Arista Networks up 54%, beating Nvidia and Cisco in AI networking

Arista Networks is positioning itself as a way for investors to benefit from AI infrastructure spending without owning the GPU maker itself. Its Ethernet switches and software address a bottleneck that grows as data centers add more processors: moving enormous amounts of data between chips, servers, and storage. In June it introduced the 7060XE7 Series, which supports 1.6 terabits per second per connection and up to approximately 100 terabits per second of total system bandwidth, and claims linear pluggable optics can cut interconnect power use by about 60% versus traditional pluggable optics. On Oct. 7 it announced a rack-scale architecture developed with AMD, Broadcom, Meta, Microsoft, and Qualcomm, designed to connect up to 144 AI accelerators in a single-hop network and scale to 1,024 accelerators across racks. That shift extends Arista from linking separate servers and racks toward letting large groups of AI processors communicate more closely over Ethernet, which could give cloud providers more flexibility in choosing equipment. CEO Jayshree Ullal described networking as the central nervous system for infrastructure, and the company's latest results — $3.03 billion in second-quarter revenue, up 37.7% year over year — suggest customers are responding. Risks remain, including competition from Cisco and proprietary networking technologies, and the chance that major cloud customers build their own gear.