About the Energy Cost of Intelligence
Intelligence is becoming a utility, and like any utility its price floor is set by physics: joules per token and silicon depreciation per token. This calculator turns the specs of a GPU inference deployment into a concrete answer - what a million tokens costs in energy, in hardware, and in total - and shows how that floor moves as electricity gets cheaper. It is the bridge between the AI buildout and the energy abundance thesis: cheap power is cheap thought.
How to use it
- Set the GPU's power draw and its sustained inference speed in tokens per second.
- Set facility PUE and your electricity price.
- Enter installed hardware cost, expected lifetime, and average utilization.
- Read Wh per token, energy cost, hardware cost, and the total per million tokens.
- Check the sensitivity chart to see what cheaper power does to the total.
How it works
Serving one million tokens takes 1,000,000 / (tokens per second x 3600) GPU-hours. Energy is GPU-hours x kW x PUE, priced at your rate. Hardware cost per GPU-hour is installed cost divided by lifetime hours times utilization - depreciation runs around the clock, but only utilized hours serve tokens. The two components sum to the physical floor of the cost per token; everything else in a commercial API price is stacked on top of it.
Worked example
A 1.2 kW GPU at 60 tok/s with PUE 1.25 uses about 6.9 Wh per 1,000 tokens. At $0.06/kWh that is $0.42 per million tokens of electricity; a $30,000 GPU over 5 years at 60% utilization adds about $5.28 of depreciation - hardware, not energy, dominates. Drop electricity to $0.02 and the total barely moves; halve the GPU price or double its speed and it plunges.
Frequently asked questions
What does a million tokens actually cost in electricity?
At 60 tokens/s on a 1.2 kW GPU with PUE 1.25 and $0.06/kWh, roughly $0.42 of electricity per million tokens - hardware depreciation usually costs several times more.
Why include PUE?
Every watt a GPU draws needs cooling and power delivery overhead. PUE scales GPU energy to real facility energy - 1.1 is excellent, 1.5 is mediocre.
Why does utilization matter so much?
Hardware depreciates whether or not it is serving tokens. At 30% utilization, each served token carries three times the capex of 90% utilization.
How does cheap energy change AI economics?
Below about $0.03/kWh, electricity nearly vanishes from the cost per token - which is why AI datacenters chase cheap solar, hydro, and eventually orbital power.
Is my data uploaded?
No - the model runs entirely in your browser.