You type a question into ChatGPT. Within seconds you have an answer. What you don’t see is the data centre humming somewhere in Virginia or Dublin that made it happen — drawing power equivalent to a small town, around the clock, every day of the year.

I’ll be honest: when I started pulling together the numbers for this piece, I expected them to be bad. They’re worse than I expected. The microwave analogy from MIT’s research is the one that stopped me — a five-second AI video clip costs as much energy as running a microwave for over an hour. Let that land for a second. One clip. Five seconds.

Here’s what the data actually shows.

One Query at a Time

Start small. A single ChatGPT query using GPT-4o uses approximately 0.34 watt-hours of electricity, according to OpenAI CEO Sam Altman. A standard Google search uses roughly 0.3 watt-hours by comparison — so for a basic text prompt, they’re surprisingly close. To put that in perspective, 0.34 Wh is less than an LED lightbulb uses in a few minutes.

But that’s the best case. Complexity changes everything fast. Attaching a document to ChatGPT pushes usage to around 2.5 Wh. A very long 100,000-token context — roughly 200 pages of text — requires close to 40 Wh. And GPT-5, which launched in August 2025, averages around 18.9 Wh per medium-length prompt according to researchers at the University of Rhode Island — more than 50 times the energy of a Google search.

Per-Query Comparison

Zoom Out: The Data Centre Picture

Individual queries are small. Scale is not. Global data centre electricity consumption hit 485 TWh in 2025, growing 17% in a single year — more than four times the rate of total global electricity growth. AI-focused data centres grew even faster, surging 50% in 2025 alone, according to the IEA’s April 2026 report.

The IEA’s central projection sees data centre electricity consumption roughly doubling to 950 TWh by 2030 — accounting for around 3% of all global electricity demand. AI-specific facilities are expected to triple their consumption in that period. For context, 950 TWh is slightly more than Japan’s entire national electricity consumption today.

Five large tech companies — think Microsoft, Google, Amazon, Meta, and their peers — collectively spent over $400 billion in capital expenditure in 2025 on data centre infrastructure, with a further 75% increase forecast for 2026.

Why a Single GPU Cluster Changes Everything

The hardware behind AI is extraordinarily power-hungry. A single NVIDIA H100 GPU — the chip powering most serious AI workloads — draws 700 watts continuously. A cluster of 100,000 H100s, which multiple companies now operate, draws 70 megawatts non-stop.

70 megawatts. Non-stop. For one cluster. There are now multiple of these. That’s not a data centre — that’s a power station dedicated to answering questions and generating images.

Goldman Sachs has identified energy availability as the single biggest infrastructure constraint in AI right now — displacing chip supply as the binding limit on how fast the industry can grow. That shift happened faster than almost anyone predicted.

Big Tech’s Answer: Go Nuclear

Renewables alone cannot solve this. Solar and wind deliver 25–35% capacity factors — meaning they only generate power a fraction of the time. Data centres need continuous, reliable power. So the world’s largest tech companies have done something that would have seemed extraordinary five years ago: they’ve gone all-in on nuclear.

Microsoft reopened Three Mile Island Unit 1 in Pennsylvania — yes, that Three Mile Island — under a 20-year, $16 billion power purchase agreement supplying 835 MW to its AI data centres. Google signed the largest corporate nuclear deal in history with Kairos Power for small modular reactors expected around 2030. Amazon is working with X-energy to deploy over 5 GW of SMR capacity by 2039. Meta is developing a 1.2 GW nuclear campus with Oklo.

Whether reopening Three Mile Island genuinely counts as clean energy is a debate worth having — but the timelines make it hard to see what else these companies could realistically do. The grid can’t keep up. Renewables aren’t reliable enough at scale. Nuclear is the only option that fits both the power density and the always-on requirements of modern AI infrastructure.

The pipeline of conditional agreements between data centre operators and SMR projects has grown from 25 GW at the end of 2024 to 45 GW today. A decade ago that sentence would have read like science fiction.

AI Video Generation: The Hidden Energy Bomb

Text queries are just one part of the picture. AI image and video generation use dramatically more power — and they’re growing fastest. Generating a single AI image uses around 3 Wh of electricity. A five-second AI video clip, according to MIT Technology Review’s 2025 investigation, uses approximately 50 Wh — equivalent to running a microwave for over an hour. Ask an AI for 15 questions, 10 images, and three short video clips in one session, and your total energy draw reaches roughly 2.9 kWh — the same as running a microwave for three and a half hours.

This is the dimension of AI energy consumption that almost nobody discusses when comparing ChatGPT to Google. The comparison assumes text. The reality in 2026 is that AI content generation — images, audio, video, and the virtual human and anime AI being built by companies like Tokyo’s AiHUB — is orders of magnitude more energy-intensive than a search query, and it’s the fastest-growing slice of AI compute demand.

AI Electricity Use Projections Explained — data centres, renewable energy, and the competition for power (Nov 2025)

What This Means for You

If you’re in Europe, AI data centres are already affecting your grid and your electricity bills — and nobody is really talking about it in plain terms. If you’re in the US, your utility company is scrambling to keep up with data centre connection requests that have years-long backlogs. Virginia electricity prices jumped 13% year-on-year in August 2024, largely because of this.

The good news — and there is some — is that efficiency is genuinely improving. The IEA notes that simple text queries now consume less electricity than running a television over the same period. That’s real progress. But AI agents — always-on assistants that run continuously, take actions, and chain tasks together — are the next wave, and they’re far more energy-intensive than a single prompt. We’re not even close to the peak of this curve.

The honest summary: using ChatGPT for text is fine. The energy cost is comparable to a Google search. But generating images and videos at scale, running agents continuously, or using the latest reasoning models is a different story entirely. The electricity bill for AI’s ambitions is going to land somewhere — on the grid, on consumers, or on the climate. Probably all three.

What Can You Actually Do About It?

Probably less than you think — and more than you’d expect. Your individual prompt usage is genuinely negligible at the scale of global data centre demand. But habits matter at the aggregate, and some choices are smarter than others.

Use text-based AI for text tasks. If you’re asking ChatGPT a question, the energy cost is comparable to a Google search — don’t stress about it. Where energy costs spike is in generation tasks: creating AI images when a stock photo would do, generating AI videos for social content, or running complex reasoning chains for simple questions. Match the tool to the task.

Choose efficient models where possible. Smaller models — GPT-4o mini, Claude Haiku, Gemini Flash — use a fraction of the energy of frontier models and handle most everyday tasks perfectly well. Reaching for GPT-5 or full reasoning mode for a simple question is the energy equivalent of driving a lorry to pick up a pint of milk.

If you’re a developer or business running AI at scale, this is where it genuinely matters. Batching requests, caching responses, and using the smallest model that solves the problem can cut your AI energy footprint by 70% or more without any loss in output quality. The companies that figure this out now will have a significant cost and sustainability advantage in three years.

Frequently Asked Questions