The 1980s Are Back Online. But AI Nostalgia Has an Environmental Cost
The 1980s have returned to social media—this time through artificial intelligence.
Across social platforms, users are turning ordinary photographs into retro portraits featuring period clothing, hairstyles, colours and photography styles. The transformation can take only seconds, creating the impression that the process is almost weightless.
It is not.
Behind every generated image is a chain of computing infrastructure: powerful processors, data centres, electricity networks, cooling systems and hardware. The environmental impact of a single image may be relatively small, but the growing scale of generative AI is making that hidden infrastructure increasingly important.
## A nostalgic image still needs a data centre
When a user submits an image-generation prompt, the request is processed by computers in a data centre. Generative models perform large numbers of mathematical calculations before producing the final image.
The process therefore requires electricity and generates heat that must be managed through cooling systems.
The International Energy Agency estimates that data centres consumed around **415 terawatt-hours of electricity in 2024**, equivalent to about 1.5 percent of global electricity consumption. In its base-case outlook, the IEA expects data-centre electricity demand to roughly double to about **945 TWh by 2030**. AI-driven accelerated computing is expected to be a major contributor to that growth.
That does not mean the entire increase will come from AI images or consumer applications. Data centres support many services, including cloud computing, storage, websites and conventional digital workloads.
But the rapid expansion of AI is changing the type and intensity of computing taking place inside them.
## Images are more demanding than simple text
AI workloads do not all consume the same amount of energy.
Research by Sasha Luccioni, Yacine Jernite and Emma Strubell compared the energy requirements of different AI tasks. Their study found substantial variation between models and tasks, with image generation among the more energy-intensive workloads examined. The study reported a mean of about **2.9 kWh per 1,000 image-generation inferences**, equivalent to roughly **2.9 Wh per inference** in that particular experimental setup.
The figure should not be interpreted as a universal electricity bill for every AI-generated picture.
Different models can require very different amounts of computation. Resolution, model architecture, hardware and other technical factors can significantly change the result. The researchers found large variation even among systems performing similar tasks.
This is particularly relevant to viral image trends because users rarely stop after one generation.
They may change the hairstyle, clothing, lighting, background or facial details and generate another version.
Then another.
The environmental question is therefore less about one nostalgic photograph than about millions or billions of repeated AI interactions.
## Water is part of the equation
Electricity is only one part of the environmental footprint.
Data centres produce substantial heat, and some cooling systems use water. Water can also be associated indirectly with electricity generation and with manufacturing the hardware used in computing infrastructure.
Research led by Shaolei Ren has highlighted the importance of considering water alongside carbon emissions when assessing AI's environmental footprint. The researchers argue that water efficiency can vary according to where and when computing takes place, meaning that the environmental consequences of AI infrastructure are not identical everywhere.
Recent estimates cited in reporting on the current AI-image trend put the electricity-associated water footprint of a typical AI-generated image at around **29 millilitres**. That is an estimate, not a fixed quantity of water physically consumed by every image-generation request. The actual figure can change considerably depending on the technology and location.
At individual scale, two tablespoons of water may sound insignificant.
At the scale of a rapidly expanding global technology industry, the calculation becomes more complicated.
## Nepal is part of a global system
The environmental consequences of AI use in Nepal are not limited to Nepal.
Most AI services used by Nepali consumers rely on computing infrastructure located outside the country. The electricity, cooling resources and hardware used to process those requests may therefore be consumed elsewhere, while the associated environmental impacts remain part of the global footprint of digital technology.
Nepal nevertheless has a particular reason to pay attention.
The country has rapidly expanded its digital and social-media participation, while its electricity system is increasingly capable of generating large amounts of hydropower.
That creates an opportunity for Nepal to position itself as a destination for lower-carbon digital infrastructure.
But abundant electricity alone does not automatically make a data centre environmentally sustainable.
The source of electricity matters, but so do water availability, cooling technology, hardware efficiency, land use and the location of facilities.
## Green electricity does not solve everything
Nepal's hydropower potential could provide an important advantage if the country eventually develops data-centre infrastructure.
However, calling a data centre “green” simply because it uses renewable electricity would overlook other environmental costs.
A facility still requires servers, networking equipment, cooling infrastructure and buildings. Semiconductor manufacturing and hardware supply chains also carry environmental footprints.
The IEA notes that servers account for a large share of data-centre electricity demand, while cooling and other infrastructure add significant additional consumption. In modern data centres, the relative share varies considerably depending on the facility's design and efficiency.
Sustainable AI therefore requires more than renewable power.
## The responsibility cannot rest only with users
It would be easy to tell people to stop generating AI portraits.
That would miss the larger issue.
An individual user has little control over where a model is hosted, what hardware it uses, how efficiently the data centre operates or whether its cooling system relies heavily on water.
Technology companies make many of those decisions.
That makes transparency important.
Users should increasingly be able to understand the environmental cost of the services they use, while companies should have incentives to improve energy efficiency, reduce water consumption and publish meaningful environmental measurements.
The original research on AI deployment costs has also highlighted how different AI models can have dramatically different energy requirements. More efficient models and hardware can therefore reduce the resource intensity of the same digital task.
## Convenience creates a scale problem
The appeal of the 1980s trend is easy to understand.
It is creative, personal and entertaining. A person can experiment with a different hairstyle or fashion era without buying clothes, arranging a photoshoot or learning image-editing software.
The environmental concern is not that people are enjoying technology.
It is that making computationally expensive activities almost effortless can encourage far more of them.
This is a familiar challenge in digital technology: when something becomes faster, cheaper and easier, people often use it more.
For AI, that could mean a future in which billions of images, videos and other generated media are created simply because the cost to the individual user feels negligible.
## Nepal needs a sustainable-AI conversation
Nepal is still at an early stage of building its domestic AI ecosystem.
That gives the country an opportunity to think about sustainability before large-scale infrastructure decisions become difficult to change.
If Nepal attracts data centres or develops domestic AI-computing capacity, environmental criteria should be considered alongside investment and electricity availability.
Possible priorities include efficient cooling systems, responsible water management, renewable electricity, energy-efficient computing hardware, transparent environmental reporting and careful selection of data-centre locations.
The country can also encourage universities, startups and technology companies to explore “green AI”—developing models and applications that achieve useful results with less computational demand.
## The real cost is the scale
An AI-generated 1980s portrait is not an environmental disaster.
Neither is a single chatbot question.
The larger issue is cumulative demand.
The infrastructure supporting AI is expanding rapidly, and data-centre electricity consumption is expected to grow substantially over the coming decade.
The nostalgic photograph is simply an unusually visible example of an invisible process.
What looks like an old-fashioned photograph on a smartphone may have travelled through modern fibre networks, powerful processors and cooling systems thousands of kilometres away before appearing on the screen.
The lesson is not that society should reject AI.
It is that digital convenience should no longer be treated as environmentally weightless.
The 1980s may be back in fashion. But as AI becomes part of everyday life, the technology powering that nostalgia needs to become more transparent, efficient and sustainable.
