Planning tool · No sign-up
How much power does your cluster actually need?
Choose a platform and a GPU count. The estimator returns IT load, facility power, rack count and the cooling class the hall must support — the four numbers that decide which sites can even take the call. It runs in your browser, and nothing you enter is stored.
Power is the filter. Everything else comes after.
Most capacity searches fail at the first question an operator asks: how many megawatts, at what density, and when. Walking in with those numbers already worked out is the difference between a screening reply and a real one.
The figures are deliberately conservative. Real draw depends on workload, power-capping, firmware and how the operator meters it — which is exactly why the contracted power definition is a term worth negotiating. For the terms themselves, see GPUaaS contract negotiation and the GPU contract terms that move the price.
If the density is the constraint rather than the megawatts, the answer is not always a new lease. Two-phase cooling can lift density inside a footprint you already pay for. When the megawatts are the constraint, start with power procurement and capacity sourcing.
NVIDIA GB300 NVL72
72 Blackwell Ultra GPUs in one NVLink domain. Liquid-cooled, heavy, and specified to the facility from day one.
Read the brief → Blackwell UltraNVIDIA B300 clusters
HGX B300 systems — allocation routes, density and the terms that matter at cluster scale.
Read the brief → BlackwellNVIDIA B200 clusters
HGX B200 — where it fits against B300, and how to contract for it.
Read the brief → HopperNVIDIA H200 clusters
H200 allocation and reallocation — when the previous generation is the right economic answer.
Read the brief →Asked about the estimate.
01How much power does a GB300 NVL72 rack need?
Published figures put a GB300 NVL72 rack at roughly 120 kW nominal, with OEM specifications up to about 140 kW. The estimator uses 135 kW per rack for planning. It is liquid-cooled by design and needs a coolant distribution unit and facility water, so an air-cooled hall cannot host it at any price.
02How many racks does a 1,024-GPU B300 cluster need?
It depends on the density the facility can carry. An HGX B300 system draws about 14.5 kW, so at 50 kW per rack you fit three systems, or 24 GPUs, per rack — around 43 compute racks for 1,024 GPUs, before network and storage. At 130 kW per rack that falls to 16. Set the density in the estimator to see the trade-off.
03What PUE should I assume?
Modern liquid-cooled AI facilities commonly design for a PUE between about 1.1 and 1.3; older air-cooled halls run higher. The estimator defaults to 1.30 as a conservative planning figure. The operator's contracted PUE, and how it is metered and billed, is a term worth negotiating in its own right.
04Is this a quote?
No. It is planning arithmetic from published system power figures. Real draw depends on workload, configuration, firmware power limits and the facility design. It tells you the order of magnitude of what you are looking for — a conversation gets you capacity, allocation and terms against your specific requirement.
05Is anything I enter stored or sent?
No. The calculation runs entirely in your browser. Nothing leaves the page unless you choose to email us the requirement.
You have the megawatts. Now find the site that can deliver them.
Send us the readout and the date you need it live. We will tell you where it can land, what the terms should look like and where the number in front of you sits against the market.