Lebanon lacks power, but your portrait looks great

Turning a recent selfie into an 1980s studio portrait takes seconds. The computing infrastructure behind it does not.

How much electricity an AI-generated image requires depends heavily on the model, hardware and image resolution.

A 2025 study by researchers at the University of Florence measured the energy consumption of 17 diffusion-based image-generation models under controlled conditions. It found consumption ranging from 0.086 watt-hours (Wh) for the most efficient model to 4.08 Wh for the least efficient, a difference of up to 46 times. Across the models tested, doubling image resolution increased energy consumption by between 1.3 and 4.7 times, depending on the model. The researchers found that longer prompts did not significantly increase energy use under their test conditions.

The finding matters because there is no single number that represents the energy cost of an “AI image”. The model, hardware, resolution and generation process all matter.

A separate benchmark published in 2024 found an average of 2.907 Wh per generated image across the image-generation systems it tested, with the least efficient system reaching 11.49 Wh. In the same benchmark, image generation consumed substantially more energy than text generation.

For a single 1980s-style portrait, these numbers may look small. At scale, they are not.

700 million images in one week

OpenAI's 2025 launch of native image generation in ChatGPT provides a measure of what that scale can look like.

More than 130 million users generated over 700 million images during the first week after the feature was introduced, according to OpenAI. That works out to an average of roughly 100 million images a day during that first week.

Applying the 0.086 to 4.08 Wh range measured in the University of Florence study purely as an illustration would put 100 million images at roughly 8.6 to 408 megawatt-hours.

At the study's highest observed consumption, that is more than 400 megawatt-hours for a single day's worth of 100 million images.

This is not an estimate of OpenAI's actual electricity consumption. The University of Florence study tested different models under controlled conditions, not OpenAI's production infrastructure. The calculation illustrates the potential scale of electricity demand rather than measuring the electricity used by ChatGPT.

The distinction is important because much of AI's environmental footprint remains difficult to measure externally. Companies generally do not publish electricity and water consumption for individual image generations, while researchers use different methodologies and assumptions.

Water is harder to count

Electricity is only part of the calculation.

Data centres generate heat and require cooling. Depending on the cooling system and location, that can involve water directly. Water is also used indirectly to generate some of the electricity that powers data centres.

Researchers therefore distinguish between water withdrawal, the amount taken from a source, and water consumption, the portion not returned to the same source, often because it evaporates.

One influential study on AI's water footprint estimated that training GPT-3 in Microsoft's US data centres could directly consume around 700,000 litres of freshwater through onsite cooling. When indirect water use associated with electricity generation was included, the estimate was substantially higher.

The same research estimated that global AI-related water withdrawal could reach between 4.2 billion and 6.6 billion cubic metres annually by 2027 under its scenarios, including direct data-centre use and water associated with electricity generation.

Those figures refer to AI as a whole, not image generation alone. They are also projections based on assumptions about future AI demand and infrastructure, rather than measurements of actual water consumption today.

That is why claims that a particular AI image “uses X litres of water” should be treated cautiously. The answer depends on where the image is generated, which model is used, the data centre's cooling technology, the electricity source and whether direct or indirect water use is being counted.

The uncertainty does not mean there is no footprint. It means the footprint is difficult to attribute to an individual image.

38% of Lebanon's internal freshwater resources

That distinction matters more in Lebanon than it might appear.

Lebanon has relatively high precipitation by regional standards, but rainfall is heavily concentrated in the winter months, while insufficient storage creates a major seasonal water gap.

Climate change is expected to make that gap worse. World Bank projections indicate that dry-season water availability could fall by around 50% by 2040, while higher temperatures, changing rainfall patterns, reduced snowpack and increased evaporation will place additional pressure on water resources.

The pressure is already visible in the country's water statistics. The latest World Bank data, for 2022, show annual freshwater withdrawals equivalent to about 38% of Lebanon's internal freshwater resources.

In Greater Beirut and Mount Lebanon, the problem is particularly acute. The World Bank says the public water system currently meets only around 24% of demand during the dry season for the roughly 1.8 million people covered by its Greater Beirut water project.

The gap pushes households towards private water sources, including tankers that the World Bank says can cost up to ten times more than piped water.

5.34 TWh of electricity, but a system still under strain

AI also arrives in Lebanon through an electricity system that has struggled to reliably meet existing demand.

Lebanon generated about 5.34 terawatt-hours of electricity in 2024. Around 56% came from fossil fuels, while solar accounted for about 30% and hydropower another 13%, according to data compiled by Ember.

The renewable share is significant, but it does not erase the underlying reliability problem.

The World Bank's assessment of Lebanon's energy sector found that Electricité du Liban was providing only around four to five hours of electricity a day in 2023, after falling to one to two hours in 2022. Households and businesses therefore continued to depend heavily on private generators and other sources of electricity.

Every AI image generated by a Lebanese user does not mean that Lebanon's electricity grid is directly powering the server that produced it. Most commercial AI services run their models in data centres located elsewhere.

That makes the Lebanese connection less direct, but not irrelevant.

The environmental cost of an AI image is determined partly by the infrastructure that serves it. Lebanon's own resource constraints therefore matter when considering what kind of AI infrastructure the country might build in the future.

A country planning for more AI

That future is already being discussed.

Lebanon adopted its National Digital Transformation Strategy for 2020–2030 in 2022. More recently, the government has begun placing artificial intelligence and advanced digital infrastructure more explicitly at the centre of its economic agenda.

In 2025, Lebanon created the position of Minister of State for Technology and Artificial Intelligence. The ministry has described AI as part of a broader effort to build a digital economy, with the country's agenda including investment in skills, data infrastructure, public-sector applications and entrepreneurship.

In September 2026, Lebanon also signed the Digital Cooperation Organization Charter, opening a framework for cooperation covering artificial intelligence, digital policy, skills, entrepreneurship, infrastructure and innovation.

The country is therefore trying to build an AI ecosystem at the same time that it is trying to repair the basic infrastructure on which any digital economy ultimately depends.

The footprint is small per image, until it isn't

An individual AI 80s portrait is not going to determine Lebanon's water future.

The more important question is what happens when image generation becomes routine.

If 100 million images require anywhere between 8.6 and 408 megawatt-hours under the range observed in one 2025 study, the difference between an efficient and inefficient model becomes enormous at scale.

And the number of images being generated is still growing. OpenAI says users are now creating more than three billion images every week across ChatGPT Images and its GPT-Image API models.

At that scale, efficiency becomes an infrastructure question rather than an individual-user question.

The same principle applies to water. There is no reliable figure for the amount of water behind one Lebanese user's 1980s photograph. But AI's overall water footprint is measurable in parts and growing, even if the exact allocation to each image remains uncertain.

For Lebanon, the issue is therefore not whether one AI portrait is “wasting” a bottle of water. It is whether a country already facing seasonal water shortages, unreliable electricity and growing climate pressure can afford to ignore the resource requirements of the digital infrastructure it hopes to build.

The 1980s photo may disappear from a user's feed within days. The infrastructure required to generate millions of them will not.