WattThe?!
Back to Data Center
Data Center tools

AI Training Energy Calculator: Power & Cost to Train an AI Model

Estimate the electricity, cost, and carbon to train a large AI model. Enter GPU count, training days, and power draw to see total energy in MWh and dollars. Free calculator from WattThe?!

Number of GPUs
Power draw per GPU(W)
GPU utilization during training 90%
Training duration(days)
Data center PUE 1.2
State (reference only — does not change rate)
Electricity rate($/kWh)

Data centers typically pay $0.06–0.10/kWh industrial rates, well below residential rates.

Grid carbon intensity(lbs CO₂/kWh)
Total facility power
27.0MW

IT load × PUE

Total energy
38,880MWh

38,880,000 kWh over 60 days

Electricity cost
$3,110,400

at selected rate

CO₂ emissions
15,941short tons

31,881,600 lbs at 0.82 lbs/kWh

Homes equivalent
3,600US homes / year

at 10,800 kWh/year avg

Results update live as you type. For planning and field-check estimates — always verify against applicable standards and equipment ratings.

How we calculate this →

From GPU count to power bill: how training energy adds up

Training a frontier AI model is one of the most energy-intensive computing workloads ever built. A cluster of 25,000 Blackwell-class GPUs drawing 1 kW each at 90% utilization produces 22.5 MW of IT load — before the building's cooling, power distribution, and lighting are added. Multiplying by PUE (Power Usage Effectiveness) captures that overhead: a 1.2 PUE means 20% of total facility power goes to non-compute infrastructure, giving 27 MW for the whole campus.

Over 60 days of continuous training that facility draws about 38,880 MWh — roughly 38.9 GWh — enough to power about 3,600 average US homes for an entire year. At industrial data-center rates around $0.08/kWh, the electricity bill alone approaches $3.1 million. Carbon follows the same math: total kWh multiplied by the grid's average CO₂ intensity (0.82 lbs/kWh on the US average mix) gives the emissions attributable to the run.

The key insight is that run time is the biggest lever after cluster size. Doubling training duration from 60 to 120 days doubles energy and cost, while improving GPU utilization from 80% to 90% reduces total energy by 11% — a meaningful saving at this scale. PUE is the multiplier that makes cooling choices consequential: dropping from a 1.5 PUE legacy facility to a 1.2 modern design saves 20% of total cost with no change to the compute cluster itself.

Frequently asked questions

Training a frontier model on a cluster of 10,000–25,000 high-performance GPUs for 30–90 days can consume anywhere from several thousand to over 50,000 MWh of electricity. GPT-4 scale training runs have been estimated at roughly 50 GWh; next-generation models training on larger clusters for longer can significantly exceed that. The dominant factors are cluster size (GPU count × power draw), training duration, and data center PUE.

Training requires computing billions of forward and backward passes through a neural network — updating trillions of parameters across thousands of GPUs simultaneously for weeks or months, with nearly all GPUs pegged at close to maximum utilization. Inference, by contrast, runs a finished model to answer one question at a time, typically on far fewer chips and for fractions of a second. A single training run can consume as much energy as millions of individual inference queries.

PUE — Power Usage Effectiveness — measures how efficiently a data center uses electricity. A PUE of 1.0 would mean all power goes to computing; a PUE of 1.5 means 50% extra energy goes to cooling, lighting, and power distribution overhead. Modern hyperscale facilities designed for GPU workloads achieve PUEs of 1.1–1.3; older air-cooled facilities can run 1.5–2.0. For a 27 MW training cluster, dropping PUE from 1.5 to 1.2 saves roughly 6 MW of continuous load — about $1 million per year in electricity at $0.08/kWh.

Related calculators

Inference Tokens per Watt CalculatorData Center Load CalculatorData Center Power Cost & TCO Calculator