Introduction

An architectural revolution is occurring in AI computing which is shifting from node-based servers to integrated rackscale systems. Examples include Nvidia’s GB200 NVL72, which combines 72 Blackwell GPUs and 36 CPUs in a single rack and Amazon’s Trainium 2 UltraServer, which spans two racks and contains 64 Trainium 2 chips. This move towards vertical integration and rack-level computing units is driven by the need to maximise performance through tightly coupled systems.

This report provides an overview of the AI server scale-up network ecosystem, highlighting the key players, both big tech and start-ups, that provide crucial connectivity silicon – from chiplet interconnects to server interconnects - to enable high-performance rackscale based AI computing systems.