Why do estimates of AI's impact differ?

Ask a simple question: how much CO2e does a written question to an AI emit? You will find 0.012 gram. You will also find 1.14 gram, more than 90 times as much. The first figure is a calculation: 0.3 watt-hour, Epoch AI's estimate for a question to ChatGPT, converted with French electricity. The second is published by Mistral AI, from a life cycle assessment carried out with Carbone 4 and ADEME. Between these 2 figures, Google gives 0.03 gram for a question to Gemini, and the same 0.3 watt-hour converted with the world average electricity mix gives 0.14 gram.

Climate

One question, 4 ways of counting CO2e

Emissions of a written question to an AI, in gCO2e, logarithmic scale · 2025

0.3 Wh, in Francecalculation, 41.4 g/kWh0.012g

Google, Geminiaccording to its contracts0.03g

0.3 Wh, worldcalculation, 458 g/kWh0.14g

Mistral AI, Le Chatlife cycle, half a page1.14g

0.010.1110

Source: Epoch AI; Ember; Google; Mistral AI · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

None of these figures is wrong, and none of their authors is cheating. They do not answer the same question. One counts the electricity of a country, the other that of a contract. One stops at the server, the other goes back to the factory that made the chips.

Going back over the company reports and the studies on which this hub is built, we counted 6 methodological choices. On their own, they explain ratios of 1 to several thousand between figures that are all accurate. Here they are in the order in which you come across them, each with the question to ask before citing a number.

A company can count the electricity of its data centres in 2 ways, and since 2015 the GHG Protocol, the international standard for emissions inventories, has required it to report the 2 results. The first applies the average carbon intensity of the grid to which the servers are connected. The second applies that of the contracts the company has signed: direct purchases from wind or solar farms, certificates of renewable origin. A company that covers all its consumption with contracts can report electricity emissions close to zero, whatever grid actually powers its machines.

Not all contracts are equal. Those that finance new installations genuinely reduce the emissions of an electricity system. Certificates bought from existing installations make no difference to the power stations already running. Microsoft stopped buying these so-called "non-additional" certificates, and its contract-based figure was multiplied by 10 in one year.

Climate

Electricity emissions, by contract or by grid

Emissions from purchased electricity (scope 2), 2 methods, in tCO2e, logarithmic scale · 2024 and 2025

Meta, contracts20241,358t

Meta, grid20245.97Mt

↑1 to 4,400the same scope 2 at Meta, depending on the method

Google, contracts20252.82Mt

Google, grid202515.1Mt

Microsoft, contractsfinancial year 20252.71Mt

Microsoft, gridfinancial year 202512Mt

1,00010,000100,0001 M10 M100 M

Source: Meta, 2025; Google, 2026; Microsoft, 2026 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

Same scope 2, 2 methods: the ratio ranges from 1:4 at Microsoft to 1:4,400 at Meta.

To know what data centres are doing to the climate today, the grid figure is the closest to physical reality. This hub makes that choice: its calculations use grid intensity. The first question to ask of an emissions figure is therefore: which electricity does it refer to?

Water has given the debate one of its most persistent images, that of a bottle of water per email. The 2 measurements that providers have published with their method come nowhere near it, and they are also very far apart.

A power station consumes water to produce the electricity for the servers, and a chip factory consumes water to make the chips. Google counts the water used in its data centres. Mistral AI counts the whole chain. Added to this is a confusion of terms:

• Withdrawing water means pumping it;
• it consuming, by contrast, means not returning it to the environment, because it has evaporated.
• Microsoft withdrew 13.3 million m³ in its financial year 2025 and consumed 8.2.withdrawn13.3 Mm³consumed8.2 Mm³
• And 1 m³ pumped in a water-stressed region does not count the same as 1 m³ pumped elsewhere. The water stress map →

Water

Water for one question, from 0.26 to 50 mL depending on scope

Water attributed to a written question to an AI, in millilitres, logarithmic scale · 2023 to 2025

Google, Geminion-site cooling, measured0.26mL

OpenAI, ChatGPTdeclaration, no methodology0.32mL

UC Riverside, GPT-3cooling and electricity10 to 50mL

Mistral AI, Le Chatfull life cycle45mL

↑a factor of 170with Google, by scope

0.1110100 mL

Source: Google; OpenAI; UC Riverside; Mistral AI · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

The second question comes in 2 parts: which water is being counted, and is it the water that is pumped or the water that does not come back?

A server is manufactured before it consumes its first watt-hour. Should that manufacturing be counted in the footprint of a question? Each organisation answers differently. Mistral AI, the GreenIT association and the EcoLogits tool count server manufacturing. Google counts the manufacture of chips and memory, about 30% of its 0.03 gCO2e per question. OpenAI and Epoch AI do not count it.

Depending on the indicator you look at, leaving out manufacturing therefore changes the result by a few percent or by more than half.

Hardware

Life cycle assessment of Mistral Large 2

Share of computation (training and use) in each indicator · Training and first 18 months of use, up to January 2025

Source: Mistral AI, Carbone 4 and ADEME, July 2025 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

Third question: is the hardware counted, and for which indicator?

Google publishes the median of its questions: half use less than 0.24 Wh, half more. This is honest information, but it cannot be used to calculate total consumption. The mean is necessarily higher, because long documents and reasoning mode pull the distribution upwards, and Google does not publish it.

Reasoning mode widens the gap further. On average, it multiplies energy use by up to 1.9 across the models tested by PEReN (the French government's centre of expertise for digital regulation), by 30 on average in Hugging Face's measurements, and by up to 700 on some models.

Uses

Reasoning mode multiplies energy use

Energy with reasoning, relative to the same request without, as a multiple · Measurements from 2025 and 2026

PEReN, average22 open models× 1.9

PEReN, codemaximum, one task× 9.5

Hugging Face, average2 model groups× 30

Hugging Face, same model3 models, with then without× 150 to 700

1101001,000

Source: Arcep and PEReN, 2026; Hugging Face, 2025 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

Fourth question: is the figure a median or a mean, and for what length of response?

The first 4 choices concern how to count a question or a company. The fifth concerns global projections. To go from a fleet of servers to electricity consumption, you have to assume how long they run, and at what load.

GreenIT assumes AI servers running continuously at 69% load. In France, RTE, the French transmission system operator, observes that data centres connected in the last 2 or 3 years draw only about 20 % of the power they have reserved: announced capacity overestimates consumption in the early years. The 2 figures do not measure the same thing: the first is an assumed server load rate, the second the share of their reserved power that recent sites actually draw. Both are a reminder that installed or announced capacity is not a measure of consumption.

This assumption, together with those on the size of the fleet, is what separates the projections for 2030. GreenIT projects 1,169 TWh for AI alone, more than the 945 TWh projected by the International Energy Agency for all data centres. The Shift Project goes up to 1,500 TWh for all data centres, cryptocurrencies included, by extrapolating capacity announcements. The IEA, for its part, applies a completion rate to announced projects and assumes efficiency gains. The comparison with Japan comes from its scenario.

Electricity

Data centre electricity consumption in 2030, according to 3 studies

Electricity consumption of data centres in 2030, in TWh per year, projection published by each study · 2023 to 2030

Source: IEA, 2026; The Shift Project, 2025; GreenIT, 2026 · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

Fifth question: what server utilisation rate does the projection assume?

The last choice is to publish nothing. No company publishes a measurement for an image, a video or a task given to an agent with its own models. The figures for images and video all come from open models measured by researchers.

Yet these uses are heavier than a short question: an image takes 1 to 3 Wh, the equivalent of 3 to 10 questions, and a 5-second video clip ranges from a few Wh to nearly 1 kWh depending on the model. Public debate therefore relies on the figures for short questions, the only ones providers publish. Our viewpoint on surface AI and deep AI starts from this observation.

Sixth question: has the use in question been measured by the company that sells it?

Taken together, these 6 choices explain why 2 accurate figures can differ by a factor of 90 for the CO2e of a question, 170 for its water and several thousand for a company's electricity. They also provide a framework for reading any figure, which we apply to every value in this hub.

For a company, the consequence is practical. The AI services it buys fall within its Scope 3, using its supplier's figures. If the supplier reports only its contract-based figure, the emissions of these uses are underestimated by a factor of 4 to several thousand. In a Bilan Carbone®, you need to ask for the 2 figures and state which one you use.

For everyone else, the rule can be stated in one line. A figure on AI given without its scope cannot be compared with any other. Anyone who cites it should be able to answer the 6 questions above, or say which ones remain unanswered.