How much electricity does it take to ask an AI a question? For two years the honest answer has been that nobody really knew, and the published guesses ranged so widely they were almost useless. On 15 June, Microsoft put a firm number on its own systems, and it is strikingly small. A typical query to a large language model, the company says, draws about 0.31 watt-hours, roughly the energy of running a 1,000-watt microwave for one to two seconds. The water used for cooling came out at less than a hundredth of a teaspoon per question.
The eye-catching part is not the figure itself but how far it sits below the conventional wisdom. Microsoft puts the middle of its range at 0.16 to 0.60 watt-hours per answer, which it calculates is between a quarter and a twentieth of estimates that have circulated widely in press coverage. If the company is right, a lot of the alarming "ChatGPT uses X times more than a web search" comparisons have been overstating the cost of a single prompt by an order of magnitude.
Where does the gap come from? Mostly from measuring the real thing rather than a lab approximation. Microsoft ran a model with more than 200 billion parameters across eight servers fitted with Nvidia H100 chips, then combined the tokens processed per second, the power each server drew, and the data centre's overall efficiency rating. Earlier estimates, it argues, often missed "batching," the practice of answering many users' questions at once, which spreads a chip's power draw across far more work. Some counted only the GPU and ignored cooling; others did the reverse. Measure a busy production system instead of a single idle chip, and the per-question number falls.
So the headline is reassuring. The trouble is that it answers a question almost nobody worried about. The strain on electricity grids was never going to come from your one query; it comes from billions of them every day, plus the enormous, separate cost of training the models in the first place. A smaller slice multiplied by a fast-growing number can still be a larger pie. Google, which has chased efficiency as hard as anyone, has watched its own emissions climb rather than fall as AI demand filled its data centres, and it has conceded it is well off track for its 2030 climate target.
This is the awkward logic that efficiency figures tend to obscure. Cheaper-per-use technologies usually get used far more, not less, and AI is a near-perfect case: make each answer trivially cheap and you invite the product to be built into every search bar, inbox and spreadsheet on the planet. Microsoft's own Copilot push is the proof. The 0.31 watt-hour answer and the soaring projections for total data centre power are not in contradiction; they are two ends of the same trend.
None of which makes the new number worthless. Honest per-query accounting is overdue, and inflated comparisons deserve to be corrected. But the figure to watch is not the cost of one answer; it is the total draw of the buildings answering all of them at once. On that, a microwave-second tells us very little.