IT load
Announced MW divided by PUE. The share of facility power that actually reaches GPUs after cooling, power conversion, and other overhead.
Calculator
Power-to-Compute Translator
Translate an announced megawatt number into a GPU-capacity range.
Data-center announcements headline a megawatt figure. Buyers and analysts want to know what that means for AI compute supply. The translator converts announced power to an IT load using PUE, sizes a rack count from your rack density, and produces a GPU capacity range with an honest band rather than a single point estimate. Announced MW is not active capacity.
Interactive calculator
Power-to-compute translator
Headline megawatts from a data-center announcement, expansion, or PPA. This is buildout potential, not active capacity.
MW
Power usage effectiveness. Total facility power divided by IT power. Modern AI data centers target 1.1 to 1.3.
IT power per rack in kilowatts. Air-cooled AI racks land near 30 to 60 kW; liquid-cooled builds reach 100 to 200 kW or more.
kW
The accelerator model you are pricing. Selecting one fills in an illustrative starting rate you can edit.
Total GPUs packed into a rack at the chosen density. Typical 8-GPU HGX servers at 60 kW give 32 to 48 GPUs per rack.
Share of installed GPU time that does useful work. Lower utilization shrinks the effective compute the site produces.
%
100 MW at PUE 1.20 translates to roughly 83 MW of IT load and a capacity range of 24,889 to 44,444 H100-class GPUs at full buildout. Announced power is not active compute.
Estimated IT load83 MW
Estimated rack count1,389
Estimated peak GPU capacity44,444 H100-class
GPUs producing output at utilization35,556 H100-class
Annual useful GPU-hours311,466,667 GPU-h
Announced MW is buildout potential. Energization timelines, supply chain, chip allocation, water and cooling permitting, and capacity ramp are not modeled here. Use the range as a directional ceiling, not a forecast of active GPU-hours.
Starting values are illustrative defaults you can edit — not live ComputeTape benchmark prices. Replace them with a real quote.
How to read the result
What the numbers mean
The translator estimates a ceiling. It tells you the largest plausible GPU capacity that the announced power could support at the assumptions you enter. The range is wide on purpose; the upper end is theoretical peak and the lower end accounts for incomplete utilization.
Rack count
IT load in kW divided by rack density. Higher density (liquid cooling, denser packing) yields fewer racks for the same load.
Peak GPU capacity
Rack count multiplied by GPUs per rack. The theoretical upper bound at full buildout, every rack populated.
GPUs at utilization
Peak capacity times the utilization input. A 100 MW announcement at 80% utilization produces materially less compute than at 100%.
Why announced is not active
Announced power is a buildout signal
A megawatt number in a press release reflects planning, financing, and permitting headroom — not chips installed today. Energization, grid interconnection, transformer supply, GPU allocation, and capacity ramp all sit between an announcement and the first useful GPU-hour. Treat the translator output as a range of what could land if everything goes well, not as a forecast.
Why power matters
Why electricity and site capacity shape AI compute markets.
What is PUE?
Power usage effectiveness and what realistic AI-data-center numbers look like.
What is liquid cooling?
How cooling choice changes rack density and the GPUs a megawatt can support.
Power Watch
Live power signals tracked with the same rights and freshness contract as provider data.