A GB300 NVL72 rack pulls 132 to 142 kW nominal, but the number that actually governs a facility's electrical design is HPE's: provision busway for 192 kW per rack, not the nameplate figure. That gap between the number on the spec sheet and the number a facilities engineer has to sign off on is where most GPU rack power and cooling 2026 planning goes wrong, and it's the same gap that makes air cooling a non-starter above roughly 40 kW per rack.
TL;DR: GPU Rack Power and Cooling 2026 for GB300 NVL72
GB300 NVL72 draws 132-142 kW nominal per rack, HPE recommends provisioning busway for 192 kW of headroom, and direct-to-chip liquid cooling is mandatory, not optional, above 40 kW.
| Spec | Figure |
|---|---|
| Nominal rack power | 132 kW nominal, up to 142 kW maximum |
| Busway provisioning | 192 kW busway target |
| Peak draw (EDPp) | roughly 155 kW |
| Liquid vs. air heat split | about 90% liquid, 10% air |
| Facility water flow (25°C-45°C) | 59 to 177 L/min |
Spheron lists GB300 NVL72 as reservable capacity rather than an hourly on-demand rate; check current GB300 NVL72 availability before committing a facility to its own cooling retrofit.
GB300 NVL72 Power Draw: 132-142 kW Per Rack, By the Numbers
Three vendors publish three slightly different numbers for the same rack, and the spread matters more than any single figure. HPE's GB300 NVL72 spec sheet lists a nominal rack draw of 132 kW, with HPE recommending busway provisioned for 192 kW (EDPp, or Electrical Design Power peak) of headroom above that. NVIDIA's own reference design rates the rack at up to 142 kW maximum, with deployed peak draw (EDPp) running to roughly 155 kW depending on workload. Lenovo's rack-scale product guide lands in between at 135 kW TDP, with the same 155 kW peak ceiling.
| Vendor | Nominal / TDP | Peak (EDPp) | Busway provisioning |
|---|---|---|---|
| HPE | 132 kW | n/a | 192 kW |
| NVIDIA reference design | up to 142 kW | ~155 kW | n/a |
| Lenovo rack-scale guide | 135 kW TDP | ~155 kW | n/a |
The practical takeaway: a GB300 NVL72 rack's actual draw under load lands somewhere in the 132-142 kW band, but the number you design electrical infrastructure around is HPE's 192 kW. That headroom exists because GPU workloads don't run flat; large training jobs spike with the compute pattern, and a busway sized to the nominal figure with no margin will trip under a real burst. For context, NVIDIA's prior-generation GB200 NVL72 draws up to 120 kW nominal, so GB300 is roughly a 10-20% step up at the rack, not a step change.
HPE's provisioning guidance also specifies the electrical path itself: 400V or 480V, 60A whips feeding eight power shelves rated up to 33 kW each, laid out in N+N redundancy. At around 142 kW nominal per rack, roughly 7 GB300 NVL72 racks consume 1 MW of IT load, a figure worth keeping in mind when a facility quotes capacity in megawatts rather than racks.
Individual GPU draw isn't an NVIDIA-published number for this generation. AM Compute's site-evaluation guide puts a single B300/GB300 GPU at roughly 1,400W, about double the H100's 700W baseline. Across 72 GPUs, that difference alone accounts for most of the jump from H100-generation rack densities to GB300's 132-142 kW.
Direct-to-Chip Liquid Cooling: CDUs, Cold Plates, and the 90/10 Air-Liquid Split
Air cooling stops being viable above roughly 40 kW per rack, which puts GB300 NVL72's 132-155 kW draw three to four times past that ceiling. Direct-to-chip liquid cooling isn't a performance option here, it's the only way the rack ships at all.
The liquid path runs from the CDU through a rear manifold, through quick disconnects, to cold plates mounted on the CPUs, GPUs, ConnectX-8 NICs, and NVSwitch components. What stays air-cooled is everything that doesn't generate enough heat to justify a cold plate: the power distribution board, M.2 storage, and front I/O.
The cold plates themselves are a materials-engineering step up from prior generations. NADDOD's technical breakdown of the GB300 cooling system describes microchannels under 0.3mm wide, machined into 3D-printed titanium alloy rated to at least 10 bar, with coolant moving through those channels at 2-4 m/s. That channel width and flow velocity is what each cold plate needs to carry roughly 1,400W of heat off a single GPU, the same per-GPU figure AM Compute's power estimate lands on, without the plate itself becoming the bottleneck.
Facilities have three CDU (Coolant Distribution Unit) options depending on scale: an in-rack CDU rated up to 250 kW for single-rack deployments, an in-row CDU rated up to 1.8 MW that can serve up to 8 racks at a 40°C supply temperature (or 6 racks at 45°C under N+1 redundancy), and an L2A (liquid-to-air) sidecar CDU rated up to 200 kW for facilities without chilled water infrastructure. The choice between them is largely a question of how many racks a facility is committing to GB300 density versus hedging with mixed densities.
Facility water temperature drives flow requirements more than almost any other variable, and the swing is steep. Lenovo's spec sheet lists facility water flow at 59 L/min for a 25°C supply (2.3 psi pressure drop), rising to 71 L/min at 30°C (3.2 psi), 89 L/min at 35°C (4.9 psi), 119 L/min at 40°C (8.5 psi), and 177 L/min at 45°C (18.4 psi). That's roughly a 3x increase in required flow between the coolest and warmest supply temperatures in the range, and the pressure drop climbs even faster, nearly 8x. A facility running warm-water loops to cut chiller load pays for that choice in pump sizing and piping diameter, not just a bigger number on the flow meter.
What a Data Center Needs Before It Can Host a GB300 Rack
Power and cooling get the attention, but three other prerequisites decide whether a facility can actually take delivery of a GB300 NVL72 rack.
Floor loading. A fully loaded GB300 NVL72 rack weighs 3,300 to 3,483 lb in a 600 x 1,068 mm footprint, working out to an average floor loading of 478 to 505 lb per square foot. That exceeds what most standard raised-floor data centers are rated for, so a GB300 deployment often means verifying or reinforcing the slab before the rack ever arrives, not just clearing electrical and cooling capacity.
Commissioning order. NVLink switch trays have to power on and establish the NVLink fabric before the compute trays boot; booting compute trays out of that sequence forces a restart of the entire rack so it can rejoin the NVLink domain correctly, which turns a sequencing mistake into a lost deployment window rather than a quick fix.
Leak detection at scale. A GB300 NVL72 rack carries roughly 280-310 quick-disconnect liquid fittings, with 18 compute trays at 14 quick disconnects each accounting for 252 of them, plus additional connections on the switch trays. Each fitting is a potential leak point, which is why deployments run a three-level detection system at the node, rack, and data-center level rather than relying on a single sensor tier.
Multi-rack deployments add one more layer: power distribution commonly follows a "4-makes-3" busbar model, four independent 1 MW busbars arranged so any one busbar can be serviced without an outage, giving N+1 redundancy at the row level rather than just the rack level.
GPU Rack Power and Cooling 2026: Reserve GB300 NVL72 on Spheron Instead of Building a Cooling Loop
None of the work above is optional, and none of it is fast. Sizing busway to 192 kW, qualifying a CDU for the right rack count, reinforcing a slab to 500 lb per square foot, and getting the NVLink boot sequence right on a rack with 300 leak-checked fittings is months of facility engineering before a single GPU runs a job. For a team that needs GB300 NVL72 capacity for a model, not a data center retrofit, that timeline is the actual cost.
Spheron lists GB300 NVL72 on its pricing page with a Reserve designation rather than a published hourly rate, consistent with how every provider sells this rack today: GB300 NVL72 is a reservation or quote-only product industry-wide, with no on-demand rate published anywhere. What differs is where the capacity comes from. Spheron sources GB300 NVL72 from its certified data center network rather than a single owned facility, so a reservation isn't gated behind one partner's cooling-retrofit schedule. The tradeoff is the one worth being direct about: Spheron doesn't publish its own busway or CDU specs for the facilities hosting GB300, because the power and cooling engineering described in this guide belongs to the data center partner, not to Spheron's own infrastructure.
For the architecture and memory numbers behind the rack, the GB300 vs GB200 NVL72 pricing and availability comparison covers the spec gap in full, and the GB200 NVL72 guide has the prior generation's 120 kW baseline this post builds on. If you're deciding between the rack-scale GB300 NVL72 and a single Blackwell Ultra GPU, GB300 vs B300 breaks down when you actually need the full rack. And if your facility is the one weighing whether to build out for this density rather than rent it, NVIDIA's MGX 1.2 spec covers the 250 kW busbar and connector standard GB300 racks are built to.
Provisioning a 192 kW busway and qualifying a CDU for 142 kW of liquid-cooled compute is a facility project measured in months, not a line item on a GPU order.
Frequently Asked Questions
HPE's spec lists a nominal draw of 132 kW per rack, NVIDIA's own reference design rates it up to 142 kW maximum, and deployed peak draw (EDPp) can reach roughly 155 kW depending on workload. HPE tells facilities to provision busway for 192 kW of headroom regardless of the nominal figure, so budget power infrastructure to that number, not the nameplate one.
Yes. Air cooling stops being practical above roughly 40 kW per rack, and GB300 NVL72 draws 3 to 4 times that. Direct-to-chip liquid cooling handles about 90% of the rack's heat through cold plates on the GPUs, CPUs, ConnectX-8 NICs, and NVSwitch components; the remaining 10%, roughly 13-14 kW, goes to room air through the power shelves and front I/O that stay air-cooled.
It depends heavily on facility water temperature. Lenovo's spec sheet lists 59 L/min at a 25°C supply, climbing to 71 L/min at 30°C, 89 L/min at 35°C, 119 L/min at 40°C, and 177 L/min at 45°C, a roughly 3x increase across that temperature range. Pressure drop rises even faster, from 2.3 psi at 25°C to 18.4 psi at 45°C, which is why warmer-water facilities need larger pumps and piping, not just more of them.
Yes, as reserved capacity. GB300 NVL72 is a reservation or quote-only product industry-wide, with no published hourly on-demand rate anywhere. Spheron lists GB300 NVL72 on its pricing page with a Reserve designation and sources that capacity from its certified data center network, so the cooling retrofit described in this guide is the partner facility's engineering work, not something a renter configures.
