Solutions
Building and upgrading AI data centers
Three build scenarios, one systems-engineering approach.
Liquid-cooling retrofits of existing halls
Container data center clusters
AI campuses
8 MW reference design
8 MW total supply covering IT, cooling and auxiliary loads. It illustrates how the systems combine and how capacity is sized; it is not a standard product specification or a delivery commitment.
3POD6 SUs deployed in phases
3072GPUs
6.09MWIT load
1.204annual PUE (design estimate)
CDU6, one per SU
Secondary loop34/44 °C (93/111 °F), about 401.2 m³/h (1,766 GPM) site-wide
Liquid-cooled load4.67 MW
Dry coolers7 × 1 MW, primary loop 32 / 42 °C
Chillers4 × 2,000 kW, 3 duty + 1 standby
Compute network800G, 8-rail
Switches32 leaf, 16 spine
TechnologyInfiniBand or RoCE
Shared storage10 PB, tiered
Tiers4 PB all-flash, 2 PB warm, 4 PB archive
Annual PUE1.204
Summer PUE1.237
Peak (design-day) PUE1.265
Peak total power (design day)about 7.70 MW
B300, 4 per rack3,072 GPUs, 168 racks, 61.8 kW per rack, 6.27 MW IT
B300, 8 per rack3,072 GPUs, 84 racks, 119.8 kW per rack, 6.09 MW IT
GB300 NVL722,304 GPUs, 56 racks, 140 kW per rack, 4.51 MW IT
Design estimates for the reference configuration, not measured project results.