01 / THE RACK
Inside an AI rack.
The GB200 NVL72 connects 72 GPUs within one rack. Its structure makes the relationship between compute, networking, and cooling visible.
FROM TRAY TO RACK
18 trays × 4 GPUs = 72 GPUs.
Each compute tray holds two GB200 superchips. Each superchip pairs one Grace CPU with two Blackwell GPUs. Nine separate switch trays connect the rack.
Drag to turn · scroll to zoom · choose a chapter to look inside
System boundaries & procurement considerations
The scale-up domain ends at the rack. The NVLink fabric provides each GPU high-bandwidth access across a 72-GPU domain while each GPU retains its own HBM. Traffic leaving that domain moves to the scale-out network, whose media, topology, and performance are procured separately; the network-tax page models that boundary.
Power is the binding constraint. Roughly 120 kW in a single rack is why liquid cooling is a requirement rather than an option, and why the binding constraint in procurement migrated from chip allocation to facility qualification. Two suppliers quoting the same GPU can differ entirely in whether their building can actually run this object.
The rack can be a commercial unit, but it is not the only one. Providers may sell full racks, nodes, instances, or capacity blocks. Ask which physical assets, failure domains, and acceptance criteria correspond to the quoted unit.
schematic geometry · 81 fallback arcs and 378 3D cable instances represent the vendor's approximately 5,000-cable spine · original hardware reference: July 2026 · memory corrected 16 October 2026: NVIDIA specifies 372 GB per two-GPU GB200 superchip, or 186 GB per GPU · NVIDIA GB200 NVL72 · NVIDIA NVLink system guide · methods