In the AI back-end fabric that ties tens of thousands of accelerators, Ethernet is pushing out InfiniBand — and Arista sells that system
Arista is the company responsible for connecting AI clusters. It does not build chips; it bundles the switch silicon made by Broadcom and Marvell with an operating system and sells it as a system. It is the primary beneficiary of open Ethernet pushing out closed InfiniBand in the AI back-end fabric. Two risks remain: concentration in two cloud titans that make up about 40% of revenue, and the chance those titans take the network in-house too.
FQ1 FY2026 revenue of $2.709B (+35.1% YoY) beat consensus, and full-year revenue guidance was raised to $11.5B (+~28%). At the same time the 2026 AI networking revenue target was lifted to $3.5B. Non-GAAP operating margin was 47.8% and billings growth accelerated to 54%. With the transition from 800G to 1.6T/3.2T Ethernet and the Ultra Ethernet Consortium standardization underway, AI back-end networking emerged as a standalone investment axis.
Connection power — Arista (#1 AI back-end Ethernet systems, 7800 universal AI spine). Component supply — Broadcom (Tomahawk/Jericho ASICs), Marvell, Credo (connectivity), with Arista as the system exit for that silicon. Offsets — customer concentration (two large customers about 40%, industry-estimated as Microsoft and Meta), hyperscaler in-house network-ASIC risk, optics/ASIC supply constraints, a demand air pocket during the 800G→1.6T transition, and an already-embedded valuation.
An AI cluster's network splits into two layers. The front-end handles user and external traffic. The back-end is the internal network where accelerators exchange data directly. The battlefield for AI performance is this back-end.
Why does the back-end matter? AI training has tens of thousands of accelerators split the same job simultaneously. The result one accelerator finishes must pass to another losslessly and without delay. If this connection clogs, expensive accelerators sit idle waiting on each other. The back-end fabric's bandwidth is cluster efficiency itself.
Here two standards collide. One is — NVIDIA's closed proprietary standard, bound to NVIDIA gear. It was long the standard for AI cluster connection. The other is — an open multi-vendor standard. The entire data center already runs on .
The flow tilts toward . The reason is simple. Hyperscalers dislike being locked to one vendor. Open Ethernet gives multi-vendor procurement and pricing leverage. As cluster scale grows, the premium of the closed standard becomes a burden and the appeal of the open one grows. That is the backdrop to Arista management declaring Ethernet the 'eventual winner.'
The AI back-end fabric is a battlefield of Ethernet versus InfiniBand, and the larger the scale, the more weight tilts toward open Ethernet. Arista is the #1 system player in that Ethernet camp.
Sign in with Google to unlock the full body of every free report instantly.
Sign in with GoogleThis site runs on ads — the tier system rewards community contributions.
This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.