Every time you press ‘send’ on ChatGPT, you’re unknowingly drawing as much power as a lightbulb burning for two minutes, and that’s just the tip of the melting iceberg. I am an avid user of AI these days and our company 4 Roads builds all sorts of things from VisionAI to custom LLMs, but we have to recognise the impact.
I’ve recently been building up some AI models for classification within communities, a story told in Vision and LLMs for moderation and labelling. We have a fairly powerful server that we use for this sort of thing that has a couple of NVIDIA A10 cards that we bought for the BEIS / DESNZ funded heat pump project we worked on. I also have a fairly good laptop that has an RTX 4060 onboard so I can also do some training locally. What I now appreciate, doing things locally, is the amount of energy that is consumed during training, my laptop is practically levitating trying to dissipate the heat.
A single search on ChatGPT uses around 0.34Wh. That is about the same as keeping an efficient lightbulb on for a couple of minutes. And that’s not including the cost of being polite to it, the “please” in a request on its own uses a fantastic amount of energy.
When you ask ChatGPT to write that report for you it uses 0.000085 gallons of water to help cool a data centre.
These numbers have come from a post Sam Altman wrote on his blog. I have no idea how he has deduced these numbers but if we take them as real then on cooling alone, ChatGPT is using 85,000 gallons (322 m3) of water a day.
That is right, there are a billion messages to ChatGPT a day.
At 1 billion queries per day, that’s 85,000 US gallons (322 m3) of cooling water every day, about an eighth of an Olympic swimming pool. Or to put this into perspective of personal consumption, our 4 bed house with 3 people uses 29,323 gallons (111 m3) a year
This cooling and energy assessment ignores the fact that AI images require much more energy and cooling to process than text queries.
The good news is that technology advancements are making these power hungry searches more efficient.
Even with the efficiencies, it is a lot of energy and water wastage for us to create pictures of us looking like Barbie characters. When the Studio Ghibli memes came around, GPUs were reportedly melting.
If unchecked, our AI curiosity could outpace the energy grids and water supplies we depend on. We must figure out greener infrastructure and more efficient models, before these conveniences become unsustainable.
What if every AI request carried a visible ‘energy cost’ label?
What if the cost of labour is actually cheaper than cost of using AI for a task?
Maybe Builder.ai "Natasha" had it right?
First published on LinkedIn on 16 July 2025.
