AI Carbon Footprint: What One Prompt, Image, Video and Model Really Emit

The AI carbon footprint of one text prompt is somewhere between 0.03 and 1.14 grams of CO₂e, depending on who does the counting. Google reports 0.03 g for a median Gemini prompt. Mistral reports 1.14 g for a Le Chat reply.

Both companies published their method, and both numbers are honest. They simply count different things.

That gap of almost 40x is the most useful fact about AI and carbon, and most guides skip it. This one starts there. Then it puts a prompt, a reasoning answer, an image, a video and a full training run on one consistent scale, adds the International Energy Agency's global picture, and works through a heavy user's year. Every figure is dated and attributed. Where a number is our own arithmetic, we say so.

The Key Takeaways

• One text prompt: 0.03 g CO₂e for Gemini on Google's market-based accounting, about 0.09 g on the same prompt at Google's own grid average, and 1.14 g for a Mistral Le Chat reply on a full life-cycle count.
• Accounting is the hidden variable: Google's 2025 electricity emissions were 2.8 million tonnes market-based and 15.1 million tonnes location-based, a 5.4x gap in the same company's own report.
• Tasks scale fast: a long reasoning answer is about 13x a normal prompt, an image about 11x, and one short video from the open WAN2.1 models about 330x to 1,500x.
• Training is the big one-off: 552 t for GPT-3, 8,930 t for Llama 3.1 405B, and an estimated 150,000 t for Grok 4, about 32,600 car-years.
• The global total: AI systems emitted 32.6 to 79.7 million tonnes of CO₂ in 2025 by one peer-reviewed estimate, the same range as New York City.

What Is the AI Carbon Footprint?

The AI carbon footprint is the greenhouse gas emitted to build, train and run AI models, measured in grams or tonnes of CO₂ equivalent (CO₂e). Almost all of it comes from two places. The first is the electricity that data centers draw to power chips, servers and cooling. The second is the carbon locked into making that hardware and building the data centers, called embodied emissions.

Climate accounting splits these into three scopes. Scope 1 is fuel burned on site, such as backup generators. Scope 2 is purchased electricity, and for AI it is the largest share. Scope 3 is everything in the supply chain, including chip manufacturing and data center construction. Google's own measurement of a Gemini prompt found that electricity dominates the per-prompt total, with embodied hardware emissions a smaller slice.

The footprint also splits by phase. Training is a one-off burst that teaches a model, and it is paid once. Inference is every answer the model gives afterwards, and it is paid billions of times a day. Inference is where your own prompts sit. For the electricity side of the same question, see how much energy AI uses per prompt.

Why the Same AI Prompt Is 0.03 g or 1.14 g

AI carbon numbers disagree because companies count different things, not because one chatbot is 40 times dirtier than another. The first choice is which electricity emissions to count, and it alone can triple a figure. The rest comes from how much of the life cycle is included, which model answered and how long the reply was.

Market-Based vs Location-Based Accounting

Market-based accounting credits a company for the clean energy it buys through contracts and certificates. Location-based accounting uses the average emissions of the grid where the electricity is actually consumed. Both are standard under the Greenhouse Gas Protocol. They answer different questions: what a company paid for, and what the grid around its data centers emitted.

Google's 0.03 g per Gemini prompt is market-based. The technical paper behind it states that it uses market-based emissions "for all metrics presented in this paper". The same paper gives Google's 2024 grid factors: 345 g CO₂e per kWh location-based, and 94 g after its clean-energy purchases.

Apply the location-based factor to Google's own 0.24 Wh prompt and the electricity share becomes about 0.083 g. Add Google's reported 0.010 g of hardware emissions (Scope 1 and 3) and the prompt lands near 0.09 g. That is Fello AI arithmetic on Google's published numbers, and it shows that accounting alone triples the figure.

The same split shows up at company scale. Google's 2026 Environmental Report lists its 2025 Scope 2 emissions as 2,815,000 tonnes market-based and 15,148,700 tonnes location-based.

That is a 5.4x difference inside one report, for one year.

Google's electricity use grew 37% in 2025, its largest load growth ever, while its market-based electricity emissions fell.

What Else Moves the Number

Model size, reply length, the boundary and the date explain the rest of the gap. Mistral's 1.14 g for a 400-token Le Chat reply comes from a life-cycle analysis built with the consultancy Carbone 4 and the French ecological transition agency ADEME. Mistral says it used location-based electricity and included hardware manufacturing. The study was peer-reviewed by two audit consultancies, Resilio and Hubblo. You can read Mistral's full disclosure.

Accounting explains about 3x of the gap between Google and Mistral. The rest comes from the model, the hardware, the reply length and how much of the supply chain each company counted. Mistral's own study found that impact rises roughly in step with model size. A model 10 times bigger generates impacts "one order of magnitude larger" for the same number of tokens.

Two more choices matter.

The first is the boundary. Google measured 0.10 Wh per prompt when it counted only the active AI chips, and 0.24 Wh once it added idle capacity, host servers and cooling.

The second is the date. Google says the carbon footprint of its median Gemini prompt fell 44x between May 2024 and May 2025, so a figure from 2023 describes a different system. Google also notes its data has not been verified by an independent third party.

The practical rule is simple. Compare two AI carbon figures only when they share an accounting basis and a boundary. Otherwise you are comparing methods, not chatbots. That matters for any "greenest chatbot" ranking, including the ones built on these two numbers. Google, OpenAI and Mistral have published a per-reply figure for their own chatbots. Microsoft has published a modelled range of 0.16 to 0.60 Wh for the large models behind tools like Copilot, but no measured Copilot figure. We found nothing from Anthropic, xAI, Perplexity or Meta.

AI Carbon Footprint Per Task: The Fello AI Carbon Ladder

A text prompt emits about 0.1 g of CO₂e on a location-based count and an image about 1 g. A short AI video runs from a fraction of a gram to more than 140 g. The Fello AI carbon ladder puts every task on one scale. It multiplies each published energy figure by the same grid factor, Google's 2024 location-based average of 345 g CO₂e per kWh. One factor keeps the rows comparable, so the differences you see come from the task, not the accounting.

The car column uses the EPA figure of about 400 g of CO₂ per mile for an average US passenger car. The CO₂e and car columns are Fello AI arithmetic. The Gemini row adds Google's reported 0.010 g of hardware emissions; the other rows are electricity only.

Text Prompts and ChatGPT

A normal ChatGPT or Gemini text prompt emits roughly 0.1 g of CO₂e on a grid-average basis. Three separate sources now land within 0.1 Wh of each other on the energy. Google measured 0.24 Wh for Gemini. OpenAI chief executive Sam Altman wrote that "the average query uses about 0.34 watt-hours", without publishing a method. A 2026 Microsoft study modelled a median of 0.31 Wh for frontier-scale models on production hardware.

The Microsoft paper also argues that many widely cited public estimates are too high by 4 to 20 times, because they were measured without production batching and serving software. All of its authors work at Microsoft, which runs one of the largest AI clouds, so read that conclusion with the source in mind. For ChatGPT's daily and yearly totals, see ChatGPT's energy use.

Reasoning Models and Long Answers

A long reasoning answer emits about 13 times more than a normal prompt. The Microsoft study found the median rose from 0.31 Wh to 3.91 Wh when a query produced about 5,000 output tokens instead of about 300. Reasoning models spend those extra tokens thinking before they answer, which is why a "think harder" setting costs more carbon as well as more time.

The driver is output length, measured in tokens, not the brand on the model. A short answer from a big reasoning model costs less than a long one from the same model. If tokens are new to you, read what AI tokens are, and see how reasoning models think for why they produce so many.

AI Image Generation

One AI image emits about 1 g of CO₂e on average, roughly ten times a text prompt. The best-known measurement comes from researchers at Hugging Face and Carnegie Mellon, who found that open image models used 2.9 kWh per 1,000 images on average in 2023. The most carbon-intensive model they tested, Stable Diffusion XL, produced 1,594 g of CO₂e per 1,000 images, which the authors compared to driving 4.1 miles.

The viral line that one AI image uses a full phone charge stretches the same paper. The authors wrote that their least efficient model used "around half a charge per image generation". No major image service, including ChatGPT, Gemini or Midjourney, has published a per-image figure, so every number for those products is an extrapolation.

AI Video Generation

AI video has the widest carbon range of any task, from under a tenth of a gram to more than 140 g for one short clip. A September 2025 Hugging Face study measured six open video models on the same hardware. AnimateDiff used 0.14 Wh in total, while WAN2.1-T2V-14B used more than 415 Wh, a factor of nearly 3,000.

The smaller WAN2.1 model used about 90 Wh per short video, which the authors put at roughly 30x an image and 2,000x a text generation. Energy grows with the square of the width, the height and the length. Doubling the length alone takes about four times the compute, and doubling both width and height takes about 16 times. OpenAI's Sora and Google's Veo have no published measurement. If you make AI video regularly, our guide to free AI video generators lists the current options.

How Much CO₂ Does Training an AI Model Emit?

Training one large AI model emits anywhere from 25 tonnes to an estimated 150,000 tonnes of CO₂e. The spread comes from model size and from the grid that powered the run. Only a handful of developers publish a figure, and each draws the boundary differently.

The car-year column divides each figure by the EPA's 4.6 tonnes of CO₂ a year for a typical US passenger car. It is Fello AI arithmetic. Meta's row shows the accounting gap at its most extreme: the same Llama 3.1 training run is 8,930 tonnes on the grid average and zero after Meta's renewable energy matching.

Is AI Training Worse Than Cars?

Training a frontier model emits as much as hundreds to tens of thousands of cars do in a year.

The famous claim that training one AI model equals "five cars over their lifetimes" comes from a 2019 study. It measured a neural architecture search, not a normal training run. A 2021 Google and Berkeley paper later found that the estimate overstated the real search's emissions 88 times, as Google actually ran it.

The bigger story is the trend. Epoch AI estimates that Grok 4 used about 310 GWh of electricity, about 240 times GPT-3's 1,287 MWh. xAI's Colossus data center in Memphis also ran largely on natural gas turbines during that run. Together, that puts Grok 4's estimate at about 270 times GPT-3's, a model trained only five years earlier.

AI Carbon Footprint of Data Centers and the Whole Industry

AI systems emitted between 32.6 and 79.7 million tonnes of CO₂ in 2025, according to researcher Alex de Vries-Gao in the journal Patterns. New York City emitted 52.2 million tonnes in 2023, which sits inside that range. De Vries-Gao, of VU Amsterdam, says the true figure is hard to pin down because tech companies do not publish AI-specific energy or water data.

The International Energy Agency tracks the wider data center sector. Its April 2026 report Key Questions on Energy and AI puts data center electricity use at 485 TWh in 2025, about 1.5% of the world's electricity. The IEA expects that to roughly double, to about 950 TWh by 2030, with AI-focused data centers tripling. Our World in Data, using IEA figures, puts AI-focused data centers at 155 TWh in 2025, or 0.49% of global electricity.

AI Carbon Footprint Compared to Other Industries

AI is still a small share of global emissions, but one of the fastest-growing. The IEA's 2025 Energy and AI report put data center electricity emissions at about 180 million tonnes a year. It said they stay below 1.5% of energy-sector emissions, while calling data centers "among the fastest growing sources of emissions". That report projected 300 million tonnes by 2035 in its base case. The IEA's April 2026 update raised the projection to about 350 million tonnes, around 2% of global electricity-sector emissions.

Single companies now report totals in the millions of tonnes. Google reported total emissions of about 14.5 million tonnes for 2025 under its own "ambition-based" method, up 18% on 2024. It attributed most of the rise to its supply chain, including building data centers. AI also has a water footprint that follows the same accounting arguments.

Is Your Own AI Use a Big Deal?

For text, no.

A heavy user sending 100 prompts a day for a year sends 36,500 prompts. At about 0.09 g each on Google's grid-average figure, that is roughly 3.4 kg of CO₂e, about the same as driving 8.5 miles. On Google's market-based figure it is about 1.1 kg. Both totals are Fello AI arithmetic.

The rest of a personal footprint is counted in tonnes, not kilograms. Food, flights and home heating make up most of it. Switching from a high-meat diet to a vegan one saves about 2.8 tonnes of CO₂e a year, based on a 2023 Oxford study, which is roughly 800 times that heavy AI year. For the full picture, see this guide to what a carbon footprint is and how to reduce it.

Video is the exception. One WAN2.1-14B clip at about 143 g equals about two weeks of that heavy text habit. Ten clips a day would emit over half a tonne a year, around one-ninth of a car-year.

The bigger lever is collective. Billions of daily prompts, rising video use and new training runs add up to the city-sized totals that de Vries-Gao and the IEA report. Your own choices matter most where the per-task numbers are largest: long reasoning, images and video.

How to Reduce Your AI Carbon Footprint

The single most effective way to cut your AI carbon footprint is to match the model to the job. Mistral's study found impact rises roughly tenfold for a model ten times bigger, and Microsoft's found a long reasoning answer costs about 13 times a normal one. Five habits follow from the data.

1. Use a small, fast model for small jobs. Save the flagship model for work that needs it. Quick rewrites, summaries and lookups run fine on the lighter models.
2. Turn on reasoning only when you need it. Extended thinking is the biggest multiplier in text. Leave it off for simple questions.
3. Ask for shorter answers. Output tokens drive the energy. Asking for three sentences instead of a full essay cuts both the reading time and the footprint.
4. Generate fewer, better images and videos. Write one careful prompt instead of ten rough ones. Keep video short and at the resolution you will actually use.
5. Batch related questions into one prompt. Mistral lists grouping queries as one of its own recommended user practices.

Matching the model to the job is easier when you do not need a second app or a second bill to do it. Fello AI lets you send the quick question to one model and the hard problem to another in the same window, with GPT, Claude, Gemini and DeepSeek behind one subscription. If you want to compare the vendors themselves, see our ranking of the most eco-friendly AI models.

The Verdict

The AI carbon footprint is small per prompt and large in total, and the counting method alone can move a headline number three to five times. A text prompt is about a tenth of a gram on a fair grid-average basis. Training runs and video are where the tonnes are, and the trend is up. Google's location-based emissions grew more than a third in one year, and AI-focused data centers are set to triple their electricity use by 2030.

So treat any single "grams per prompt" figure as half a fact until you know its accounting basis. For the bigger question of whether that trade-off is worth it, read whether AI is bad for the environment.

FAQ