Every time you ask ChatGPT a question, somewhere a massive computer is running at full tilt. Actually, many massive computers. And they need electricity. A lot of it.
This isn't a secret, but it rarely gets specific. Let's change that.
The numbers (as far as we know)
Exact figures are hard to come by — the major AI providers aren't particularly forthcoming on this. But researchers have published estimates that roughly look like this:
- One Google search: around 0.3 watt-hours (Wh)
- One ChatGPT request (short answer): around 1–10 Wh, depending on length and complexity
- A longer conversation with several messages: correspondingly more
The often-cited ballpark figure: ChatGPT uses roughly 10 times as much energy per request as a normal Google search.
That sounds dramatic. It is — but it needs context.
For comparison: what else uses electricity?
- An LED light bulb (10W): in one hour = 10 Wh
- HD video streaming: around 1 Wh per minute
- Sending an email with an attachment: around 50 Wh (that's the whole system around it)
- A computer running: 50–150 W, so 50–150 Wh per hour
A ChatGPT request is comparable to a few minutes of video streaming. Not zero, but not apocalyptic either.
The real problem: scale
The problem isn't the individual request. The problem is that millions of people make millions of requests every day.
OpenAI reported over 100 million daily active users in 2024. If even a fraction of those send several requests per day, that adds up to a serious energy problem fast.
Then there are training runs: training one large model is estimated to consume thousands to tens of thousands of megawatt-hours — equivalent to the annual consumption of hundreds of households.
Don't forget water and CO2
Data centers don't just need electricity — they need cooling. And that often runs on water. Microsoft admitted in 2023 that its water consumption had risen sharply due to AI training — specifically: several million liters for training GPT-4.
On CO2: it depends on where the electricity comes from. In regions with lots of renewable energy, the picture looks better. In coal-heavy regions, worse. Some of the large data centers now use green energy — some don't.
When is it "worth it"?
That's the actually interesting question. An honest answer:
AI makes sense energy-wise when:
- The task is more complex than a simple search
- The AI saves real time that would otherwise go into research
- The result is qualitatively better than a quick Google search
AI is overkill when:
- You're asking what the weather will be tomorrow (weather app)
- You want to solve a simple math problem (calculator)
- You want to look up a word (dictionary)
Sounds obvious. But take a look at how often you actually use AI for things that could be solved much more simply.
No panic, but no "who cares" either
The goal here isn't to make anyone feel guilty. AI has real value and the technology is becoming more efficient.
But: "other things use electricity too" isn't an argument for ignoring consumption. And if you use AI daily, you should at least know what that means.
Mindful use is worth more than a guilty conscience.
Numbers based on publicly available research estimates (including Goldman Sachs Research 2024, International Energy Agency 2024). Exact figures vary by model and request type.