The evidence

Impacts

AI uses electricity and water, contributes to carbon emissions and depends on material infrastructure. The figures below remain attached to the systems and boundaries they describe.

01

Energy

Electricity demand from data centres is large and growing. AI is a major driver, but the best macro figures do not isolate it from other digital services.

415TWh in 2024
≈945TWh in 2030

Both figures describe global data-centre electricity, not AI alone. The first is an estimate and the second is the IEA base-case projection. Per-query estimates also vary greatly with model, task and assumptions.

Sources: IEA, Jegham and co-authors

02

Water

Water can mean direct cooling, electricity-generation water, semiconductor manufacturing or a full lifecycle result. Location also changes the significance of every litre.

Google Gemini0.26 mLmedian prompt, direct freshwater
Mistral Le Chat45 mL400-token response, lifecycle water

The raw values differ by about 173 times. This is not a provider ranking. They use different functional units and boundaries, so the distance between them partly reflects what each source counts.

Sources: Google, Mistral AI

03

Carbon emissions

Carbon results depend on electricity use, grid conditions, hardware and accounting choices. A precise figure can still describe only a narrow part of the system.

Google Gemini0.03 gCO₂emedian prompt, market-based electricity
Mistral Le Chat1.14 gCO₂e400-token response, location-based lifecycle method

The raw values differ by 38 times. The prompts, models, lifecycle boundaries and electricity accounting are not aligned. Dividing the two values does not establish that one provider is 38 times more efficient.

Sources: Google, Mistral AI

04

Materials

AI hardware depends on mining, refining, semiconductor manufacturing, construction and end-of-life systems. These impacts are often outside operational measurements.

No reliable AI-only material total

Mistral reports 0.16 mg Sb eq of abiotic resource depletion for one Le Chat response. This is a lifecycle impact indicator, not the physical mass of minerals used. ADEME and ARCEP quantify material inputs for the wider digital sector, not for AI alone.

Sources: Mistral AI, ADEME and ARCEP