1980s AI photo trend: How much energy and water does one AI image use? Experts answer

The 1980s are back on social media, this time with the help of artificial intelligence.

Users are uploading selfies to AI tools and turning them into retro portraits with period hairstyles, vintage clothes, old studio backdrops and film-like textures. The trend has spread rapidly across Instagram and other platforms, with people experimenting with multiple versions of the same photograph.

The process looks almost effortless. Upload a photograph, type a prompt and wait a few seconds.

But there is a physical infrastructure behind that digital image.

AI image generation requires computing power, usually from specialised processors housed in data centres. Those machines consume electricity and generate heat, which in turn creates a need for cooling. Depending on where the data centre is located and how it is powered and cooled, the process can also carry a water footprint.

So, how much does one AI-generated image actually use?

The answer is not one fixed number.

One AI image can consume fractions of a watt-hour to several watt-hours

Dr Srinivas Padmanabhuni, CTO of AiEnsured, estimates that generating an AI image can consume roughly 1.2 watt-hours of electricity.

Ritwik Batabyal, CTO & Innovation Officer at Mastek Group, said the number can vary significantly, from fractions of a watt-hour to several watt-hours, depending on the model and how it is deployed.

Jaspreet Bindra, Co-founder and CEO of AI&Beyond, gave a broader estimate of roughly 0.1 to 4 watt-hours per image, again depending on the model and computing setup.

More from Business

Independent research also shows why there is no universal number.

The International Energy Agency has noted that energy consumption varies significantly across AI applications and that image generation is much more energy-intensive than simple text-based tasks. A 2026 United Nations University assessment estimated that a typical AI-generated image requires around 1,450 times the energy of a basic text-classification task.

The IEA has also highlighted the importance of factors such as model size, hardware, utilisation and the way workloads are run. That means two people asking different AI systems to create what looks like the same image can generate different energy footprints.

Why one image is not the real issue

At an individual level, even a few watt-hours is a small amount of electricity.

The problem emerges when image generation becomes a mass activity.

Users rarely create just one image. A first version may be rejected because the face looks wrong. A second prompt may change the clothes. A third may alter the lighting. More versions may be created before the final image is shared.

A viral trend can therefore translate into millions of inference tasks.

The UNU estimates that more than 15 billion images were generated using text-to-image systems during 2022 and 2023, with roughly 34 million AI images being generated each day in 2023. These are historical industry-wide estimates and cannot be used to calculate how many images have been created specifically as part of the current 1980s trend.

There is currently no reliable public number for how many retro 1980s images have been generated through the latest social-media trend. Assigning a total energy or water footprint to it would therefore be speculative.

The water footprint is harder to see

Electricity is only part of the story.

Data centres need cooling, and some cooling systems use water directly. Water can also be associated with the generation of the electricity consumed by the data centre.

The UNU’s 2026 assessment projects that data centres powering AI could require around 945 terawatt-hours of electricity a year by 2030. The associated water footprint could reach about 9.3 trillion litres, while the land footprint associated with the electricity demand could exceed 14,500 square kilometres.

Again, these are figures for the broader AI infrastructure, not the amount of water consumed by one 1980s selfie.

The water required for an individual AI task can vary substantially depending on the location of the data centre, the cooling technology, the local climate, the electricity mix and the efficiency of the infrastructure.

Padmanabhuni said that when AI image creation is repeated at scale, the water used for cooling becomes part of a wider resource question.

‘Digital convenience does not mean zero physical cost’

Batabyal said the debate should not be about asking people to stop using AI.

“Generative AI has made image creation incredibly accessible, but we should not confuse digital convenience with zero physical cost,” he said.

For him, the bigger issue is efficiency across the technology stack.

“We need to look at the entire AI stack — from model architecture and compute to data-centre infrastructure — and continuously improve its efficiency,” Batabyal said.

He argued that as companies move from experimenting with AI to deploying it at scale, responsible AI should also mean resource-efficient AI.

“The next generation of AI should deliver more intelligence with less compute, less energy and, ultimately, less environmental impact,” he said.

Bindra offered a similar view. “AI has made creating an image almost effortless,” he said. “But there is a physical world behind that digital magic.”

He said the environmental cost should not be viewed as a reason to stop using AI, but as a reason to think more carefully about how the technology is designed and used.

“The real measure of progress should not simply be how much AI can create, but how efficiently and responsibly it can do so,” Bindra said.

The AI boom is also an infrastructure story

The concern is growing as AI pushes up demand for data-centre capacity.

The IEA said in April 2026 that global data-centre electricity consumption surged in 2025 and is expected to keep rising sharply as AI expands. Capital expenditure by five major technology companies exceeded $400 billion in 2025 and was expected to rise by another 75 per cent in 2026, according to the agency.

The growth has implications beyond electricity bills.

More computing capacity means more servers and specialised chips, larger data centres, additional cooling systems and greater demand for electricity infrastructure. The UNU has argued that AI therefore needs to be assessed as a physical system involving energy, water, land and supply chains rather than simply as a digital service.

The report estimates that 80 to 90 per cent of AI energy use occurs during inference, when deployed models are actually being used, rather than during the original training process. It also estimates around 2.5 billion ChatGPT prompts a day, underscoring how everyday AI use can accumulate into a substantial infrastructure load.

The cost is not just about the model

Hakimuddin Wadlawala, Founder of Aquila I, said AI photo generation is becoming popular while its infrastructure costs remain largely invisible to users.

“Creating each image requires significant computing power, electricity and cooling, placing pressure on data centres, power networks and water resources,” he said.

Wadlawala said organisations will need to pay greater attention to energy efficiency, more efficient cooling systems and infrastructure that can operate with alternative energy sources.

He also pointed to the need for secure and resilient infrastructure as AI workloads expand.

“The key to a sustainable scale of AI is efficient, secure and resilient infrastructure,” he said.

Does that mean you should stop making AI images?

Not necessarily.

The environmental footprint of one AI-generated image is relatively small compared with many everyday sources of emissions and resource use. The more important question is what happens when AI makes image generation so easy that users create billions of images, often generating several versions before settling on one.

That is where efficiency gains matter.

AI models can be made more efficient. Hardware can improve. Data centres can use different cooling technologies. Operators can choose where facilities are built and what power sources they use. Users and companies can also choose models and workloads that use less computing power for a task that does not require the most powerful system.

The UNU has explicitly argued for more transparency around AI’s environmental footprint and has called for standardised reporting of energy, water and land impacts.

The current 1980s photo craze is therefore unlikely to be remembered as a major environmental event.

Its significance lies elsewhere.

It shows how quickly a computationally intensive technology can become an everyday consumer habit.

A person may see only a retro photograph appearing on a screen. Behind that image are processors, servers, electricity, cooling equipment and, in many cases, water.

One AI image may not use much energy. The challenge is what happens when everyone starts making them.