On the die
Direct contact on GPU, HBM, and power MOSFETs. Copper micro-fin channels, quick-disconnect fittings with dripless mating. Leak detection at every junction.
Cold plates on the chip. Nothing else matters.
Direct-to-chip liquid cooling at rack level, with the GB300 NVL72 reference architecture validated before your cluster arrives. The cooling circuit is closed, the water is ours, and the cold plates are already mounted on the dies.
Liquid cooling
Racks · CDU · dry coolers
Cold plates on the chip, CDU heat exchangers, dry coolers at the perimeter. Cyan loop animates when the figure enters view; hotspots pulse rack → CDU → rejection.
FIG. COOL · Closed primary loop · warm water 45°C supply · no evaporation on site
45°C
Supply temp
Warm water loop, high delta-T, efficient heat rejection.
136kW
Per rack
GB300 NVL72 with cold plates on every die.
1.15–1.25
Target PUE
ASEAN average sits near 1.55.
0L
Municipal make-up
Closed loop; reservoir owns the budget.
2N A / B loops
N+1 redundant CDUs. 1.2 MW per unit, 8 per module. Solid-state pumps, variable speed, pressure-compensated flow.
Dry coolers
Warm water (45°C supply, 50°C return) rejects heat to ambient air. No chillers, no compressors.
Air wall · N+2
Hot aisle sealed, ceiling return. Cold aisles open. Positive pressure in the hall, ISA 71.04 G1 corrosion rating.
GB300 NVL72 Service
The rack is validated before your cluster arrives
REF · Module cooling topology for 32 × GB300 NVL72
Reference architecture
72 GPUs per rack, cold plates mounted on the die, 136 kW per rack. The rack is assembled, tested, and validated before your cluster arrives. You bring the servers, we provide the rack and the power.
Service model
Choose turnkey delivery with the rack already commissioned, or pick up at white-floor and let your team do the fit-out. Either way, the reference architecture is DSX-validated and Tier III concurrent maintainable.
800 V DC tap
45°C supply
600 kW pipe reserved
Compatibility
Beyond B300
GPU compatibility — GB300 NVL72, B200, H100/H800 with adapter cold plates. Custom sleds for non-NVIDIA accelerators.
Network — Quantum-X800 NDR InfiniBand, 400 GbE. Spine-leaf topology, oversubscription ratios configurable.
Cooling — 45°C supply, 50°C return. Compatible with all major CDU manufacturers.
Power — 800 V DC busway, 800 A per rack. PDU shelves drop to 48 V DC at the rack.
Rack envelope — 2.1 m tall, 1.2 m wide, 1.2 m deep. Fits standard raised floor or slab-mount.
Commissioning — Factory acceptance test, site acceptance test, burn-in cycle, warm handover.