Arista Networks Sees Shift from Nvidia to Mixed Accelerators as AMD Gains 20–25%

Arista Signals Changing Accelerator Mix

Networking equipment maker Arista Networks is seeing a tangible shift in the mix of AI accelerators customers deploy, with company executives describing a move away from near-total reliance on one vendor toward a more heterogeneous stack. Jayshree Ullal, Arista’s president and CEO, tells investors that deployments which were “pretty much 99% Nvidia” a year ago now include roughly 20% to 25% AMD accelerators in new installations. The change reflects broader architecture decisions by hyperscalers and cloud providers as they balance performance, cost and vendor lock‑in.

Networking Strategy Shifts as AMD Gains Ground

Arista’s Ethernet switches are central to tying together high‑performance AI chips for training and inference, so changes in accelerator selection carry direct implications for its product design and interoperability work. The company is adapting its software and hardware compatibility to support AMD accelerator traffic patterns and integration requirements while continuing to optimise for Nvidia’s dominant footprint, which still represents the majority of the market. Arista’s emphasis is on ensuring low-latency, high-throughput fabric performance across mixed-accelerator clusters to meet evolving customer designs.

The push toward AMD follows moves by Nvidia to develop its own networking stack, Spectrum‑X, which is winning design engagements at large cloud customers such as Meta and Oracle and prompting some organisations to specify alternative networking and co‑design approaches. Arista is responding by diversifying partnerships and co‑developing customised AI cluster references — including a recently announced collaboration with AMD — to keep its platforms relevant as customers adopt multi‑vendor accelerator strategies.

Alliances and Architecture

Arista formalises its AMD collaboration to produce tailored AI cluster configurations, aiming to give cloud operators and enterprises validated, interoperable options beyond single‑vendor stacks. The company positions these engineered references as a way to shorten deployment time and reduce integration risk when customers mix accelerators and networking fabrics.

AI Demand and Competitive Pressure

Executives point to the rapid acceleration of AI workloads since the launch of OpenAI’s ChatGPT and note additional competitive dynamics from Google’s Tensor Processing Units as factors prompting heterogeneity. Industry analysts characterise the shift as significant but manageable for Arista, saying Nvidia is being designed in “less often, if not fully, to a degree,” while the overall market continues to expand.