PUE & AI Efficiency Calculator
PUE tells you how much of the power bill actually reaches the servers. The AI-era companions — TFLOPS per watt and tokens per watt — tell you what the compute produces per unit of energy.
total ÷ IT
cooling + power losses
of total facility power
inference productivity
Results update live as you type. For planning and field-check estimates — always verify against applicable standards and equipment ratings.
PUE and what comes after it
Power Usage Effectiveness is total facility power divided by IT power. A PUE of 1.35 means that for every 100 kW reaching servers, another 35 kW is spent on cooling, power conversion losses, and building systems. The theoretical floor is 1.0; the global fleet averages around 1.5–1.6, while hyperscale builds with economization and liquid cooling report 1.1–1.2. The metric's blind spot is that it says nothing about whether the IT power itself does useful work — an idle cluster at PUE 1.1 is still waste.
That's why AI operators increasingly track compute productivity per unit of energy. TFLOPS per watt measures raw arithmetic capability against the full facility draw, useful for comparing hardware generations — each new accelerator family has roughly doubled it. Tokens per watt (or per watt-hour) is the inference-era metric: how much model output the whole facility produces per unit of energy, which folds in model efficiency, batching, quantization, and utilization, not just silicon.
Together the three numbers form a chain: PUE tells you how efficiently power reaches compute, TFLOPS/W tells you how capable that compute is, and tokens/W tells you how much value comes out. Anyone underwriting an AI facility should be asking for all three.