# AI energy use: why does it need so much power? By Amanda, cofounder of [Wistkey](https://wistkey.com) · 17 Aug 2026 · AI Canonical: https://styamanda.ai/blog/ai-energy-use/ Players keep asking whether an AI reply burns a small forest. This is not how the damage counter works. A short text answer can be fairly modest; the serious load comes from enormous models, heavier jobs, always-ready infrastructure, and millions of people pressing the summon button at once. ## Why AI uses so much electricity **AI uses so much electricity because it performs a huge number of mathematical operations on specialised chips, moves data through memory, and relies on servers, networking, power equipment, and cooling around those chips.** Training a model can be an intense one-off computation. Running it for users—called inference—repeats computation for every prompt and every generated token. One request may be small, but billions of requests, larger models, and demanding tasks add up. That distinction matters. The International Energy Agency’s current [Energy and AI analysis](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary) says simple text queries have become dramatically more efficient. It also says video generation, extended reasoning, and agentic tasks can use hundreds or thousands of times more energy per request. There is no honest universal number for “an AI query” because asking for one sentence and generating a video are not the same spell. > The energy problem is not one prompt in isolation. It is computation × task size × users × frequency, plus the building kept ready to serve them. ## Where the electricity actually goes An AI data centre is a physical factory for computation. The product looks weightless on your screen, but the machinery is very real. Its electricity bill has several main parts. ### 1. The chips do a vast amount of arithmetic Language and image models are built from large collections of numerical parameters. To train or run one, processors repeatedly multiply and add large grids of numbers. GPUs and other AI accelerators are designed to perform many of these operations in parallel. That makes them much faster than a general-purpose processor for the job, but a packed rack of accelerators still demands substantial power. Bigger is not automatically better, but more parameters, more input, more generated output, and more reasoning steps usually mean more computation. This is also why the token count discussed in my field note on [what AI tokens are](https://styamanda.ai/blog/ai-tokens-explained/) is not just a billing curiosity. Longer prompts and replies keep the machinery working for longer. ### 2. Moving data costs energy too The model’s numbers must be available to the chips at high speed. That requires memory, storage, networking, and host processors. Feeding an accelerator can be as important as the calculation itself; a very fast chip waiting for data is expensive furniture with fans. A useful measurement therefore counts the whole serving system, not only the accelerator while it is busy. Google’s published [inference measurement method](https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/) includes host CPUs, RAM, idle capacity held for traffic spikes, power distribution, and cooling. Its figures describe Google’s own Gemini service at a point in time, not every model everywhere, but the accounting boundary is the important lesson. ### 3. Heat has to leave the building Nearly all the electricity used by computing equipment eventually becomes heat. Dense AI racks create a lot of it in a small space, so data centres use air or liquid cooling to keep hardware within operating limits. Pumps, fans, chillers, and power-conversion equipment consume additional electricity. Some cooling systems also use water, which means local climate and water stress matter as well as the headline power figure. ### 4. The service must be ready before you ask People expect an AI assistant to answer immediately. Providers therefore keep spare capacity available for busy periods and failures. A server that is powered and waiting is not doing zero work from the grid’s point of view. Reliability has a resource cost, just as keeping extra healers at the save point does. ## Training is not the whole story Training gets attention because it concentrates an enormous job into weeks or months: the system processes examples, compares predictions with desired results, adjusts its parameters, and repeats. Fine-tuning repeats a smaller version of that process. My [prompting versus fine-tuning guide](https://styamanda.ai/blog/prompting-vs-fine-tuning/) explains why most teams should not reach for model training first. But a trained model may then serve users for years. Each response is an inference job. At sufficient scale, the repeated cost of inference can rival or exceed the original training run. New features also change the balance: generating video, using long contexts, trying several reasoning paths, or letting an agent call tools repeatedly requires much more work than returning a short text answer. ## Is one AI prompt environmentally disastrous? Usually, no. Treat viral per-prompt comparisons cautiously. Energy varies with the model, hardware, request length, output length, data-centre efficiency, utilisation, and task type. Carbon emissions vary again with where and when the electricity was generated. Water estimates depend on the cooling design and local conditions. A precise-looking number without those details is often costume armour. The system-level trend is still important. The IEA estimates that data centres used 485 terawatt-hours of electricity globally in 2025 and projects roughly 950 terawatt-hours in 2030, around 3% of global electricity demand. Those totals include more than AI, but the agency identifies AI as the fastest-growing part. It also stresses uncertainty: efficiency is improving quickly while usage and the appetite for heavier tasks are growing too. Local impact can be sharper than the global percentage suggests. A large facility arrives in one place, asks for a large and reliable grid connection, and may need new generation, transmission, and water infrastructure. The relevant community question is not merely “Does AI use electricity?” Everything digital does. It is “Who pays for the new capacity, what powers it, how flexible is the demand, and what trade-offs were agreed in public?” ## What users and teams can do You do not need to ration every useful prompt. You do need to stop treating all AI work as equally cheap. A practical efficiency checklist is: - **Use the smallest capable model.** Route routine extraction, classification, and drafting to a lighter model; reserve the largest model for jobs that need it. - **Keep prompts and context relevant.** Do not resend an entire archive when five paragraphs contain the answer. - **Limit output and retries.** Ask for the format and length you need, then fix the instruction instead of generating ten vague alternatives. - **Match the medium to the job.** Text, images, video, deep reasoning, and autonomous agents have very different compute profiles. - **Measure the workflow.** Teams should track model, tokens, latency, retries, and task success. Cost is an imperfect but useful signal for wasted computation. - **Ask providers for system-level evidence.** Look for energy, emissions, water, location, and methodology—not a single flattering per-prompt number. At [Wistkey](https://wistkey.com), the useful design question is the same one I use for every inventory slot: does this tool earn the resources it occupies? The greenest AI feature is not necessarily the tiniest model. It is the feature that solves a real problem with the least capable system that can do the job reliably. ## Efficiency will help, but it may not shrink the total Chips, models, and serving software are getting more efficient. That is real progress. It does not guarantee that total electricity use falls. When a task becomes cheaper, more people use it, providers add more demanding features, and yesterday’s premium capability becomes today’s default button. Efficiency lowers the cost per quest; popularity spawns more quests. So the honest answer is neither “every prompt is a climate disaster” nor “better chips will make the problem disappear.” AI is becoming more efficient per unit of work while the total amount and intensity of work are climbing. Both can be true. I have logged this under *physical limits the interface forgot to mention*. The cloud was always someone else’s machine. It turns out the machine has a power cable.