Eight schematic views / campus to memory

Inside an AI data center, from power grid to GPU memory

Follow power, heat, and data through a hypothetical campus and a documented GB200 NVL72 rack. Drag each 3D model to rotate it; use +/− to zoom and Reset view to return. Views are not to scale. Select a number for part details or a red ! for a common mistake.

Campus → row: scenario1 GW, hall size, pod layout, and GPU total are examples, not one built site.
Rack → HBM: GB200 referenceCounts and rack power follow NVIDIA's hardware guide; shapes are simplified.
Cooling flow: calculated~146 L/min assumes 85% of 120 kW into liquid and a 10 K coolant rise. It is not a vendor flow setting.
This interactive page requires JavaScript and WebGL.
GB200 NVL72 / as of Sep 2026
GPUs
72
Power
~120 kW
Liquid
~146 L/min model
Mass
~1.32 t
Detail
Reference →
POWER · ~120 kW → busbarCOOL · ~146 L/min model ⇄ manifoldDATA · 130 TB/s NVLink domain
03 / Rack18 compute trays inside one NVLink domainORv3 envelope proxy · GB200 contents sourced
Schematic · not to scale
1
2
3
4
1 GW · all-GB200 arithmetic example
  1. 1,000 MW ÷ PUE 1.2 = 833 MW IT
  2. 833,333 kW ÷ 140 kW per rack* ≈ 5,952 racks
  3. 5,952 × 72 ≈ 428,544 GPUs
1 / 3
*140 kW = 120 kW rack + 20 kW allowance. Assumes all IT power serves these racks; omits storage and other loads. This is not a site forecast.
1
2
3
4
1
2
3
4
1
2
3
4
Rack generations · as of Sep 2026
SystemGPUs / NVLink domainRack powerCoolingHBM / GPURack massVolume status
DGX H100 ×48 per server≤40.8 kW for four serversair80 GB HBM3not publishedshipping by 2023
GB200 NVL7272~120 kW~85% liquid186 GB HBM3E~1.32 tshipping by 2025
GB300 NVL7272not published hereliquid288 GB HBM3Enot published hereshipping by 2026
Vera Rubin NVL7272not published hereliquid288 GB HBM4not published hereproduction ramp in 2026

Four DGX H100 servers make four 8-GPU domains. NVL72 systems make one 72-GPU rack domain.

1
2
3
4
1
2
3
4
1
2
3
4
1
2
3
4

Common Mistakes