Every time I ask an AI tool to help draft something for work, the answer just… appears ! No noise, no visible effort, nothing to suggest anything strenuous just happened. This is a strange contrast to my day job. Everything runs on infrastructure you can hear and see, with conveyors rattling and forklifts beeping. However, with AI , is it the same? what is really going on in behind the scene?
AI’s infrastructure is the exact opposite: it’s enormous, and it’s completely invisible from where you’re sitting typing a question.
Somewhere, possibly thousands of miles from wherever you’re reading this, a hall full of specialised microchips has just worked flat out to produce those few paragraphs. Those chips draw real electricity, throw off real heat, and need genuinely vast industrial buildings just to keep functioning.
That hidden effort is where a once-famous claim came from: that asking an AI chatbot a single question burns through roughly ten times more electricity than a normal search engine query. There’s a grain of truth buried in that figure — but it’s also years out of date, and the real story turns out to be more interesting than the soundbite.
What Does a Watt-Hour Actually Mean?
Quick bit of housekeeping before any comparisons: a watt measures power, which is how fast electricity is being drawn at any given moment. A watt-hour measures energy, the total amount used over a stretch of time. Run a small ten-watt reading lamp for an hour, and you’ve used exactly ten watt-hours. When researchers work out the footprint of a single AI answer, they’re dealing in tiny slivers of a watt-hour — which is exactly why “ten times more power” was always a slightly loose way of putting it. What’s really being compared is the total energy needed to finish the job.
Where Did the “Ten Times More” Figure Actually Come From?
Traditional web search is remarkably efficient. Back in 2009, Google itself estimated that a typical search took around 0.3 watt-hours. When generative AI chatbots first became mainstream, early independent estimates put a typical AI-generated answer at somewhere around 2.9 watt-hours — call it 3 for round numbers. Divide the two, and you land almost exactly on “ten times more energy,” which is why that figure spread so widely: it was a genuinely clean, quotable number.
The trouble is that it’s now stale. AI systems have become dramatically more efficient in a short space of time. Google published its own detailed measurements of its Gemini assistant in August 2025, and found the median text prompt used just 0.24 watt-hours — roughly what a microwave draws in a single second, and squarely back in line with those old search-engine estimates. Google also reported that this figure had fallen roughly 33-fold over the previous twelve months alone, thanks to more efficient models and smarter software underneath.
Worth flagging honestly: not everyone is convinced these self-reported numbers tell the whole story. Some independent researchers have pointed out that figures like this typically leave out things such as the energy used to actually train the model in the first place, the network carrying your request back and forth, and the device you’re reading the answer on — all real costs, just not counted in the headline number. So treat any single figure, including this one, as a reasonable estimate rather than gospel. Either way, there’s no fixed rule saying every AI question costs exactly ten search queries’ worth of electricity. Simple questions barely register; complicated ones can use a genuinely large amount.
Searching Finds an Answer. AI Has to Build One.
Ask a search engine “why is the sky blue?” and it behaves like an extremely well-organised librarian — it doesn’t write you an answer, it scans a vast, pre-built index of the web, finds pages that already explain it, ranks them, and hands you the links.
A chatbot does something genuinely different. It doesn’t retrieve a stored file — it builds a brand-new response, word by word, working out what should come next based on billions of patterns absorbed during training. That process is called inference, and generating every word (or fragment of a word, known as a token) means running your question through a huge neural network, firing off millions of calculations in the process. A long, detailed answer means round after round of that work before the final full stop appears.
Where the Electricity Actually Goes: Chips and Cooling
Ordinary computer chips aren’t built to cope with that volume of parallel calculation. AI systems instead rely on specialised processors — accelerators, or graphics processing units (GPUs) — originally designed to render video game graphics, and rather good, it turns out, at doing thousands of small mathematical tasks simultaneously. Modern AI data centres pack tens of thousands of these into enormous server halls.
Run that many chips flat out and two things happen at once: they draw a serious amount of electricity, and they generate a serious amount of heat. Let a server room get too warm and chips start slowing down or failing outright, so data centres run industrial-scale cooling around the clock — huge fans, chilled water systems, air handlers — just to keep temperatures in check. The true energy cost of an AI answer, then, isn’t just the electricity flowing into the chip itself; it includes all the cooling infrastructure keeping the building liveable for the hardware.
Not Every Question Costs the Same
A single average figure for “an AI query” is genuinely misleading, because different requests demand wildly different amounts of work. Compare “what’s the capital of France?” with “read this 200-page report, compare five years of financial performance, identify the major risks, and write me a detailed summary.” Both arrive as one message in a chat box. The second one requires thousands of times more computation than the first.
What actually determines the energy cost:
- Model size. Larger, more capable models generally need more electricity per word generated.
- Length. Longer questions and longer answers mean more rounds of processing.
- Reasoning and tools. If the AI has to pause, double-check itself, read an attached file, or search the live web, the cost climbs.
- Media type. Generating an image, a voice clip, or a video takes vastly more energy than writing a paragraph of plain text.
Efficient Doesn’t Mean Small-Scale
Here’s the genuinely encouraging bit: individual AI tasks keep getting cheaper and faster, at a fairly remarkable pace. Chipmakers keep delivering more computing power per watt, and engineers have found clever ways to wake up only the specific part of a model actually needed for a given task, rather than powering up the whole system every time.
So why do energy analysts remain worried, if individual questions are getting lighter? Because of scale. Swap an old incandescent bulb for an efficient LED, and that one bulb uses a fraction of the electricity — genuinely good news. But install a thousand new LED bulbs where there used to be a single old one, and your total electricity bill still climbs regardless. That’s more or less the situation AI is in. Millions of people now use it daily, it’s being folded into search engines, phones, office software, and customer service — and, crucially, people are shifting away from quick one-line questions towards asking AI to handle sprawling, multi-step jobs like deep research or video generation.
According to the International Energy Agency, data centres worldwide used around 415 terawatt-hours of electricity in 2024 — about 1.5% of global electricity use — and that figure is projected to more than double, to roughly 945 terawatt-hours by 2030, a number the IEA notes is close to Japan’s entire current annual electricity consumption. AI is the single biggest reason behind that growth.
It’s Not Just Electricity — There’s Water and Carbon Too
Electricity is only half the equation; where it comes from matters just as much. A data centre running on solar, wind, or hydroelectric power has a very different footprint from one drawing on a coal-fired grid, and because AI facilities run around the clock, securing steady clean power has become one of the industry’s biggest headaches.
Water is the other hidden factor. Many facilities cool their servers using evaporative cooling towers. A single quick text answer might account for only a handful of drops, but multiply that across billions of daily interactions — some of them in genuinely drought-prone regions — and it becomes a real, physical issue for the communities living near those buildings, not just an abstract statistic.
So What’s the Actual Hidden Cost?
For most of us, AI feels weightless. There’s no engine noise beside your desk, no exhaust pipe, nothing to suggest effort. You type a question, and words appear. Behind that clean little text box, though, sits one of the largest infrastructure build-outs in recent memory — sprawling facilities packed with processors, high-speed networking, and heavy-duty cooling, running continuously.
The genuinely good engineering news is that every individual calculation is getting cleaner and more efficient by the month. The harder problem is that we’re collectively asking AI to do vastly more calculations than ever before, and that number shows no sign of levelling off. The real question isn’t how much electricity a single chatbot reply costs any more. It’s what happens to the power grid once billions of people, businesses, and automated systems are asking AI to think on their behalf, billions of times a day, every single day.