When to use: Estimate the power density (kW/rack) of a GPU or compute rack from server count and per-server nameplate wattage, then check that density against the rack's PDU/circuit capacity. AI/GPU racks commonly run 20-130+ kW/rack — far above the 5-10 kW/rack typical of traditional enterprise IT, which is the core driver behind the shift to liquid cooling.
This calculator estimates power density (kW per rack) for GPU/AI compute racks from server count and nameplate wattage, then checks that load against the rack's PDU circuit capacity — the two numbers that determine whether a rack can be air-cooled or needs liquid cooling assistance.
A traditional enterprise 1U server draws roughly 500-1000W, so a fully populated 42U rack of general-purpose compute lands around 5-10 kW. A GPU training server (e.g. an 8-GPU NVIDIA H100 SXM node) can draw 10+ kW by itself — put four to eight of those in a rack and density climbs to 40-80+ kW/rack, and next-generation rack-scale systems like NVIDIA GB200 NVL72 push a single rack past 120 kW. This step-change in density is the single biggest reason AI data centers require fundamentally different electrical and mechanical infrastructure than a traditional enterprise data hall.
Server power supplies are rated above typical operating draw to handle transient spikes (e.g., GPU boost states, simultaneous training-job ramp-up). A 15-25% design margin over nameplate/typical draw is standard practice for circuit and PDU sizing so the electrical infrastructure isn't running at its continuous limit under normal peak conditions.
The rack PDU circuit capacity is calculated using the NEC continuous-load rule: a circuit intended to carry current for 3+ hours continuously must be sized so the load doesn't exceed 80% of the breaker rating. Two redundant circuits (A+B feed) per rack is standard practice for concurrently-maintainable or fault-tolerant (Tier III/IV) facilities, letting either circuit alone carry the full IT load if the other fails or is taken down for maintenance.
Air cooling is limited by how much airflow can physically move through a rack and how much heat a given CFM of air can absorb per degree of temperature rise. Past roughly 30-50 kW/rack, the airflow required becomes impractical (loud, high fan power draw, hot-aisle temperatures too high) which is why direct-to-chip liquid cooling or immersion cooling becomes standard at higher densities — liquid carries far more heat per unit volume than air.
Nameplate power is the maximum rated draw printed on the server's power supply label — a worst-case ceiling, rarely sustained continuously. Design power for circuit sizing typically starts from nameplate (or a vendor-published typical/TDP figure) and adds a margin for headroom, since undersizing circuits based on optimistic "typical" power alone risks nuisance breaker trips under real peak load.
Redundant A+B power feeds let maintenance or a fault on one circuit/PDU/UPS path occur without dropping the rack's IT load — the other circuit alone must be able to carry the full rack load, which is why each circuit is typically sized for 100% of the design load, not 50/50 split.
Rack density (kW/rack) is the IT load side of the PUE equation multiplied out per rack. Higher rack density with liquid cooling can still produce a lower overall PUE than lower-density air cooling, because liquid cooling systems are often more energy-efficient per watt of heat removed. See the PUE Calculator in this studio.