You have the megawatts. What actually fits
Rack density went from 6 kW to 600 kW in about four years, so every sizing spreadsheet built before 2024 is now wrong by an order of magnitude. Put in a power budget or a fleet size. Get the other side, plus the heat you have to reject and the floor it lands on. Every constant is in the open, so read the method if you want to argue with one.
Three things broke at once
Capacity used to be a floor-area question. It is now power first, heat second, floor last, and most planning conversations still run in the old order.
Density broke the rule of thumb
A hall planned at 10 kW per rack and one planned at 130 kW are different buildings. Any estimate carried over from the old number is off by 10x.
Power became the unit of account
Nobody buys square feet anymore, they buy megawatts, and PUE plus held-back capacity eat the difference before a single GPU turns on.
The binding constraint moved
Below roughly 40 kW per rack, floor area binds and air works. Above it, the liquid loop binds. Which side you are on changes the entire build.
Four inputs, in the order they actually get decided
Power first, because you cannot conjure it. Density second, because you choose it. Efficiency and margin last, because you argue about them.
Which end do you know?
Power in is the siting conversation: a fixed service, and you need to know what it holds. Fleet in is procurement: the racks are decided, and you need the megawatts before you promise a date.
Pick the rack, not the chip
The highest-leverage input. Moving from 45 kW air to a 130 kW GB300 changes the cooling, the floor loading, the distribution and the schedule. Pick the closest archetype or go Custom; 50 kW and up is treated as liquid cooled.
The tax between the fence and the floor
Total facility power over IT power. At 1.2, every 100 MW at the fence delivers about 83 MW to the servers. New builds hit 1.1 to 1.2; much of the installed base sits nearer 1.5.
Capacity you build and refuse to sell
Nobody fills a hall to the last kilowatt. Network, storage and redundancy take a slice first. 15 percent is an honest default; 0 is the theoretical ceiling.
→ divide by PUE 1.20 = 25.0 MW of IT capacity
→ hold back 15 percent = 21.25 MW usable for compute
→ divide by 120 kW per GB200 NVL72 = 177 racks
→ times 72 GPUs = 12,744 GPUs
→ 21.24 MW of heat / 3.517 kW per ton = 6,039 tons, of which 5,435 go to the liquid loop
→ 177 racks x 40 sq ft = 7,080 sq ft of white space, roughly 14,000 sq ft of building
The density ladder
Published figures for the systems halls are being planned around. Air-cooled numbers are practical ceilings, not hard limits. Select any archetype to load it into the calculator.
| Archetype | kW / rack | GPUs | Cooling | Notes |
|---|
Every constant, in the open
A sizing tool that hides its assumptions is a sales tool. Here is the whole model. If you disagree with a constant, you now know exactly which one to argue with.
Usable IT = IT capacity x (1 - margin)
Racks = floor(usable IT kW / kW per rack)
Heat to reject = racks x kW per rack, in kW
Cooling tons = heat kW / 3.517
Liquid loop = 90 percent of heat, for liquid archetypes
White space = racks x 30 sq ft air, or x 40 sq ft liquid
Total shell = white space x 2.0
Energy per year = deployed kW x PUE x 8,760 hours
- Heat equals power. Essentially all electricity into a rack leaves as heat.
- The ton is 3.517 kW. Liquid archetypes send 90 percent of heat to the loop, 10 percent to room air, roughly what cold plates leave behind.
- White space is 30 sq ft per air rack, 40 per liquid, including aisles and CDUs. Shell doubles it for electrical and mechanical.
- Energy is billed on what you run: deployed load x PUE x 8,760 hours. Training clusters run close to continuous.
- Homes equivalent is 10,500 kWh per US home per year. Legible, not precise.
- A planning heuristic for the first conversation, not an engineering submittal. Nothing here replaces a mechanical engineer.
I design the console and ship it to production myself
I am Adrian Mucha, a product designer for the software most designers avoid: data centers, AI tooling, operator consoles. Recent work: a repair console saving around $10M a month, a data-access portal cut from 10 days to 48 hours, production code shipped as pull requests.
If your engineers are drawing your operator UI because you have not hired a designer yet, that is the gap I fill. Two-week console sprints or an embedded retainer.