- Voltage, phase, frequency and connection
- Breaker loading and dual power paths
- rPDU, busway, UPS and generator capacity
- Protection and monitoring interfaces
Engineering calculator · Updated July 2026
How Much Power and Cooling Does an AI GPU Rack Need?
AI racks can span from conventional enterprise densities to more than 100 kW per rack. The useful answer is not a universal threshold: it is a calculation based on the maximum input of the actual server system, auxiliary loads, design headroom and the fraction of heat captured by liquid.
Planning calculator
Convert the equipment plan into facility loads.
Use the maximum input power for one complete server or rack system—not GPU TDP alone. Results are planning-level and require OEM and engineering confirmation.
The first distinction
Capacity design and energy forecasting are different calculations.
Capacity chain
Rack design IT load
= (maximum system input × quantity + auxiliary load)
× (1 + explicit design headroom)This value informs rack power paths, rPDU and busway selection, UPS capacity and cooling design. It should not be reduced because the GPUs are expected to average 60 or 70 percent utilization.
Energy chain
Expected facility energy
= design IT kW × expected average load
× operating hours × expected PUEThis second chain is useful for energy, operating cost and sustainability estimates. Keeping the two chains separate prevents an optimistic energy assumption from under-sizing critical infrastructure.
Cooling decision
Rack kW is a screening variable—not a universal liquid-cooling threshold.
Cooling architecture depends on more than heat density. The server manufacturer may require air cooling, direct-to-chip liquid cooling or a hybrid arrangement. The site must then match allowable inlet conditions, airflow, coolant temperature, water quality, redundancy and maintenance access.
Air coolingCheck required airflow, containment, fan energy and outdoor design conditions.
Hybrid air + liquidSize the liquid loop and the residual room air load separately.
Direct-to-chipCoordinate cold plates, manifolds, secondary loop, CDU, facility water and heat rejection.
Existing-site retrofitVerify electrical paths, floor loading, pipe routes, leak detection and service clearances.
Heat split
Liquid cooling does not eliminate the room air load.
If 80 percent of rack heat is captured by liquid, 20 percent still enters the room. Power supplies, network switches, storage and other components may remain air cooled. That residual load must be included in the in-row or room-cooling design.
Liquid-side heat = rack design IT load × liquid capture ratio
Residual air heat = rack design IT load × (1 − liquid capture ratio)The heat-capture ratio must come from the equipment architecture and project design—not a generic default.
Existing-facility readiness
Check the whole path before approving the rack.
- Airflow or coolant inlet conditions
- CDU and facility-water temperatures
- Heat rejection at design ambient
- Residual room cooling and humidity control
- Floor loading and rack anchoring
- Delivery and service access
- Pipe routing and leak detection
- Expansion footprint
- Redundancy during maintenance
- Alarm and controls integration
- Water-quality management
- Commissioning and acceptance tests
Common sizing failures
Five shortcuts that create expensive rework.
- Adding GPU TDP while ignoring the complete server.
- Using average utilization to size breakers or UPS capacity.
- Multiplying by PUE as if it were a UPS redundancy factor.
- Ignoring network, storage and residual air heat.
- Selecting a CDU without defining the facility-water interface.
AI rack readiness review
Have RackMade review the assumptions behind your result.
Send the server schedule, rack count, site location, utility conditions, redundancy target and deployment date. We will identify missing inputs and the next engineering decision.