Arista Networks has launched a new family of routing platforms built for a problem it says is coming next: AI inference spread across many data centers instead of concentrated in a few giant ones. The Etherlink R4 lineup was announced on 17 November, timed to the Supercomputing 2026 show, where the company will display it

2

3

1

.

Why inference is spreading

Arista frames the spread of inference as driven by three things. Agentic AI chains many model calls into a single request, so latency compounds across the chain. Regulators increasingly require data to stay inside national borders. Companies, meanwhile, are putting inference wherever power and space allow

3

1

.

Brendan Gibbs, Arista's vice president of product line management, said the forces that spread training out will do the same to inference. Roughly 30% of Arista's AI business is already scale-across, he said, which he offered as evidence that distributed AI is becoming ordinary rather than experimental

1

. By Kerravala's account the company ranks among the biggest winners of the AI back-end buildout, with customers ranging from hyperscalers to neoclouds to enterprises building their own local AI

1

.

What Arista is shipping

The R4 lineup spans four product categories: compact 7020HR4 and 7020R4 routers, a 7282R4 with dual supervisor engines for redundancy, and a 7280R4 built for coherent transport

2

3

1

. A pluggable optical line system moving 25.6 Tbps rounds out the family

2

3

1

. All of it is available now apart from the 7020HR4, which ships in early 2027

2

3

1

.

Source: SiliconANGLE

Source: SiliconANGLE

The platforms ship with integrated MACsec and IPsec encryption, multi-tenancy, segment-routing virtual private network services, traffic engineering for AI workloads and network timing

2

3

1

. The 7280R4 also supports 400 Gbps and 800 Gbps coherent optics

2

3

1

. The company says the new chips cut power consumption by up to 60% compared with alternatives built on older 7nm or 16nm silicon

2

3

1

. One outlet identifies that silicon as Broadcom Jericho and Qumran

1

. Per the company, the platforms let operators scale distributed inferencing across geographies with predictable, low-latency performance, including keeping sensitive inference traffic inside national borders

3

1

.

Kerravala frames the launch as Arista going downmarket. Its recent announcements centered on dense 800G and 1.6T systems, and the company previously lacked a viable option in this smaller, cheaper front-end tier

1

. The 7280R4, he reports, was co-designed with a hyperscaler looking to bridge older gray optics to coherent ones

1

. The entry-level routers' 400G uplinks may look like overkill today, though in his view the headroom is cheaper than a forklift upgrade later

1

. The optical line system, though, is resold rather than Arista-built, since Arista is not an optical vendor

1

.

A new customer, and a bet

The hardened 7020HR4 is built for 5G and AI radio-access-network cell sites and outside-plant deployment, which points at telcos. Telcos have not historically been a core Arista market, so per Kerravala any carrier business the 7020HR4 wins would be additive to what the company already has

1

. He thinks the telco play likely works better outside the US, naming BT, NTT and Telstra as carriers better positioned for it

1

.

Early customers sounded convinced. "We need a compact platform that supports the services our business and carrier customers rely on, and these products look like a good fit for that role," said Jon Vradenburg, chief technology officer of Douglas Fast Net

2

. Takayuki Sakou, general manager of Japan's STNet, cited expected higher throughput and improved power efficiency "while maintaining the quality of the EOS software we trust"

1

.

Kerravala says AI networking's back end is largely settled and its front end is not; Arista, in his telling, is betting its routing credibility can make it the default there

1

. He expects the bet to pay off

1

.

© 2026 TheOutpost.AI All rights reserved