Remember those old family photographs from the 1980s — slightly faded colours, big hair, simple clothes and that unmistakable film-camera look? Now, you can recreate the same look in seconds.
AI image generators are turning modern photographs into 1980s-style portraits, complete with retro clothes, hairstyles, backgrounds and the grainy look of an old camera. For users, it is mostly a matter of uploading a photo, typing a prompt and waiting a few seconds.
But behind that simple interaction is something most people don’t see: computers working in data centres, consuming electricity and, in many cases, water. A June 2026 report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) estimates that generating a standard-resolution AI image has an electricity-associated water footprint of about 28.6 millilitres. It also estimates that generating the image requires around 2.9 watt-hours of electricity and produces about 1.22 grams of carbon dioxide equivalent.
Twenty-eight millilitres doesn’t sound like much. It is roughly two tablespoons of water. But the number becomes interesting when you stop thinking about one photograph and start thinking about millions of them.
People don’t always generate just one image. They try different prompts, change the clothes, alter the background, regenerate the face or ask the AI to make the image look more realistic. A single person may create several versions before deciding which one to keep.
Now multiply that behaviour across millions of people. That is where AI’s water footprint starts becoming a much bigger question. According to UNU-INWEH, global data centres consumed about 448 terawatt-hours of electricity in 2025. The report expects that figure to rise to 945 terawatt-hours by 2030 as AI and other data-heavy services grow. The associated water footprint could reach about 9.3 trillion litres by then.
And it isn’t just the training of AI models that matters. Once an AI model has been built, it still needs to run every time someone asks it a question, creates an image or generates a video. This process is known as inference. UNU-INWEH estimates that inference could account for around 80% to 90% of total AI energy use.
Images can be particularly demanding. The report estimates that generating a typical AI image can use around 1,450 times more energy than a basic text-classification task. Video generation can require even more computing. The researchers behind the report say the environmental impact needs to be understood at scale.
“AI’s environmental footprint is not just an outcome of physical infrastructure; it is the cumulative result of countless daily decisions,” the report’s co-author Mir Matin said. That is perhaps the simplest way to look at the issue.
Generating one 1980s-style photograph isn’t going to empty a water tank. And the exact amount of water associated with an AI image can vary considerably depending on the model, image resolution, data centre, cooling system and source of electricity.
So the 28.6 ml figure shouldn’t be treated as a fixed price tag attached to every AI image. It is an estimate that gives us a sense of what is happening behind the screen. There is also a much larger water story developing around AI. A 2023 study, “Making AI Less ‘Thirsty’”, estimated that global AI demand could lead to 4.2–6.6 billion cubic metres of water withdrawal in 2027, depending on how AI systems are deployed and operated. The irony is hard to miss.
AI lets us recreate a photograph from four decades ago in a few seconds. The experience feels completely digital — no film, no camera, no physical materials. Yet producing that image still depends on very physical things: servers, electricity, cooling systems, data centres and water. So the next time you turn your selfie into an 1980s photograph, there is more happening behind that retro filter than meets the eye. The image may look like it came from an old film camera.
But somewhere behind it, a modern data centre did the work.
