A breakdown of what actually determines a GPU rack's power draw — GPU count and generation, server density, networking overhead — and how that translates into the electrical and cooling capacity a data center design has to provide per rack.
Rack Power Is a Function of What's Actually Inside It
A GPU rack's power draw isn't a fixed number — it's the sum of every GPU, server chassis, networking switch, and piece of supporting hardware physically installed in that rack, multiplied by how hard that hardware is actually working. Two racks that look identical from the outside can draw meaningfully different power if one is running at full GPU utilization during active training and the other is partially idle. That said, for design purposes, electrical and cooling engineers size around the rack's rated maximum draw, not its average — a facility has to have enough capacity for every rack to run at full power simultaneously, because that's exactly what happens during a large, synchronized training job.
What Drives the Number Up
The single biggest factor is GPU count and generation. Each individual AI accelerator GPU can draw several hundred watts to over a kilowatt on its own under full load, and a dense GPU server can pack 8 or more accelerators into a single chassis. Multiply that by the number of server chassis a rack physically holds, add the power draw of CPUs, memory, storage, and high-speed networking interfaces in each server, and add the rack's own networking switches — and the total climbs quickly. This is why modern AI GPU racks commonly draw anywhere from roughly 40 kW at the lower end to well over 130 kW for the densest current deployments, compared to the 5-10 kW a traditional enterprise server rack draws.
Why Rack Power Determines Everything Else About the Rack
A rack's power draw isn't just an electrical number — it directly dictates the cooling method that rack requires. Air cooling remains workable up to a certain power density (roughly the 10-20 kW range, depending on the specific air-cooling design), but beyond that, removing heat fast enough with air alone becomes impractical, which is why many high-density AI racks require direct-to-chip liquid cooling or immersion cooling instead — see How Data Centers Decide the Liquid-vs-Air Cooling Split. Rack power also determines what electrical distribution the rack needs (whether standard PDU circuits suffice or higher-capacity busway taps are required, see Data Center Busway Explained) and how many racks a given electrical distribution segment can actually support before hitting its own capacity limit.
Working the Numbers Yourself
The studio's Rack Power Calculator lets you work through a real rack's power draw from its component counts — GPU count and per-GPU wattage, server chassis count, and networking overhead — rather than relying on a single rule-of-thumb number, since the honest answer to "how much does a rack use" always depends on exactly what's installed in it.
Planning for What Comes Next
Because successive GPU generations have consistently drawn more power per accelerator, a rack power figure that's accurate today is unlikely to remain accurate through a facility's full operational life. Data center electrical and cooling designers increasingly plan headroom into rack-level distribution and cooling capacity specifically so a future GPU refresh doesn't require ripping out and replacing electrical infrastructure that was sized too tightly to today's hardware.