Most published answers to the question of how much energy artificial intelligence uses are either a single dramatic number with no arithmetic behind it, or a per-query figure quoted without the fleet context that gives it meaning. Both are avoidable. Electricity consumption in a data centre is a function of four inputs the industry already discloses: installed capacity, utilisation, cooling overhead and hardware efficiency. This piece sets out that arithmetic, applies it to the current build-out, and provides a simulator readers can run against their own assumptions.
Key Takeaway
THE ARITHMETIC, IN FULL
Start with installed IT capacity, expressed in gigawatts. That is the power the computing hardware itself can draw. Multiply by utilisation, because no fleet runs flat out every hour. Multiply by power usage effectiveness, the ratio of total facility power to IT power, which captures cooling, power conversion and everything else in the building. Multiply by the eight thousand seven hundred and sixty hours in a year. The result is annual consumption.
A worked example. One gigawatt, seventy-five percent utilisation, a PUE of 1.25. That is 1 times 0.75 times 1.25 times 8,760, which gives 8,213 gigawatt hours, or about 8.2 terawatt hours. For scale, that is roughly the annual household electricity consumption of a mid-sized American city, produced by a single campus cluster.
Two adjustments follow. Hardware efficiency improves each generation, so the useful work per watt rises and the same task costs less energy over time. And installed capacity grows, usually faster than efficiency improves. The net direction of total consumption depends entirely on which of those two curves is steeper, which is why the simulator below exposes both as sliders rather than assuming an answer.
WHAT A SINGLE QUERY COSTS
Per-query figures published over the last two years cluster between a fraction of a watt hour and a few watt hours for text generation, depending on model size, context length and how effectively requests are batched. Image and video generation sit an order of magnitude higher.
At around one watt hour, a single text query is roughly six minutes of a ten watt LED bulb, or about one fifteenth of a smartphone charge. Individually trivial. The number only becomes material when multiplied by query volume, and query volume is precisely the figure that grows when a capability becomes a default feature inside search, office software and operating systems rather than a destination people visit deliberately.
Nobody worries about the bulb. The question is how many bulbs, left on, everywhere, by default.
RUN THE NUMBERS
The simulator below converts capacity into annual terawatt hours, share of national generation, emissions, cost and per-query context. Load a regional preset, then override any input. The standalone version is at /calculator/ai-power-demand-simulator.
Fleet and grid
Grid context and pricing
Projected annual consumption
Bars are terawatt hours per year, including cooling and overhead at the chosen PUE, net of the efficiency gain accruing across the horizon.
Verdict
AI load becomes a visible line item in national demand. Expect capacity auctions, ratepayer disputes and behind-the-meter generation deals.
Year one
99 TWh
Year 6
173 TWh
Share of grid now
2.29%
Share at horizon
4.03%
Emissions now
0 MtCO2
Power bill
$7.7B/yr
Year-one consumption equals the electricity used by about 9.4m households, and could serve roughly 110bn AI queries.
One query, in context
THE CONSTRAINT IS NOT GENERATION
National totals understate the problem because they average across a grid that cannot move power freely. AI capacity does not distribute itself evenly. It concentrates where land, fibre and tax treatment align, which produces local load growth of a kind distribution networks were not planned for.
The binding constraints, in the order operators actually encounter them, are interconnection queue position, transformer and turbine lead times, transmission capacity into the specific substation, and only then generation. A project can have a signed power purchase agreement and still wait years for the equipment that connects it. This is why behind-the-meter generation, gas turbines on site, and long-dated nuclear agreements have moved from novelty to standard practice in siting decisions.
It is also why ratepayer politics has arrived. When a single campus adds load equivalent to a city, someone has to fund the network reinforcement, and the argument over whether that is the operator or the general customer base is now live in several jurisdictions.
WHERE THE THRESHOLDS SIT
As a share of national generation, below about four percent AI load remains a local planning issue. Between four and ten percent it becomes a visible line item in national demand forecasts, with capacity auctions and tariff disputes following. Above ten percent, the conversation stops being about chips and becomes about building power stations, which operates on a decade-long clock rather than a procurement cycle.
Set the growth and efficiency sliders in the simulator against each other and the crossover is easy to find. At twenty percent annual capacity growth, a thirty-five percent efficiency gain spread over six years does not stabilise consumption. It slows the climb. Consumption falls only when capacity growth approaches zero, and nothing in the current commitment book suggests that.
Key Takeaway
RELATED WORK
The capital behind this build-out is mapped in AI circular financing, explained with the actual numbers, and the ring itself is editable in the Circular Financing Map. For the systemic framing, read The Circle Tightens.
This article was researched and written by human editors with analytical assistance from AI tools. All conclusions, interpretations, and editorial decisions are independently reviewed by the CALCULATORiQ Editorial Team before publication.
For questions about our editorial process, see our Editorial Standards page.
Share this brief
