Planning · Power and cooling
Plan the power and cooling before the cluster arrives.
Most server rooms weren’t built for the power and cooling a GPU cluster needs. We help you plan for these requirements, so your cluster runs the day it arrives.
Why it matters
GPU clusters have outgrown most server rooms.
In 2026, the typical data center rack runs at about 7.8 kW, up just 0.3 kW from a year earlier. For example, a cluster of four 8-GPU servers draws about 57 kW, more than seven typical racks’ worth, before you add the network switches, storage and head node. Rack space is rarely the problem. Power, cooling, weight and depth usually are.
The good news: every one of these is solvable when it’s planned early. We review your space, specify what you need, and work with your facilities team and electrician, so there are no surprises on install day.
The limits, and how we plan for them
What a GPU cluster asks of your room.
Size the circuits to the cluster
Each 8-GPU server draws about 14.3 kW, and a 208 V, 30 A circuit gives you about 5 kW you can use. For four servers, that means four 208 V, 60 A three-phase circuits per side, plus power for the network, storage and head node. We specify every circuit in writing your electrician can work from.
Make redundancy actually redundant
With A and B feeds, each side has to carry the whole cluster on its own. We size every feed for the full load, so a failure on one side doesn’t take your cluster down.
Plan the cooling, not just the power
Nearly every watt becomes heat: about 16 tons of cooling for four 8-GPU servers, before the rest of the cluster. We work out the load with your facilities team and recommend the airflow, containment or liquid cooling to match.
Plan by capacity, not rack units
Four 10U servers fit in one 42U rack, but together they draw about 57 kW, more than seven times what a typical rack is built for. We lay out the cluster across racks around power, cooling, weight and depth.
Planned before the servers arrive
From site review to power-up, we work through every limit with you before you order, and we rack, cable and power up the cluster on site.
Sources: server power from NVIDIA DGX B200; size and weight from the NVIDIA DGX B200 User Guide; typical rack density from the Uptime Institute 2026 Global Data Center Survey, with the 2025 comparison from Data Center Knowledge. Figures are for planning; we confirm them for your configuration.
Before you order
Questions for your facilities team.
The answers help us size everything correctly. Don’t have them all? That’s what a site review is for, and we’ll find them with you.
Power
What feeds each rack?
Voltage, amps, single- or three-phase, and how many circuits. Are there separate A and B feeds, and can the UPS and generator carry the new load?
Cooling
How much heat can the room remove?
Cooling capacity per rack and for the whole room, in kW or tons, and how much of it is already in use.
Airflow
Where does the hot air go?
Whether the room has hot and cold aisles, and whether they’re contained. Modern compute and GPU nodes move a lot of air, front to back.
Floor and racks
What can they hold?
Load ratings for the floor, especially a raised floor, and for the racks themselves. Four GPU servers can weigh over 1,200 lb.
Rack depth
Will the servers fit inside?
The usable depth with the doors closed. GPU servers run about 35 inches deep before power and network cables.
Getting it in
How will it reach the rack?
Doorways, elevators, the loading dock, and whether a server lift is available on site.
When the room isn’t ready
There’s almost always a way.
A room that can’t take a modern cluster today usually can with the right plan. These are the options we work through with you.
Spread the load
Fewer servers per rack, so each rack stays within the power and cooling it actually has.
Upgrade the circuits
208 V, 60 A three-phase feeds sized for each server, with separate A and B sides.
Add targeted or liquid cooling
Rear-door heat exchangers, in-row cooling, or direct liquid cooling for the densest racks. We’ll recommend what fits, and can help source it if you’d like.
Use a colocation facility
A high-density colocation facility can supply the power and cooling your room can’t. We’ll plan the install there just as we would in your own room.
Right-size the hardware
A 4-GPU or PCIe GPU configuration may deliver what you need on the power you have today.
Phase it in
Start with what your room supports now, and add nodes as power and cooling upgrades land.
Planning a GPU cluster? Start with your room.
Engineering the future of HPC and AI
info@advancedclustering.com(866) 802-82223148 Roanoke Road, Kansas City, MO 64111