Power density planning for GPU racks is the work of sizing power distribution, cooling, and floor capacity for racks that draw far more than conventional design assumes. Most halls were laid out for 5 to 10 kilowatts per rack, a figure that has held for general-purpose compute for years.
A densely populated GPU rack draws 40 kilowatts and upward, with current accelerated platforms pushing past 100. That is not an increment. It means a single rack can consume what an entire row was provisioned for, and every downstream assumption — breaker sizing, busway capacity, airflow, floor loading — needs to be recalculated rather than adjusted.
What Power Figure Should You Plan Against?
Plan against sustained draw under the real workload. Nameplate ratings on power supplies describe maximum capability, not consumption, and sizing against them overprovisions substantially. Sizing against idle figures underprovisions catastrophically. Neither is the number to plan with.
The figure that matters is sustained draw under the workload the hardware will actually run. Training workloads hold accelerators near maximum for hours; inference is burstier and lower on average. A rack sized for inference and then used for training will trip breakers.
Ask the vendor for measured consumption at realistic utilization for the specific configuration, not the platform family. Accelerator count, memory configuration, and cooling option all move the number materially.
Then add headroom for transients. GPUs exhibit sharp power excursions during workload transitions, and those spikes are brief, large, and entirely capable of tripping a breaker that the average draw sits comfortably beneath. Circuit protection sized under the National Electrical Code (NFPA 70) responds to instantaneous current, not to a fifteen-minute average.
Can Your Power Distribution Reach the Rack?
Capacity at the utility feed is not capacity at the rack position, and the constraint is usually somewhere in between.
Trace the whole path: utility service, switchgear, UPS, power distribution units, busway or whip, and the rack PDU. Each element has a rating, and the lowest one governs. A facility with abundant utility capacity may have busway that cannot deliver 40 kilowatts to a single position.
High-density racks generally require three-phase power at the rack, with connector types and amperages that differ from the single-phase distribution common in older halls. This means new whips, possibly new busway taps, and electrical work scheduled well before equipment arrives.
Redundancy multiplies the requirement. A rack requiring 2N power at 60 kilowatts needs 120 kilowatts of provisioned capacity, and the redundant path must independently carry full load. Designs that share capacity across paths fail on the day the design is tested.
UPS runtime falls as load rises. A battery plant sized for fifteen minutes at the old density delivers a fraction of that at the new one, which changes what the generator start sequence has to achieve.
How Do You Cool a 40-Kilowatt Rack?
Not with conventional room cooling alone. Power delivered becomes heat, and removing 40 kilowatts from a single rack is beyond what room air can do.
Air cooling has a practical ceiling in the range of 15 to 25 kilowatts per rack even with well-managed containment, and reaching the upper end requires disciplined hot and cold aisle separation, blanking panels in every empty position, and sealed floor penetrations. Above that, air runs out regardless of how much of it is moved.
Rear-door heat exchangers extend the range considerably by removing heat at the rack before it enters the room. They require facility water to the rack position, which is a plumbing project rather than an electrical one.
Direct-to-chip liquid cooling is where the highest-density platforms are going, as the Open Compute Project's cooling work reflects, and it changes the facility requirements fundamentally: coolant distribution units, secondary loop piping, leak detection, and the maintenance implications that follow. Anything installed this way carries those constraints for its whole life, including when it eventually has to be drained and relocated.
Where the density exceeds what an existing hall can support, the answer is usually a data center relocation to a facility designed for it rather than a retrofit. Plan cooling capacity against the same sustained draw figure used for power, and confirm the room can reject the heat rather than simply move it around.
Weight and Space
Dense racks are heavy, and the weight arrives with the power.
A fully populated GPU rack can exceed 3,000 pounds, and liquid-cooled configurations add more. That pushes directly against raised floor and structural limits, and the concentrated load under four casters is the specification most likely to be exceeded.
Depth and clearance often exceed what older racks and aisles were built for. Deeper chassis need deeper cabinets, and service access for liquid-cooled equipment requires more room at the rear than a cable-managed air-cooled rack.
Counter-intuitively, high density frequently means leaving positions empty. A row whose power and cooling supports 40 kilowatts per rack cannot fill every position if the room total is fixed, and the layout ends up with deliberate gaps. Planning for this in advance avoids buying racks that cannot be energized.
Commission Before You Rely On It
The assumptions above should be tested before production workloads depend on them.
Load bank testing at the rack position validates that the power path delivers what the design claims, under sustained load, with the redundant path proven independently.
Thermal validation under real workload confirms the cooling works at the density installed. Inlet temperatures checked against ASHRAE TC 9.9 guidance, and component temperatures under a sustained job, will find the recirculation problems that a walkthrough does not.
Instrument for the long term. Per-rack and per-circuit power monitoring, and inlet temperature sensing at the rack, turn capacity from an estimate into a measurement — which matters because the next generation of accelerators will ask for more, and the planning conversation will start from whatever data exists.