ANET

Walk into the networking room of a hyperscaler data center — a space separate from the server halls, often climate-controlled independently, where rows of network switches are stacked in racks connected by thousands of fibre optic cables. Each cable carries data at 400 or 800 gigabits per second — soon 1.6 terabits. The switches directing that traffic are overwhelmingly Arista.

The physical architecture follows a hierarchy. At the bottom, "leaf" switches sit in each server row, connecting individual servers to the network — one switch might connect 32 or 64 servers. Above them, "spine" switches connect the leaf switches to each other, creating a fabric where any server can reach any other server. In the largest AI clusters, there may be a third "super-spine" tier connecting multiple spine pods together. Arista makes switches for every tier.

For AI-specific clusters — where GPUs communicate via the data center's Ethernet network (as opposed to NVLink, which operates within a rack) — Arista's switches handle the "east-west" traffic: data flowing between GPU servers during distributed training. This traffic is enormous. In a training run, every GPU needs to share gradient updates with every other GPU after each computation step. The speed and latency of Arista's switches directly affect how fast the model trains. Meta built a 24,576-GPU AI cluster on Arista's switching fabric. Three of Arista's four hyperscaler AI customers have deployed more than 100,000 cumulative GPUs on Arista Ethernet, with the fourth migrating from InfiniBand.

GPU clusters require dramatically more networking than traditional cloud workloads. A traditional cloud server might need one or two 100G network connections. An AI GPU server needs multiple 400G or 800G connections — one for each GPU's network path. The networking spend per GPU is 3-5x higher than networking spend per traditional server. Arista's AI networking revenue target of $3.25B for 2026 — up from near-zero three years ago — is a direct consequence of this multiplication.

Arista is a pure-play data center networking company. Full year 2025 revenue was $9.0B, with core cloud/AI products at 65% of revenue, campus and routing adjacencies at 18%, and subscription software and services at 17%. The company had approximately 5,200 employees exiting 2025 and more than 10,000 cumulative customers. Cloud and AI titans contributed 48% of 2025 revenue, enterprise and financials 32%, and AI and specialty providers (including Oracle and emerging neo clouds) 20%.