Rack-Scale System

A rack-scale system is a pre-integrated AI computing unit designed, shipped and operated as one rack, not as separate servers. Its GPUs, CPUs, networking, power delivery and liquid cooling are all designed into that single rack. NVIDIA's GB200 NVL72 is the best-known example: 72 GPUs linked by NVLink in a single rack drawing roughly 120-132 kW, far above the 5-10 kW of a traditional enterprise rack. Because the rack is the unit of purchase, data centers must supply direct-to-chip liquid cooling, busway power at that density and floors able to carry the weight, so facilities are now marketed as "rack-scale ready." Next-generation platforms such as Vera Rubin push density higher still, which is why new AI buildings are specified around liquid cooling from the design stage.