Server & Rack Integration
The companies that physically build the boxes hyperscalers buy.
What this layer does
Nvidia ships GPUs; someone has to build them into a server, plumb in liquid cooling, certify the rack, and roll it onto a truck. This is what server OEMs and Taiwanese ODMs do — and at GB200 NVL72 scale, “the rack” is now the actual unit of sale, weighing 1.5 tons and costing ~$3M each.
The economics here are thin (single-digit operating margins) but the volume is huge, and integration complexity is rising fast. The structural debate is: do ODMs (Foxconn, Quanta, Wiwynn) keep eating share from US OEMs (Dell, HPE) as hyperscalers move toward whitebox? Super Micro is the chaos variable.
Sub-categories
Brand-name servers sold with support contracts. Enterprise buyers, some neoclouds, sovereign clouds.
The actual factories behind hyperscaler servers. Lower-margin but enormous volume.
Companies that assemble full NVL72-class liquid-cooled racks. Overlap with OEMs and ODMs, but a distinct skill set.
Electronics manufacturing services building the boards going into AI servers and switches.
Server-equivalent layer for AI storage — high-throughput NVMe arrays, parallel file systems.