Sustainability

Our climate footprint

Every time an AI colleague replies, a data centre uses electricity and water. Here we show how much we estimate our AI use has cost so far — in the simulations, in the tool and in building it — and how we calculate it.

Electricity: 375 kWh; Range 95 kWh–948 kWh
Emissions CO₂ equivalents: 139 kg; Range 24 kg–426 kg
Cooling water: 412 L; Range 19 L–1.7 m³

About the same as driving 2,080 km in an electric car, or emitting as much as 1,150 km in a new petrol car.

The figures are estimates with a range, not measurements: none of the AI providers we use state how much electricity their models need. We therefore show a medium estimate, with the low and high estimates beside it.

Calculated 5 Oct 2026, 16:29. The figures are updated every six hours.

The breakdown

Where the footprint comes from

We divide the AI use into three: what students meet in the sessions, the tool educators use, and the work of building it all.

The simulations

1.3 g CO₂eunder 1% of the emissions

20 Wh · 22 mL · 127 AI calls

The AI colleagues' replies in the sessions, speech in and out, and the notes after the session. The conversations go to Mistral AI on the EU endpoint.

The tool and the website

8.2 g CO₂eunder 1% of the emissions

22 Wh · 25 mL · 12 AI calls

The builder assistant, the help chat, the preview and the AI test. Here we mostly use Claude from Anthropic, which runs in the USA.

Development

139 kg CO₂eover 99% of the emissions

374 kWh · 412 L · 32,866 AI calls

When we build and fix the tool with the AI assistant Claude Code. This is the largest item, and it does not depend on how many students use the tool.

Counted from September 2026.

Most of the footprint comes from development, not from the students. Building a tool with AI uses far more than running it — and we would rather show that figure than hide it.

Before 2 October 2026 we stored usage only for the conversations in the sessions and for the AI test. The other usage by the tool before that date is missing from the figures.

One session

What does a session cost?

A worked example: 30 students in a written session of 45 minutes. We use typical figures for how much the students write and how long the replies are.

AI replies: 2,565
The whole session: 25 g; 705 Wh · 776 mL
Emissions per student: 0.82 g; CO₂e

Spoken session with speech in and out: 1.1 g per student.
The same written session on an average American electricity grid instead of the European one we assume for Mistral: 261 g — about 11 times as much. The ratio depends on which electricity figures you choose; see “Why the figures vary”.^1,2,4

Did you know

Did you know …

Some figures that have shaped our choices. Each point refers to the sources at the bottom of the page.

The electricity grid matters more than you think

Counting the whole life cycle of the power stations, electricity production in 2025 emitted about 28 grams of CO₂e per kWh in Norway, 41 in France, 210 on average in the EU and 384 in the USA. The same AI reply can therefore have many times the emissions, depending only on where the data centre is. Counting only the emissions from the power stations themselves, France was as low as 19.6 grams.^1,2

Student conversations stay in Europe

The AI colleagues reply using Mistral AI, a French company, which stores and processes data in the EU by default. The website itself runs at Upsun in OVHcloud's data centre in Gravelines in northern France, which Upsun puts at 58 grams of CO₂e per kWh.^5,6

One AI reply is small — many are a lot

Google measured that a typical text request to Gemini used 0.24 Wh in May 2025, and Epoch AI estimates about 0.3 Wh for an ordinary ChatGPT reply. That is like a 10-watt LED bulb for a minute and a half to two minutes. Anthropic, which makes Claude, does not publish such figures, so we have to use other companies' figures as a reference.^7,8

Training and the machines count too

Mistral AI had a life-cycle analysis made of its Large 2 language model together with Carbone 4 and the French environment agency ADEME. With training and the production of hardware included, a 400-token reply came to 1.14 grams of CO₂e and 45 millilitres of water. The training and the first 18 months in use together came to 20,400 tonnes of CO₂e and 281,000 cubic metres of water. We use the method as an upper bound under “How we calculate”.⁹

Water is harder to count than electricity

Data centres use water for cooling. Google states 0.26 millilitres per Gemini request — about five drops. Researchers at the University of California estimated that GPT-3 used half a litre for 10–50 replies when the water used in power generation is included. The difference shows how much it matters what is counted.^7,10

Images cost more than text

Researchers from Hugging Face and Carnegie Mellon measured open models and found that generating an image used on average more than 60 times as much electricity as generating a text reply. That is why we count the images in “Create your avatar” separately.¹¹

Data centres are growing fast

Data centres used about 415 TWh of electricity worldwide in 2024, around 1.5 per cent of all electricity use, and the International Energy Agency expects the figure to roughly double by 2030. In the USA, 4.4 per cent of all electricity went to data centres as early as 2023.^12,13

We do not call ourselves climate neutral

From 27 September 2026, the EU prohibits claiming that a service is climate neutral because its emissions are offset with purchased credits. Norway has adopted the same rule, but it has not yet entered into force. The Norwegian Consumer Authority already advises against such claims. We therefore state what we do, and how much — not that the footprint is gone.^14,15,16

The method

How we calculate

Four steps from each AI call to the figures at the top. The factors below are the same ones the calculation uses, so they can never say anything different from the figures.

Tokens
Each AI call stores how many tokens (pieces of text) went in and out, which model replied and which provider ran it.
Electricity
Tokens are multiplied by an estimate of electricity per 1,000 tokens, depending on the size of the model. Input tokens count for less than output tokens, and reused tokens for almost nothing.
Emissions
The electricity is multiplied by the emissions per kWh of the electricity grid where the provider runs the model — not by purchased guarantees of origin. If the provider does not state the country, we choose a range that covers the countries concerned.
Water
The electricity is multiplied by how much water data centres on average use for cooling per kWh. Water used in the power stations is not included.

Electricity per 1,000 output tokens^7,8

Model size	Low	Medium	High
SmallMistral Small, Claude Haiku	0.05 Wh	0.15 Wh	0.5 Wh
MediumMistral Large and Medium, Claude Sonnet	0.15 Wh	0.5 Wh	1.5 Wh
LargeClaude Opus and Fable	0.3 Wh	1.2 Wh	3 Wh

Emissions per kWh where the models run^2,3,4

Provider	Low	Medium	High
Mistral AIEU (Mistral does not state the country)	15 g	35 g	175 g
Speechmatics (speech in)EU	20 g	175 g	250 g
Anthropic (Claude)USA	250 g	370 g	450 g
ElevenLabs (speech out)Unknown, USA or EU	20 g	250 g	450 g
Google (images)USA and global	125 g	370 g	450 g

Why the figures vary^1,2,3,4,5,6

Emissions per kWh for the same country can vary a lot, depending on what is counted. We therefore show a range and explain how we have chosen.

Direct or life cycle: French electricity emitted 19.6 grams of CO₂e per kWh in 2025 when only the power stations themselves are counted. Including the construction of power plants, mining and fuel, the figure is 41 grams.
Annual average or hour by hour: Upsun states 58 grams for its data centre in France, based on the International Energy Agency and Electricity Maps. Such figures often give more weight to hours of high demand.
Unknown location: Mistral only states that data is processed in the EU, not in which country. We have therefore set the medium estimate at 35 grams, in line with Sweden and France on a life-cycle basis, and use the EU average (about 175 grams) as the high estimate.
Market-based or location-based: Some companies count the electricity they have bought guarantees of origin for, which gives lower figures. We use the grid where the data centre is located.
For the USA we use 370 grams, based on the direct emissions from electricity generation (EIA). On a life-cycle basis the figure is about 384 grams.

Cooling water: 1.1 litres per kWh (range 0.2–1.8).⁷

Upper ceiling by Mistral AI's life-cycle method, with training and hardware: 187 kg CO₂e. It is not added to the figures at the top, but shows how high the footprint can get when everything is counted.⁹

Not included

The training of the models (only in the life-cycle ceiling above).
The running of the website itself at Upsun and the storage at Scaleway.
The students' and educators' own machines and networks.
The water used where the electricity is produced.

Offsetting

How we offset the climate footprint

We have not offset the footprint yet. When we do, it will be stated here: what we have bought, from whom, for how many tonnes, and when.

How we think about it

We offset for the high estimate, not the medium one, so that we would rather offset too much than too little.
We prefer permanent removal of CO₂, such as biochar, to forest credits where the carbon can be released again.
We do not call the tool climate neutral. Offsetting does not remove the emissions; it is an addition to keeping them low.

What it would cost today

Our highest estimate so far is 426 kg CO₂e (the high electricity estimate or the life-cycle ceiling, whichever is larger). This is what it would cost to offset it:

Type	Price per tonne	For us now
Forest and nature credits (not permanent)	7–24 €	NOK 32–NOK 110
EU allowances that are cancelled	80–86 €	NOK 367–NOK 395
Nature-based removal	30–150 €	NOK 138–NOK 688
Biochar (permanent removal)	105–200 €	NOK 482–NOK 918

Prices read on 3 October 2026 from the providers' own websites and S&P Global (EU allowances).

Sources

The sources behind the figures

The type tells you how far you can trust the figure: research is peer reviewed, while companies' own figures have not been checked by anyone outside.

Analysis instituteEmber, via Our World in Data: Carbon intensity of electricity generation (life cycle, g CO₂e per kWh) (June 2026).
Public authorityRTE (French transmission system operator): Bilan électrique 2025 — emissions from electricity generation (February 2026).
Public authorityEuropean Environment Agency (EEA): Key trends and drivers in greenhouse gas emissions in the European Union (April 2026).
Public authorityU.S. Energy Information Administration (EIA): How much carbon dioxide is produced per kilowatthour of U.S. electricity generation? (December 2024).
The company's own figuresMistral AI: Where do you store my data or my organization’s data? (2026).
The company's own figuresUpsun: Regions — fr-3, OVHcloud Gravelines (April 2026).
The company's own figuresElsworth et al., Google: Measuring the environmental impact of delivering AI at Google scale (August 2025).
Analysis instituteJosh You, Epoch AI: How much energy does ChatGPT use? (February 2025).
The company's own figuresMistral AI, with Carbone 4 and ADEME: Our contribution to a global environmental standard for AI (July 2025).
Peer-reviewed researchLi, Yang, Islam and Ren: Making AI Less “Thirsty” (Communications of the ACM 68(7)) (June 2025).
Peer-reviewed researchLuccioni, Jernite and Strubell: Power Hungry Processing: Watts Driving the Cost of AI Deployment? (ACM FAccT 2024) (June 2024).
Analysis instituteInternational Energy Agency (IEA): Energy and AI (April 2025).
Public authorityShehabi et al., Lawrence Berkeley National Laboratory: 2024 United States Data Center Energy Usage Report (December 2024).
RegulationEuropean Parliament and Council: Directive (EU) 2024/825 on empowering consumers for the green transition (February 2024).
RegulationStorting (Norwegian parliament): Act of 19 June 2026 no. 34 amending the Marketing Control Act (not yet in force) (June 2026).
Public authorityNorwegian Consumer Authority: Avoid claims of “climate neutral” or “climate compensated” (joint Nordic position, in Norwegian) (May 2024).

Last updated 5 October 2026. The factors are checked at least every six months, and whenever a provider publishes its own figures.

Do you see an error, or have a better source? Write to kontakt@virtualaicorp.no.