6 steps, 4 teams: how a bare rack becomes a live AI cluster.
The rack, a compute tray, the NVLink spine, power, liquid cooling, rack-to-rack networking, storage, switch-on and the full hall, in 3D.
This walkthrough needs WebGL, which this browser has turned off.
Procurement, assembly, delivery, install of GPU racks
Fleet-scale GPU cluster provisioning, multi-tenant
One control plane over Ethernet + InfiniBand + DPUs
Blackwell / Blackwell Ultra / B200, validated configs
Click a stage above to see what happens, who runs it, and the key facts.
Same 4 layers every time — the step you pick lights up where it happens and drops in the key numbers.
An illustrative reference model of a typical AI-factory build-out, assembled from publicly documented industry patterns: NVIDIA reference architectures, the Kubernetes GPU Operator, Slurm-on-Kubernetes and GPU network-automation practice. Roles are shown generically; they do not describe any specific customer, partner, vendor contract or project. Simulations are simplified for explanation, not measured results.