AI resource consumption

Energy per query

The figure depends directly on token length. In 2025, Epoch AI published the cleanest token-based model:

Source: Epoch AI, 2025. Calculation model: GPT-4o, ~100 billion active parameters, NVIDIA H100.

The 10x myth: origin and context

The widespread claim that an AI query uses 10 times as much energy as a Google search goes back to a single estimate by Alphabet chairman John Hennessy (Reuters interview, February 2023). No measurement, no study, a casual “likely”.

From there the chain ran on: de Vries quoted it in a Joule paper, the IEA adopted it, the UN Environment Programme too. Since then it has circulated as an established figure.

The second problem: the reference value of “0.3 Wh per Google search” comes from a Google blog post from 2009. Current estimates put a search at ~0.04 Wh. Both sides of the calculation were overestimated by a factor of 10.

Source: Engineering Prompts, 2024

Efficiency leap 2024 to 2025

Between May 2024 and May 2025, Google reduced the carbon emissions per Gemini prompt by a factor of ~23. The cause: algorithmic optimisation plus better hardware utilisation. Studies based on GPT-3/GPT-4 data from 2022/2023 cannot capture this leap.

Source: Elsworth et al. (Google), arXiv 2508.15734, August 2025. Direct measurement in the production Gemini infrastructure.

Three footprints that do not always move in the same direction

The UN report (UNU-INWEH, June 2026) introduces a central finding that is almost entirely absent from public debate: carbon, water and land use are three independent dimensions. A decision that lowers one of these footprints can raise another.

Practical consequence: “carbon-neutral” is not the same as “ecologically harmless”. A data centre running on nuclear power has a low carbon footprint but high water use. A solar-powered data centre in a water-scarce region can cause a land conflict despite very low carbon emissions.

Source: UNU-INWEH: Environmental Cost of AI's Energy Use: Carbon, Water and Land Footprints (June 2026). Authors: Aczel, Chamanara, Matin, Farsi, Marwala, Madani.

Server location as a decision criterion

Users can influence the energy mix underlying their AI use through their choice of provider and region. The major providers now allow European server regions to be selected.

• Anthropic (Claude): API access runs via AWS regions. EU regions (e.g. eu-central-1, Frankfurt) are available and run on an electricity mix with a significantly higher share of renewables than US regions such as us-east-1 (Virginia) or us-west-2 (Oregon).
• OpenAI (GPT): EU data processing is possible via Azure regions. European customers can configure data residency in the EU.
• European electricity mix: The average carbon intensity of the European electricity mix is well below that of the US, primarily due to high shares of wind, solar and nuclear power. Texas and Virginia, two of the most important US data-centre locations, have a higher share of gas.

Organisations that want to make their AI use transparent in their sustainability communication can document server location and energy mix as concrete criteria in their choice of provider.

Global scale

Water use: context

Water does not physically “disappear”. What matters: is it still available where it was taken from? Data centres deliberately evaporate water in cooling towers. For the region it is gone for good, which is especially problematic in drought areas.

Data centres draw 80 to 90 per cent on high-purity “blue water”: directly from rivers, lakes or the drinking-water network, the same water as in the local water supply. Avocados and grain, by contrast, live largely on rainwater (“green water”) within the natural cycle.

This means: the global water demand of AI is minimal compared to industrial agriculture (approx. 70 % of all freshwater withdrawals worldwide). The real problem is local and qualitative, not global and quantitative.

Sources: UNU-INWEH, June 2026; GI study, June 2025

Why the figures diverge so widely

The range in the press is not a measurement error but a methodological problem.

System boundaries: is only the active GPU chip measured (underestimates real consumption by a factor of 2.4), or the entire site including upstream electricity generation? Both approaches yield valid figures, but for different questions.

Outdated efficiency assumptions: Many studies use hardware nameplate values from 2022/2023 and ignore algorithmic optimisations of the past two years.

Lack of transparency: No provider publishes AI-specific environmental metrics. Amazon (AWS) does not publish aggregated water-consumption data at all.

Methodological basis: Frontiers in Communication, March 2025 (14 models, 7B–72B parameters, direct measurement on NVIDIA A100).

Practical consequences

Based on current research, concrete behaviours can be derived that reduce resource consumption:

• Prompt discipline. Shorter prompts save energy proportionally. Pleasantries towards AI systems have measurable, if small, costs.
• Use AI images and videos sparingly. A single high-resolution AI image uses as much energy as a full smartphone charge.
• Limit agentic loops. Autonomous agents without a stop condition multiply consumption exponentially. Build in clear limits and manual approval steps.
• Small model first. For classification, extraction and structured tasks, a smaller model needs about 30 times less energy than a large one.
• Opt out of training data. Anyone using the OpenAI API should actively object to their input data being used for training purposes.