. Gibbs threw shade at the competition.
“I would consider them legacy routing vendors because they’ve been around for 20 years, and that doesn’t mean that they’re necessarily setting the pace for innovation anymore,” Gibbs said.
Arista claims its new routers use up to 60% less power than competing platforms. The comparison is against platforms built on 7-nm or 16-nm chips. R4 uses 5-nm silicon from Broadcom. Arista also said its routers offer both MACsec and IPsec encryption on every port, and that its EOS software has fewer published vulnerabilities than rival network operating systems, based on its own count of public CVE data as of Oct. 21.
For organizations trying to stretch the power available for GPUs and other AI accelerators, that saving could be compelling, said Bob Laliberte, principal analyst at Liberte Research Group.
Rivals are chasing the same traffic. Cisco launched its 8223 router and Silicon One P200 chip for links between AI data centers in October 2025. HPE pitched its Juniper MX301 edge router at inference close to where data is generated. Cisco also has a compute partner at the edge: AT&T is running Cisco’s AI Grid with Nvidia in commercial pilots. Arista’s announcement names no compute partner.
Arista is packaging Broadcom Jericho-based hardware and EOS in form factors better suited to the edge, Boujelbene said.
“Cisco, HPE Juniper and Nokia will pursue the same opportunity, but Arista is trying to turn its cloud routing track record into production wins in metro and carrier networks,” Boujelbene said.
“Arista’s strongest argument is operational consistency for customers that already use EOS, combined with a broader range of routing platforms,” Laliberte said.
That makes the new routers most attractive to organizations already running Arista in AI data centers that want to extend familiar tools and skills to more locations, he said.
“An operator with an established competing environment will need to weigh those benefits against migration and integration costs,” Laliberte said.
Arista enters the fight with momentum. Third-quarter revenue rose 58.5% year over year to about $3.66 billion, its fastest growth since its IPO.
PR Archives: Latest, By Company, By Date