Capacity is an envelope, and the envelope has a plane
A site's power envelope is the most it may draw. It is set by the grid connection, the switchgear and the contract, and it rarely moves for years.
The number in a grid or colocation agreement is almost always total facility power. A capacity plan for the data hall usually counts IT load at the rack input. The two differ by the site's overhead, which is what PUE measures.
Choose where you count before computing anything. At a PUE of 1.25, a 10 MW facility envelope leaves 8 MW for IT equipment. At 1.40 it leaves about 7.1 MW. On tight arithmetic, that difference decides whether a requirement fits.
How racks become megawatts
Three numbers turn racks into megawatts. Accelerators per rack, from the manufacturer's specification. Rack draw at the rack input, which covers the whole rack. And the share of that draw that is accelerator.
The rack draw is the number most often misread. It includes host processors, memory, networking, fans and conversion losses as well as the accelerators. Plan against the whole rack, and round down to whole racks.
The accelerator share matters because a saving at the GPU acts only on that share. It is often the largest single lever in the arithmetic, so use your own.
Headroom inside the envelope you hold
Headroom is power you already hold that your hardware is not using. When each accelerator draws less for the same work, the difference becomes room for more equipment inside the same envelope.
Converting a saving at the die into racks takes two numbers that are yours: the accelerator share of your rack draw and, for a facility envelope, your PUE.
One check catches an input error: the conversion never returns a larger percentage at the rack than at the die.
Do not count the same watts twice
The energy you stop buying and the headroom you free are two uses of the same watts. Fill the headroom with new racks and those racks draw the power that was saved.
Bank one and plan against it. A capacity plan that counts both has counted the same watts twice.
When more megawatts are on order
A bigger grid connection keeps its own timetable. While it does, the envelope you already hold can carry more of the fleet.
Work from your own queue date. A deployment waiting for new capacity can run part of its fleet inside today's envelope, if the arithmetic on your own numbers says it fits.
Grid flexibility, in plain words
A flexible load is one whose draw can be reduced on instruction and restored afterwards. Grid flexibility programmes pay for that ability, under their own rules for notice, duration and measurement.
A data centre offers flexibility only if it can show the reduction where the programme measures it, which is usually the meter rather than the chip.
Where the gain lands
Bursty and part-loaded fleets gain the most, and the guide to reducing GPU power consumption shows how to check yours with a week of readings.
Where power is passed through to customers, the bill is not the argument and the freed capacity is. A silicon refresh changes the rack numbers, so rerun the arithmetic on the new parts.
Questions people ask
- What is data centre power capacity?
- The most power a site may draw, set by its grid connection, switchgear and contract. It is usually stated as total facility power, which includes cooling and conversion as well as IT equipment.
- How do you free power capacity without a new grid connection?
- Reduce what the existing equipment draws for the same work, cut the facility overhead, or both. The power released becomes headroom for more equipment inside the same envelope.
- Can a GPU saving count as both energy saved and capacity freed?
- Each watt counts once, as energy you stop buying or as room for more equipment in the same envelope. Fill the headroom and the new equipment draws the power that was saved, while each unit of work still takes less energy.
Where ATHLAZ fits
ADAPT frees headroom inside the envelope
ADAPT, AI-Driven Adaptive Power Technology, is a per-GPU software runtime. It reduces what each accelerator draws while the work runs, so more compute fits inside the envelope you hold.
Measured: up to 21% less GPU die power on NVIDIA H100 NVL, over a continuous 48-hour window, with managed and baseline arms under an equal power cap. Any rack or facility figure built on it is modelled on your inputs.
Grid flexibility is on the ATHLAZ roadmap as the ATHLAZ Power Network, which builds on sites already running ADAPT.