Skip to content

UPDATED 12:57 EST / NOVEMBER 17 2026

INFRA

Arista takes aim at the next AI network battleground: distributed inference

Awareness of the network’s role in AI has steadily increased, with much of the focus on the back end – larger clusters, faster ports, and scale-across fabrics that stitch GPUs together across data centers. Arista Networks Inc. has seen continued, accelerated growth and has been one of the biggest winners of that buildout.

This week at SC26, Arista announced a shift in focus to the other side of the network. The company also announced a new set of Etherlink R4 routing platforms designed for what it calls distributed inferencing, along with a pluggable optical line system for data center interconnect. The lineup includes:

  • 7020R4 and 7020HR4: Compact 1RU access routers with 25G to 400G ports, deep buffers, and optional wire-speed encryption on every port. The HR4 versions are temperature-hardened and add advanced timing for cell sites and outside plant.
  • 7282R4: A fixed-edge and core router with dual supervisors for control-plane redundancy.
  • 7280R4: A 25.6 terabit “in-betweener” with 64 400G ports, half supporting ZR/ZR+ coherent optics, to bridge local fabrics to the wide area network.
  • Pluggable optical line system: A full C-band amplifier carrying 25.6 Tbps over a single fiber pair, plus muxes and a pluggable fiber tester.

All are based on Broadcom’s Jericho and Qumran silicon, designed to deliver up to 60% lower power consumption than competing 7nm and 16nm platforms. Everything is available now, except the 7020HR4, which ships in early 2027. The speeds and feeds always make headlines, but the implications of the launch point to where AI infrastructure is heading.

Inference is about to get distributed

During a pre-briefing, Brendan Gibbs, Arista’s AVP of product line management, argued that the forces driving training across multiple sites will have the same effect on inference. Gibbs also said that roughly 30% of Arista’s AI business is now scale-across, indicating that distributed AI is rapidly becoming commonplace.

Inference will be distributed for three reinforcing reasons:

  • Latency. A centralized cluster works well for a chatbot that streams text line by line. It doesn’t work for agentic AI, where a single request can trigger dozens of chained model calls and every millisecond compounds.
  • Sovereignty. Regulators increasingly want inference requests and data to stay within national borders. European IT leaders have little appetite for sending data to U.S. cloud providers.
  • Power and space. Companies will deploy inference wherever they have capacity, and the network must support it.

The implication is that the front-end network, long treated as plumbing, now needs back-end attributes: deep buffers, traffic engineering, multi-tenancy, and encryption, because traffic now leaves the building.

Arista is going downmarket

This launch is the opposite of the descriptors we typically associate with Arista. Recent announcements focused on dense 800G, 1.6T, and liquid cooling. The products launched at SC26 are smaller and less expensive. Arista dominates big iron and is now rounding out its portfolio. Arista’s customers range from hyperscalers to neoclouds to enterprises building local AI. The front end needs rich features, not massive bandwidth, and Arista hasn’t had a viable option for that tier until now.

These routers are designed as entry-level but ship with 400G uplinks, which may seem like overkill today but offer investment protection. Gibbs said the goal is a 10-year life cycle, with customers swapping optics rather than chassis. No one knows what agents will do to traffic volumes, and I suspect companies that consider this overkill today will think differently in three or four years. Headroom is cheaper than a forklift upgrade at a remote site. Also notable: Arista co-designed the 7280R4 with a hyperscaler seeking a gray-to-coherent bridge, a sign that it’s building with its biggest customers.

The service provider wildcard

The most intriguing product is the hardened 7020HR4, built for 5G and AI-RAN cell sites. Telcos have largely sat out the AI buildout, yet they own the last mile and real estate near users, exactly what low-latency inference requires. This gives a carrier the backbone to offer a managed AI factory.

My view is that this performs better outside the U.S., where enterprises don’t trust carriers beyond connectivity. In Europe and Asia-Pacific, managed services are far more common, and providers such as BT, NTT, and Telstra have the footprint to make it work. With sovereignty requirements, the opportunity for telcos to move up the stack has never been better. Telcos have never been a core market for Arista, so any growth from that audience is additive to their already massive TAM.

Quality and security are the quiet differentiators

Since 2019, Arista has reported a far lower CVE count than competitors, even after the surge in AI-assisted discovery following Anthropic’s Mythos. I’m skeptical of CVE comparisons because counts depend on testing rigor, installed base, and reporting practices, and some vendors now bundle multiple issues into a single CVE.

The more relevant benefit is operational. Customers often don’t apply patches because upgrades cause downtime and break things. If they trust the software, they patch faster, and the risk window shrinks. With Mythos accelerating the retirement of end-of-support gear, quality and longevity become buying criteria, not nice-to-haves. On the call, Gibbs agreed it’s hard to draw conclusions from CVE counts but explained that CVE counts were a measure of Arista’s software quality. This has been Arista’s calling card since it was a startup. While CVEs provide one metric, the bigger proof point is its historical track record of quality software.

Also, wire-speed MACsec and IPsec on every port is a differentiator because AI brings different security requirements. However, integrated security does not replace firewalls. I asked Gibbs about this, and he said that Arista has no intention of replacing firewalls and expects Palo Alto Networks to continue to sit behind it in the line. I’d like to see that formalized into validated designs for distributed inference, like what Nvidia-validated designs did for the back end.

The optical line system is a similar simplification play. Arista isn’t an optical company and is reselling these products, but offering dark fiber customers a turnkey, EOS-integrated way to move 25.6 Tbps across a metro network removes a real headache.

Customers have responded positively to Arista’s entry into this market. Douglas Fast Net CTO Jon Vradenburg said the company needs “a compact platform that supports the services our business and carrier customers rely on.” Takayuki Sakou of Japan’s STNet cited “higher throughput and improved power efficiency while maintaining the quality of the EOS software we trust.”

Considerations for network professionals

  • Map where inference will run, not just where it runs today. If your AI roadmap includes agents, assume that latency and sovereignty will push inference closer to users, and plan the front end accordingly.
  • Treat the front end as part of the AI fabric. Evaluate buffering, traffic engineering, and multi-tenancy at the WAN edge with the same rigor you apply to GPU clusters.
  • Make encryption the default. Inference traffic between sites carries prompts and proprietary data. Line-rate encryption should be table stakes.
  • Buy for longevity. Prioritize platforms with uplink headroom, modern silicon, and a software track record that lets you patch without fear.

Final thoughts

The back end of AI networking is largely a solved architecture at this point. The front end is not, and that’s where the next round of competition will play out. With R4, Arista is betting that its routing credibility, built quietly within hyperscaler backbones, can make it the default there as well. Given how much traffic agentic AI is likely to generate, that’s a bet that, like most Arista bets, will likely pay off.

Zeus Kerravala is a principal analyst at ZK Research, a division of Kerravala Consulting. He wrote this article for SiliconANGLE. 

Image: Arista

A message from John Furrier, co-founder of SiliconANGLE:

Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.

  • 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more
  • 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network

Are you an AWS customer?  Support SiliconANGLE financially by buying your AWS services from our Marketplace portal page and links: https://siliconangle.com/aws-marketplace/

 

About SiliconANGLE Media
SiliconANGLE Media is a recognized leader in digital media innovation, uniting breakthrough technology, strategic insights and real-time audience engagement. As the parent company of SiliconANGLE, theCUBE Network, theCUBE Research, CUBE365, theCUBE AI and theCUBE SuperStudios — with flagship locations in Silicon Valley and the New York Stock Exchange — SiliconANGLE Media operates at the intersection of media, technology and AI.

Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.

Send us a news tip

Send us a News Tip

  • This field is for validation purposes and should be left unchanged.
  • Max. file size: 244 MB.

Sign in

SIGN IN

Bio

Ethics statement

Extract the signal from the noise

Get SiliconANGLE updates and analysis.

Contact us

Partner with us

Contact us

Guest inquiry