AI's own footprint: does generative AI eat fashion's sustainability gains?

In August 2025 Google published a number that was meant to settle an argument. The median text prompt in its Gemini Apps, the company said, uses 0.24 watt-hours of electricity, emits 0.03 grams of CO2 equivalent and consumes 0.26 millilitres of water, roughly five drops. Three months earlier, MIT Technology Review had measured an open-source video model and found that a single five-second clip took about 3.4 million joules, more than 700 times the energy of a high-quality image. Both figures are correct. They describe very different uses of the same technology.

That gap is the heart of the question for fashion. The industry has spent a decade promising to cut emissions, water use and waste, and it is now adopting generative AI for design, imagery, search, forecasting and customer service. The honest answer to whether AI eats those gains is: not at the level of a single prompt, and not yet at industry scale, but the answer depends on what a company uses AI for, how often, and whether the tools actually reduce physical production and shipping. The hardware footprint is real and growing. The savings are plausible but mostly unmeasured.

This matters now because the numbers are moving in opposite directions. Per-query efficiency is improving quickly, while total demand for computing is rising even faster. Microsoft and Google both reported in 2025 that AI and cloud growth pushed parts of their footprints up. Fashion companies that report Scope 3 emissions will increasingly have to account for the computing they buy, and to show that it pays its way.

How much electricity does AI actually use?

The most comprehensive public estimate comes from the International Energy Agency. Its Energy and AI report, published in April 2025, put global data centre electricity consumption at about 415 TWh in 2024, around 1.5% of global electricity use. In its base case, that figure more than doubles to about 945 TWh by 2030, slightly more than Japan uses today. AI is the main driver: according to Scientific American's reading of the report, AI servers accounted for 15% of total data centre energy demand in 2024, and the United States, Europe and China together accounted for 85% of data centre consumption.

In emissions terms, the IEA estimates that data centres produce about 180 Mt of CO2 a year from the electricity they use, excluding backup generation. That is about 0.5% of global CO2 from fuel combustion. In the base case, those indirect emissions rise by almost 80% over the decade to about 1% of combustion emissions; in a faster "Lift-Off" case they grow 2.5 times to 1.4%, the largest emissions growth of any sector in that scenario. The agency is careful to frame these as scenarios rather than forecasts, and outside researchers think AI's share may be understated. Alex de Vries of VU Amsterdam told Scientific American that the AI growth "could be a serious risk for our ability to achieve our climate goals".

What does one AI prompt, image or video cost in energy and water?

Per-task figures vary by an order of magnitude or more, depending on the model, the hardware, the data centre and what is counted. Google's 0.24 Wh figure, published by Amin Vahdat and Jeff Dean in August 2025, includes idle machines kept on standby, host processors and memory, and data centre overhead such as cooling. Counting only active AI chips, Google says the same prompt would be 0.10 Wh, a method it calls "an optimistic scenario at best". The company also reports that energy per median prompt fell 33 times and carbon per prompt 44 times between May 2024 and May 2025. It notes that the data have not been verified by an independent third party.

OpenAI's chief executive Sam Altman wrote in June 2025 that an average ChatGPT query uses about 0.34 Wh and 0.000085 gallons of water, roughly one fifteenth of a teaspoon, but did not explain how the figures were calculated. Independent measurement is only possible for open models. MIT Technology Review, working with researchers, found that Meta's Llama 3.1 405B used about 6,706 joules per response, an image from Stable Diffusion 3 Medium about 2,282 joules (around 0.63 Wh), and a five-second clip from the video model CogVideoX about 3.4 million joules, equivalent to running a microwave for over an hour. Sasha Luccioni of Hugging Face, quoted in that investigation, put the problem bluntly: "We should stop trying to reverse-engineer numbers based on hearsay."

For a fashion business, the practical reading is that text work is cheap per task, image generation costs a few times more, and video is in a different class. A merchandising team of 50 people each sending 100 Gemini-class prompts a working day would use, on Google's figure, about 1.2 kWh a day. A single five-second video from the model MIT tested uses close to that amount. The choice of task matters more than the choice of chatbot.

Why are tech companies' emissions rising if each prompt is getting cleaner?

Because volume is growing faster than efficiency. Microsoft's 2025 sustainability report says its total emissions across Scopes 1, 2 and 3 were 23.4% above its 2020 baseline, driven by "growth-related factors such as AI and cloud expansion". Its energy use rose 168% over the same period. Direct and purchased-energy emissions fell 29.9%, but Scope 3 emissions, those in its value chain, rose 26%. Google's 2025 environmental report showed data centre electricity demand up 27% in 2024; it said data centre energy emissions nevertheless fell 12%, while supply chain (Scope 3) emissions rose 22%.

This is the familiar rebound effect. Cheaper and cleaner queries invite more queries, larger models and new uses such as video and autonomous agents. The IEA notes that rebound effects could offset AI's efficiency gains elsewhere in the economy. For fashion buyers of AI services, it means a supplier's falling per-query footprint does not guarantee a falling total, and the emissions embodied in chips and buildings will appear in their own Scope 3 inventories over time.

There is a geographic dimension as well. The IEA expects the United States and China to account for nearly 80% of data centre electricity growth to 2030, and grid mixes differ widely between and within those countries. The same workload can therefore carry very different emissions depending on where a provider runs it, which is information few fashion buyers of AI services currently ask for.

How does AI's footprint compare with fashion's own emissions?

Fashion is a far larger emitter. McKinsey and the Global Fashion Agenda estimated in 2020 that the industry produced about 2.1 billion tonnes of greenhouse gases in 2018, about 4% of the global total. The United Nations, citing a wider range of estimates, said in March 2025 that the industry is responsible for up to 8% of global greenhouse gas emissions and uses about 215 trillion litres of water a year. On those numbers, fashion's emissions are roughly 12 times the 180 Mt of CO2 that the IEA attributes to the electricity of every data centre in the world, although the two estimates use different years, scopes and methods and should not be read as a precise ratio.

Water tells the same story at a different scale. On Google's figure of 0.26 mL per median prompt, a million prompts would consume about 260 litres of water. The UN's estimate for the fashion industry is about 215 trillion litres a year, equal to 86 million Olympic-sized swimming pools. Per-prompt water figures leave out the water used to generate the electricity and to make the chips, so the true total is higher, and data centres can strain local supplies in dry regions even when their global share is small. But for a fashion company deciding where to focus, the water in its fibres and mills dwarfs the water in its software.

Most of fashion's footprint sits upstream, in fibre production, dyeing and garment manufacture, which is why the main abatement levers are physical: renewable energy in mills, better materials and producing fewer unsold goods. McKinsey's analysis found that only 60% of garments are sold without a markdown and estimated that brand operations, including cutting overproduction, could abate 308 Mt of CO2 equivalent in 2030. That is where AI could matter most, if forecasting and assortment tools translate into smaller production runs rather than just faster ones.

Where can AI cut fashion's footprint?

The strongest evidence so far is in returns. Zalando, the Berlin-based online fashion platform, says online fashion return rates in Europe can reach about 50%, with size and fit responsible for up to half of them. Its size and fit tools, which combine size flags, recommendations, a body measurement tool tested by more than 1.5 million customers and a virtual fitting room, prevented 8% of size-related returns in 2025, according to a company update in June 2026. Pilots of the virtual fitting room reduced return rates by up to 40%, and Zalando aims to roll it out to all customers this year. These are company figures and do not report the carbon saved.

Returns are concentrated, which makes targeted tools more valuable. A study of almost 10,000 European consumers by Vrije Universiteit Brussel, reported by FashionUnited in June 2026, found that 15% of consumers who return excessively account for almost 60% of return-related CO2 emissions, producing almost 20 kg of CO2 a year each through returns compared with about 3 kg for other shoppers. Vendors of sizing software make larger claims for their products, but few have been independently audited.

The second lever is overproduction. The IEA estimates that existing AI applications in end-use sectors could cut about 1,400 Mt of CO2 in 2035 if widely adopted, several times more than data centres would add. But the agency adds a warning that applies directly to fashion: "there is currently no momentum that could ensure the widespread adoption of these AI applications." Better forecasts only cut emissions if buying and production plans change in response.

Does AI-generated imagery save more than it costs?

Imagery is where fashion's AI use is most visible. Zalando told Reuters in May 2025 that around 70% of its editorial campaign images in the fourth quarter of 2024 were AI-generated, cutting production time from six to eight weeks to three to four days and costs by 90%, according to Matthias Haase, its vice-president of content solutions. Haase framed the change as a matter of speed and relevance: "We are using AI to be able to be reactive." Zalando did not publish an environmental comparison, and none of the sources reviewed for this article quantify the footprint of a conventional fashion shoot.

The arithmetic suggests that still images are a small item. On MIT Technology Review's measurement, 1,000 images from Stable Diffusion 3 Medium would use roughly 0.63 kWh, less than a single five-second video from the model MIT tested. The risk lies elsewhere: faster, cheaper content invites far more of it, and the shift to AI video multiplies energy use by hundreds. Companies that replace one shoot with thousands of generated variants and videos may end up with a larger computing bill than the travel they avoided, without anyone having measured either side.

What are the risks, limits and counter-arguments?

• The data are thin and company-controlled. The most detailed per-prompt figures come from Google and OpenAI themselves, and only Google has published a methodology. Neither discloses total query volumes, so per-query numbers cannot be scaled to a real footprint.
• Accounting choices change the answer. Google's carbon figure uses market-based accounting that reflects its clean power purchases. MIT Technology Review cites a Harvard preprint finding the electricity used by US data centres was 48% more carbon intensive than the US average.
• Water and hardware are undercounted. Per-prompt water figures usually cover cooling only, not the water used to generate electricity or make chips, and it was Scope 3 emissions, up 26% at Microsoft, that pushed its total up.
• The savings are mostly potential. Fashion's main AI benefits, lower returns and less overproduction, are rarely reported in tonnes of CO2, and the IEA's 1,400 Mt depends on adoption that is not yet happening.
• The counter-argument has weight. For text and analytics, the footprint per task is genuinely small next to a garment's. A team that uses AI to avoid producing a few thousand unsold units almost certainly comes out ahead.

What should fashion companies do now?

1. Inventory your AI workloads List which teams use text, image and video generation, and roughly how often, before setting any targets.
2. Ask suppliers for disclosed figures Request per-task energy, water and emissions data with methodology, including whether figures are market-based or location-based.
3. Treat video as a separate budget Generated video uses hundreds of times more energy than images; require a business case for it.
4. Measure savings in physical units Track returns avoided, units not produced and markdowns cut, then convert them to CO2 with the same rigour as emissions.
5. Link forecasts to buying decisions Make sure better demand predictions reduce order quantities, not only lead times.
6. Report AI in Scope 3 Include purchased cloud and AI services in emissions inventories so that rebound effects become visible over time.