Part of the day: Rack-scale systems

Assembled by the nightly build from this catalog’s own entry for Rack-scale systems (4.01, Layer 4 — Servers, racks & integration). The tables below are rebuilt every night; first published 2026-09-17, refreshed 2026-09-24.

A rack-scale system is the fully integrated cabinet of compute trays, switch trays, power shelves, cooling manifolds and busbar that ships as one wired unit rather than as discrete servers.

The rack-scale system is the top-level assembly layer: it takes the GPU baseboards (4.03), switch trays (4.04), power shelves (4.12) and liquid-cooling hardware from layer 5 and integrates them into a single cabinet wired for one scale-up fabric domain (layer 3). Nvidia and AMD publish the reference architecture and mechanical/electrical spec; the ODMs and OEMs listed against this part build to that spec, badge the result, and compete mainly on integration quality, lead time and support rather than on the underlying design. AI training and inference clusters push this part into existence at all: a conventional enterprise server rack has no equivalent, because the point of a rack-scale system is to turn dozens of GPUs into one coherent domain, something only large models require. For investors, the interesting split is between the reference-design owners (Nvidia, AMD), who capture the architecture value, and the ODM/OEM tier, which competes on manufacturing scale and qualification speed with each new rack generation. Each new generation resets the qualification race, so watch which ODMs get design wins early and which fall behind on ramp.

AI delta: AI training and inference workloads are the reason rack-scale systems exist at all — no other server class assembles dozens of GPUs into one wired, liquid-cooled domain.

Who makes it

The full entry — sub-categories, news, videos and sources — lives on the Rack-scale systems page.