Data Center Power Estimator

Buying AI hardware is a one-time shock; powering it is a bill that arrives every month for years. At the scale of modern clusters, the electricity to run the silicon has become comparable to — sometimes greater than — the amortized cost of the silicon itself, which is why infrastructure decisions now start with a power-and-cost model rather than a hardware spec sheet.

The arithmetic is straightforward but unforgiving. The IT load is the accelerators times their power plus the supporting CPUs, networking and storage. PUE multiplies that into the total the facility pulls from the grid, adding the cooling and distribution overhead — a 1.3 PUE means thirty cents of overhead on every IT dollar, burned continuously. Energy is facility power times the hours times utilization, cost is energy times price, and carbon is energy times the grid's intensity. Four multiplications, eight-figure consequences.

Two levers dominate, and they're different for cost and carbon. For cost, PUE and the electricity rate rule — which is why hyperscalers chase low-PUE designs and cheap-power regions, and why this console shows the PUE overhead explicitly. For carbon, the grid's intensity dominates: moving the same cluster from a coal-heavy grid to a hydro- or nuclear-rich one can cut emissions several-fold, a far bigger effect than any efficiency tweak. Reporting cost and carbon separately, and letting you set the grid intensity, makes that distinction actionable.

Because these bills are budgeted and reported in local currency, this tool converts every monetary figure — the electricity rate you enter and the annual cost it produces — into your chosen currency with correct locale formatting. Pair it with the Energy Per Inference console for workload-level efficiency, and the Fab ROI console for the capital side of the ledger.