greenai.guide

Home › Articles › What is AI's carbon footprint?

Article

What is AI's carbon footprint?

Updated · 8 min read

In short

AI's carbon footprint is the greenhouse gas emitted to make the hardware and buildings AI runs on (embodied carbon) and to generate the electricity that trains and runs its models (operational carbon). Per text prompt it is small: Google reports 0.03 g of CO₂ equivalent for a median Gemini prompt. But AI emissions add up: training GPT-3 was estimated at 552 tonnes, and Microsoft's total emissions rose from about 12.9 to 20.3 million tonnes between 2020 and 2025 as it built out AI. Where the electricity comes from matters as much as how much is used.

Key points

  • Google reports 0.03 g CO₂e for a median Gemini text prompt in May 2025, market-based, a 44-fold fall in a year; about a third of it is from hardware and other non-electricity emissions.
  • Mistral AI's life-cycle assessment put a 400-token Le Chat response at 1.14 g CO₂e, and training plus 18 months of use of Mistral Large 2 at 20,400 tonnes.
  • Training GPT-3 was estimated at 552 tonnes CO₂e; Meta reports 11,390 tonnes location-based for the Llama 3.1 family, and zero market-based.
  • Microsoft's emissions rose from about 12.9 million tonnes in its 2020 financial year to about 20.3 million in 2025.
  • The choice of model, data centre and processor can cut the carbon of training by around 100 to 1,000 times, according to Patterson and colleagues.
  • US data centres' 2023 electricity was linked to about 74 million tonnes CO₂e, an average of 0.34 kg per kWh, according to Lawrence Berkeley National Laboratory.

What is AI's carbon footprint? It is the greenhouse gas released to build the hardware and buildings AI runs on, and to generate the electricity that trains and runs its models. Per prompt, the AI carbon footprint is small; in total, AI emissions are growing fast, because data-centre electricity use and construction are growing fast. This page sets out the published figures, why they differ, and how to estimate your own.

How big is AI's carbon footprint per prompt?

For a short text prompt, a fraction of a gram of CO₂ equivalent by the providers' own accounts; more when the whole life cycle is counted. CO₂ equivalent (CO₂e) adds other greenhouse gases to carbon dioxide, weighted by their warming effect.

Figure What it covers Source
0.03 g CO₂e per median Gemini text prompt Electricity (market-based), plus hardware and other emissions; May 2025 Google, 2025
1.14 g CO₂e per 400-token Le Chat response Life-cycle assessment of inference, excluding the user's device Mistral AI, 2025
about 0.34 Wh per average ChatGPT query Energy only; no carbon figure published OpenAI chief executive, 2025

There is no official carbon footprint of ChatGPT per query: OpenAI has given an energy figure but not a carbon one, and the carbon depends on the grids its data centres draw on.

Google's figure breaks down into 0.023 g from electricity and 0.010 g from its Scope 1 and Scope 3 emissions, which include manufacturing the hardware. It also reports a 44-fold fall in carbon per median prompt over one year. Mistral's figure is larger partly because it counts more of the life cycle and uses a different method; the two are not directly comparable.

Heavy use changes the picture. Zeke Hausfather estimated that his own intensive use of an AI coding agent, about 1.1 MWh of electricity a year, would emit about 370 kg CO₂e a year. See how much energy does AI use?

What is the difference between operational and embodied carbon?

Operational carbon comes from the electricity used to run AI; embodied carbon comes from making, shipping and disposing of the hardware and buildings. Both belong in an honest footprint.

  • Operational carbon depends on how much energy the work uses and how clean the electricity is.
  • Embodied carbon comes from chip fabrication, server manufacture, data-centre construction and end-of-life disposal. It is shared across every task the hardware performs in its life.

Embodied carbon is a large share. Wu and colleagues at Meta reported in 2022 that it was roughly 30% of the footprint of Meta's large-scale machine learning. Luccioni, Viguier and Ligozat estimated BLOOM's training at about 24.7 tonnes CO₂e counting only the dynamic power used during training, and about 50.5 tonnes once everything from equipment manufacturing to operational energy was included. As electricity gets cleaner, the embodied share grows, which makes keeping hardware in use for longer, and keeping it busy, more important. See account for embodied carbon of AI hardware.

How much carbon does training an AI model emit?

Training emissions for large models published so far range from tens to tens of thousands of tonnes CO₂e. Model cards (the documentation released with a model) are the main place to find them, though many frontier models publish nothing.

Model Training emissions Basis Source
BLOOM (176 billion parameters, 2022) about 24.7 t CO₂e; 50.5 t including manufacturing and all operational energy Independent life-cycle estimate Luccioni, Viguier and Ligozat, 2022
GPT-3 (175 billion parameters, 2020) 552 t CO₂e; 1,287 MWh Estimate by Google researchers Patterson et al., 2021
Llama 3.1 family (8B, 70B, 405B, 2024) 11,390 t CO₂e location-based; 0 t market-based Model card, 39.3 million GPU-hours Meta, 2024
Llama 3.1 405B alone 8,930 t CO₂e location-based Model card Meta, 2024
Mistral Large 2 (training plus 18 months of use) 20,400 t CO₂e; 281,000 m³ of water Life-cycle assessment Mistral AI, 2025

These are one-off costs, but they are not the whole story. Patterson and colleagues cite industry estimates that 80–90% of machine learning workload is inference, so for a popular model, running it over its life can outweigh training it. The EU AI Act now requires providers of general-purpose AI models to document known or estimated energy use for regulators. See measure the carbon footprint of training a model.

How does the electricity grid change AI's emissions?

The same kilowatt-hour can carry very different carbon depending on where and when it is used. Carbon intensity is the CO₂e emitted per kilowatt-hour of electricity generated. A grid running mostly on hydro, nuclear, wind or solar has a low intensity; one running on coal or gas a high one.

  • Lawrence Berkeley National Laboratory estimated that US data centres' 2023 electricity was linked to about 74 million tonnes CO₂e, an average of 0.34 kg CO₂e per kWh, given the grids where they are located.
  • Patterson and colleagues found that choosing the model, data centre and processor can reduce the carbon footprint of training by around 100 to 1,000 times. Location is one of the biggest of those levers.

This is why where you run AI, and when, matters. A training job that can wait can be scheduled for hours when the grid is cleaner. See pick a low-carbon cloud region, schedule training when the grid is greener and find grid carbon intensity data.

Location-based or market-based: which number is right?

Both are legitimate, but they answer different questions, and the gap can be huge. Under the Greenhouse Gas Protocol, a company reports Scope 2 emissions from electricity in two ways:

  • Location-based uses the average carbon intensity of the grids it actually draws from.
  • Market-based reflects the clean energy contracts and certificates the company buys.

Meta reports Llama 3.1 training as 11,390 tonnes location-based and zero market-based. Microsoft's 2025 financial year shows Scope 2 emissions of about 12.0 million tonnes location-based against 2.7 million market-based. Alphabet reported 15.1 million tonnes location-based against 2.8 million market-based for 2025. When a company calls its AI "carbon neutral" or "100% renewable", check which method it is using. See check whether my AI runs on renewable energy and honest green claims.

Are tech companies' emissions rising because of AI?

Yes: the largest AI providers report rising total emissions, mainly from building and powering data centres. Their annual sustainability reports show the trend.

  • Microsoft reports total emissions of about 12.9 million tonnes CO₂e in its 2020 financial year, the baseline for its goal to be carbon negative by 2030, and about 20.3 million tonnes in 2025 (on its management-criteria basis). Its electricity use more than tripled over the same period, from about 10.8 TWh to 37.0 TWh.
  • Google reported that its operational emissions (Scope 1 and market-based Scope 2) fell 2% in 2025 while its electricity demand rose 37% and its supply-chain emissions rose 25%. Alphabet's assured figures for 2025 show about 15.9 million tonnes of Scope 3 (supply chain) emissions, more than five times its 2.9 million tonnes of operational emissions.

The pattern is the same in both: buying clean electricity is holding down operational emissions, but the manufacturing and construction behind new AI capacity is pushing supply-chain emissions up.

At national level, the evidence annex to the UK Compute Roadmap estimated that emissions from AI compute between 2025 and 2035 could be 34–123 million tonnes CO₂, around 0.9–3.4% of projected UK emissions over the period. The wide range reflects uncertainty about how much compute will be built and how clean the grid will be.

Why do AI carbon estimates disagree?

Because they count different things, use different grid figures and date quickly. The main sources of difference:

  1. Boundary. Chips only, or the whole server, cooling and idle capacity? Hardware manufacture? Training shared across uses?
  2. Accounting method. Location-based or market-based electricity, as above.
  3. Grid and time. The same model run in a different region, or a year later, can have a very different footprint.
  4. Workload. A short prompt, a long reasoning answer and an agent session differ a thousandfold in energy.

A figure is only meaningful with its boundary, method, location and date. See measurement is the bottleneck.

How can I estimate the carbon footprint of my own AI use?

Multiply energy by grid carbon intensity, then add a share of embodied emissions. This is the structure of the Software Carbon Intensity (SCI) standard, ISO/IEC 21031:2024, which expresses the result per unit of useful work, such as per request or per user.

  1. Estimate the energy of the AI work in kilowatt-hours: from a meter or tool if you run it yourself, or from per-task figures for a hosted service.
  2. Find the grid's carbon intensity where the work runs, in kg CO₂e per kWh, ideally for the hour it ran.
  3. Multiply the two for operational emissions.
  4. Add embodied emissions: the hardware's manufacturing footprint divided across its working life.
  5. State the boundary, the method (location- or market-based) and a range rather than one number.

Tools help. CodeCarbon and Green Algorithms estimate emissions for code you run; EcoLogits estimates the impact of calls to hosted AI models. They are estimators, not meters. Step-by-step guides: work out my personal AI carbon footprint, calculate an SCI score for an AI feature and find my cloud provider's carbon data.

What to do next

Frequently asked questions

What is the carbon footprint of ChatGPT?

OpenAI has not published a carbon figure per query. Its chief executive gave about 0.34 Wh of energy for an average query; the carbon depends on the grid supplying the data centre. Google reports 0.03 g CO₂e for a median Gemini text prompt, by its own market-based accounting.

How much CO₂ does training an AI model produce?

It ranges from tens to thousands of tonnes. Training BLOOM was estimated at about 24.7 tonnes CO₂e (50.5 including manufacturing and all operational energy), GPT-3 at 552 tonnes, and Meta reports 11,390 tonnes location-based for the three Llama 3.1 models.

Are AI emissions going up?

For the largest providers, yes. Microsoft's total emissions rose from about 12.9 million tonnes in its 2020 financial year to 20.3 million in 2025. Google reported that its operational emissions fell 2% in 2025, but its supply-chain emissions rose 25% as it built data centres.

What is embodied carbon in AI?

The emissions from making, shipping and disposing of chips, servers and data-centre buildings. Wu and colleagues at Meta estimated it at roughly 30% of the footprint of Meta's large-scale machine learning, and it grows as a share when the electricity is clean.

Does AI running on renewable energy have zero emissions?

Not in reality. Market-based accounting lets a company report low or zero electricity emissions if it buys renewable energy certificates, while the grid it draws on may still burn fossil fuel. Meta reports Llama 3.1 training as 11,390 tonnes location-based but zero market-based.

How do I calculate the carbon footprint of my AI use?

Multiply the energy used, in kilowatt-hours, by the carbon intensity of the grid supplying it, in kg CO₂e per kWh, then add a share of the hardware's embodied emissions. Tools such as CodeCarbon, Green Algorithms and EcoLogits help estimate each part.

Is AI bad for the climate?

It adds to emissions, mainly through rapid growth in data-centre electricity and construction. How much depends on how it is used, how efficiently it runs and how clean the grid is. AI can also help cut emissions elsewhere, but those benefits are not automatic and are harder to measure.

Sources

  1. Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
  2. Our contribution to a global environmental standard for AI, Mistral AI (2025)
  3. The gentle singularity, Sam Altman (OpenAI) (2025)
  4. Carbon emissions and large neural network training, Patterson et al. (arXiv 2104.10350) (2021)
  5. Llama 3.1 8B model card, Meta (2024)
  6. Estimating the carbon footprint of BLOOM, a 176B parameter language model, Luccioni, Viguier and Ligozat (arXiv 2211.02001) (2022)
  7. Sustainable AI: environmental implications, challenges and opportunities, Wu et al., MLSys (2022)
  8. 2026 environmental data fact sheet, Microsoft (2026)
  9. Read our 11th annual Environmental Report, Google (2026)
  10. Alphabet FY2025 environmental indicators assurance letter (Schedule of select environmental metrics), Alphabet (2026)
  11. 2024 United States Data Center Energy Usage Report, Lawrence Berkeley National Laboratory (2024)
  12. Scope 2 guidance, Greenhouse Gas Protocol (2015)
  13. Regulation (EU) 2024/1689 (AI Act), Annex XI, European Union (2024)
  14. UK Compute Roadmap: evidence annex, Department for Science, Innovation and Technology (2025)
  15. The real energy use of agentic AI, Zeke Hausfather, The Climate Brink (2026)
  16. ISO/IEC 21031:2024 Software Carbon Intensity (SCI) specification, ISO/IEC and Green Software Foundation (2024)
  17. CodeCarbon, CodeCarbon project (2020)
  18. Green Algorithms: quantifying the carbon footprint of computation, Lannelongue, Grealey and Inouye, Advanced Science (2021)
  19. EcoLogits, GenAI Impact (2025)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy