Arista Networks (ANET) Wells Fargo 8th Annual TMT Summit summary

AI networking market outlook and revenue targets

• Committed to $750 million AI back-end network revenue for fiscal year 2025, with an additional $750 million expected from front-end network pull-through, totaling $1.5 billion.

• AI back-end networks require undersubscribed, high-capacity architectures to support 100% GPU throughput for extended periods.

• Networking spend as a ratio of total GPU cluster cost has decreased due to rising GPU costs, but absolute networking requirements have grown.

• Optics and cables are significant cost drivers; efforts focus on reducing CapEx and OpEx through proximity and reliability improvements.

• Engineering, architecture, and multi-standard interoperability are key differentiators in delivering high-performance AI networks.

Scaling, technology evolution, and cluster architecture

• AI clusters are rapidly scaling from 10,000 to 100,000+ GPUs, with future growth limited by power and space constraints.

• Large-scale deployments may shift to interconnected pods or islands, with synchronous and asynchronous parallel technologies.

• Technology is advancing to support 400Gb, 800Gb, and eventually 1.6T per GPU, with current systems already 800Gb ready.

• The distinction between AI and traditional data center networking is blurring as similar products serve both.

• Connecting geographically dispersed clusters is becoming necessary due to physical limitations.

Ethernet vs. InfiniBand and Ultra Ethernet Consortium

• Ethernet is favored over InfiniBand due to supply chain diversity, lower cost per bit, and broad industry support.

• Ultra Ethernet Consortium (UEC) is developing enhancements like Ultra Ethernet Transport (UET) to improve scalability, reliability, and integrated security.

• UEC's first specification is expected in Q1, with many features available as software upgrades to existing products.

• Ultra Ethernet aims to lower NIC costs, improve load balancing, and enable secure, multi-tenant AI workflows.

• Growth in AI networking is projected to come from Ethernet, with InfiniBand expected to remain flat or decline.