Arista adds routers to connect AI inference across data centers and remote sites
The Etherlink R4 release combines encrypted routing with optical transport. Most products are available now, while a temperature-hardened model follows in early 2027.
Loading page…
The Etherlink R4 release combines encrypted routing with optical transport. Most products are available now, while a temperature-hardened model follows in early 2027.
Listen to this story
At SC26 on November 17, Arista introduced Etherlink R4 routers and optical equipment for connecting distributed AI inference sites. The four router models target different needs, from site aggregation and cell-tower deployments to redundant control and 400/800-Gbps optical transport; Arista claims up to 60% lower power than 7nm or 16nm alternatives. Most equipment is available now, while the temperature-hardened 7020HR4 is expected in early 2027. Arista says scale-across accounts for about 30% of its AI business, but that is not a measure of distributed inference adoption.
The 7020R4 includes MACsec and IPsec; the 7282R4 adds dual supervisor engines for control-plane redundancy.
Arista describes a 10-year lifecycle goal in which customers replace optics rather than the router chassis.
Brendan Gibbs said the products are not intended to replace firewalls.
Arista Networks is extending its AI networking lineup beyond centralized computing clusters to the connections serving users and remote sites. Announced November 17 at SC26, its new Etherlink R4 routers and optical line system target distributed inference: running AI workloads across locations rather than relying on one central cluster. Arista’s pitch combines faster responses with encrypted links between sites.
Inference is the work of running a model to produce a response. Arista says users far from centralized compute can experience slower chatbot responses or delayed execution by AI agents.
Brendan Gibbs, Arista’s AVP of product line management, argues that bringing compute closer to users reduces round-trip delays, including for chained inference calls. Arista also presents distribution as a way to keep sensitive traffic within national borders.
Running Arista’s EOS operating system, the platforms combine encryption, separation of multiple customers’ traffic, and traffic engineering to steer workloads along selected network paths. Deep packet buffering and redundant control systems also feature in the portfolio.
The accompanying optical line system includes a 25.6-terabit-per-second amplifier, equipment for combining and separating optical channels, and a pluggable fiber-testing instrument. The tester supports validation while fiber connections remain in service.
Arista says its 5nm R4 platforms consume up to 60% less power than alternatives using 7nm or 16nm chips. This is a company-supplied comparison.
In a pre-briefing described by analyst Zeus Kerravala in SiliconANGLE, Gibbs said roughly 30% of Arista’s AI business is now “scale-across,” connecting infrastructure across sites. That business-mix figure is not a measure of distributed inference adoption.
Gibbs also described a goal of a 10-year lifecycle, with customers replacing optics rather than the router chassis. Integrated encryption has a boundary, however: he told Kerravala that Arista does not intend these products to replace firewalls.
Most of the announced equipment is available now. The exception is the 7020HR4: customers planning temperature-hardened deployments near cell sites must wait for its expected availability in early 2027.
Loading discussion...
Join the conversation
Explain whether long-term reuse or flexibility would matter more to you.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.