Illustration of a man pulling back a red curtain to reveal a data centre beside a river and power lines

Position Paper • AI & Sustainability

Stop Counting Prompts:
A Practical View of
AI and the Environment

By Seyha Nop • 12 November 2026

Illustration of a man pulling back a red curtain to reveal a data centre beside a river and power lines

Position Paper • AI & Sustainability

Stop Counting Prompts:
A Practical View of
AI and the Environment

By Seyha Nop • 12 November 2026

At a recent client workshop, someone asked what all this AI means for the environment. It is a fair question. The answers often go in two directions: every prompt drains a bottle of water, or the impact is so small that it barely matters. Neither gives companies much help in deciding what to do.

Our view is that counting prompts alone tells you very little. A short text exchange can use little electricity. At scale, training and running models requires substantial electricity and water, and more demanding uses can increase that footprint quickly. For a company, the useful question is how much computing its AI use requires, and whether the work justifies it.

The prompt is small

In 2025, Google published a detailed measurement of its Gemini assistant. It reported that the median text prompt used 0.24 watt-hours of electricity, about five drops of cooling water, and emitted 0.03 grams of CO2 equivalent (Google Cloud, MIT Technology Review). The International Energy Agency reached a similar conclusion this year: a simple text query typically uses less electricity than a television running for the same length of time. If every conventional internet search became a simple AI text query, the total would stay under 4 TWh a year, less than 1% of current data centre electricity use (IEA, Key Questions on Energy and AI, 2026).

To make that more concrete, imagine a hundred employees each sending fifty prompts every working day. Over 220 days, that is 1.1 million prompts. If each used 0.24 Wh, the total would be 264 kWh, around 8% of the 3,500 kWh annual household benchmark cited by Energuide. That is an illustration of scale, rather than an estimate of a company’s actual consumption.

There are limits to that comparison. Google reports a median for one service, which cannot tell us the average across other models or tasks. Its cooling-water figure excludes the water used to generate electricity. Training is outside the study’s scope, although its carbon calculation does include allocated hardware emissions and credits for clean electricity purchases. Google also reports a 33-fold drop in energy per median prompt over twelve months. Efficiency is improving, but that does not mean total consumption is falling.

The system is large

Data centres consumed about 485 TWh of electricity in 2025, up 17% in a year. Consumption at AI-focused data centres grew by 50%. The IEA expects the total to roughly double to 950 TWh by 2030, around 3% of the world’s electricity (IEA). This expansion is putting pressure on climate targets. As The Guardian reported, Microsoft, Amazon and Google’s combined emissions rose nearly 20% in a year, and Microsoft alone reported a 25% increase, according to ESG Dive. Google has also acknowledged, as eWeek notes, that its AI expansion is moving faster than the decarbonisation of the grid. UN University researchers estimate that the electricity used by data centres had a water footprint of 4.5 trillion litres in 2025, potentially rising to roughly twice that by 2030 (UNU-INWEH report).

Data centre electricity use is set to double by 2030

All data centres, in terawatt-hours per year. About a third is AI-focused today.

415 TWh

485 TWh

950 TWh

2024

measured

2025

measured

2030

projected

Chart: Merlin Innovation. Data: IEA, Energy and AI (2025) and Key Questions on Energy and AI (2026). The 2030 figure is the IEA central projection.

These figures describe different scales. A small footprint per task can still add up when demand grows. Data centres also support much more than AI: cloud services, streaming, storage and company IT. Our World in Data estimates that non-AI facilities accounted for about two-thirds of data centre electricity in 2025, and AI-focused facilities for about a third. That estimate does not tell us how the AI share divides between everyday chat, training, video, reasoning and agents.

About a third of data centre electricity goes to AI-focused facilities

Global data centre electricity use in 2025, estimated split

485 TWh in total

Non-AI data centres

about 323 TWh

Cloud services, streaming, storage, company IT

AI-focused data centres

about 162 TWh

Chat, training, video, agents: split unknown

For comparison, a hypothetical scenario on the same scale

Every internet search as a simple AI text query: under 4 TWh

The scenario is not part of the split above and is not a measure of today’s chat use.

Chart: Merlin Innovation. Data: IEA (2026) and Our World in Data. The under-4-TWh figure is a hypothetical search-replacement scenario, not measured text-chat consumption.

This is close to home. Data centres already take about 4% of Belgian electricity, twice the European average, and BCG expects that share to pass 10% by 2035. Elia has raised the idea of capping how much grid capacity data centres can claim, so that other industry keeps its place, Reuters reported. And according to The Brussels Times, the first phase of a new AI data centre in Neder-over-Heembeek is expected to use around 3% of the Brussels region’s electricity and up to 140,000 cubic metres of water a year.

The full life cycle

A per-prompt figure does not necessarily cover the full life cycle. Training, chip manufacturing and data centre construction also need to be considered, and studies draw their boundaries differently. One useful example is the assessment Mistral published in 2025 with Carbone 4 and ADEME. It reports about 20 kilotonnes of CO2 equivalent for Mistral Large 2 over the assessed period. In its breakdown, training and use account for roughly 85% of emissions, and hardware manufacturing for about 11% (Mistral AI). Microsoft, meanwhile, attributes much of its recent emissions increase to infrastructure construction (Techstrong). These findings can coexist: the source of a company’s annual emissions increase can differ from the largest share of a model’s life cycle footprint.

Training and running the model account for most emissions and water use

Share of one model’s life cycle footprint by stage (Mistral Large 2)

Stage

Emissions

Emissions

Water

Water

Materials

Materials

1 Model conception

1 Model conception

Training data and developers’ equipment

<1%

<1%

<1%

2 Data centre construction

2 Data centre construction

Putting up the buildings, before any computing

<1%

<1%

1.5%

3 Hardware

3 Hardware

Making, shipping and retiring servers

11%

5%

61%

4 Training and inference

4 Training and inference

Electricity and water to run data centres. Our focus here

85.5%

91%

29%

5 Network traffic

5 Network traffic

Requests and responses in transit

<1%

<1%

<1%

6 User devices

6 User devices

Laptops and phones, made and powered

3%

2%

7%

7 Enabled impacts

7 Enabled impacts

What the AI is used for. Not measured in the study

N/A

N/A

N/A

Bars show each stage’s share of the total.

Chart: Merlin Innovation. Framework and data: life cycle assessment of Mistral Large 2 (Mistral AI, Carbone 4 and ADEME, 2025). Columns do not sum to exactly 100% because of rounding.

This article focuses on training and running models. In the Mistral assessment, these stages account for most emissions and water use. Hardware dominates the materials footprint. The breakdown will vary across models and infrastructure, but companies can already act on how they use AI. Construction, hardware and the wider effects of the work AI enables also deserve attention.

Where your footprint will come from

A short text exchange is generally among the lighter ways to use AI. The IEA notes that some video generation, reasoning and agent tasks can use hundreds or thousands of times more energy than simple text generation. One independent benchmark found that problem-solving tasks with reasoning used about 25 times more energy per response, on average, than text conversation. That compares different tasks and output lengths; it is not a fixed multiplier for switching reasoning on (University of Michigan, ML.ENERGY).

In this benchmark, a short AI video used about 1,900 times the energy of a text answer

Energy per generation in watt-hours, open-source models

Text answer

0.047 Wh

Image

2.9 Wh

Short video

90 Wh

Chart: Merlin Innovation. Data: Hugging Face (Delavande et al., preprint, 2025). One open-source model per task, so read it as an order of magnitude.

As companies move from assistants to agents, this distinction becomes practical. Rephrasing an email can require little computing. An agent working through a task for an hour and calling a large model repeatedly can require much more. A company can therefore increase its AI footprint substantially without treating that increase as a separate decision. We expect this to become an important sustainability question as adoption grows.

What we recommend

We recommend four practical steps.

1. Match the tool to the task

Use the smallest model that does the job well. Use reasoning modes, video and agents when the task benefits from them. Cost can help identify demanding workloads, but it is only a rough signal: pricing, energy use and carbon emissions do not move together consistently. Where possible, compare the computing needed to complete the task to the required standard.

2. Ask your vendors where your workloads run

Where the computing happens affects its carbon footprint. The electricity figures cited here are about 41 grams of CO2 equivalent per kWh in France and 384 in the United States (Ember data, via Projet Celsius). These are country averages, rather than measurements of a particular workload. Ask vendors which regions they use, what environmental data they can provide, and whether you can choose a region with lower emissions. Local water stress matters too. The European Commission has also proposed an energy and water rating for data centres above 500 kW, with first labels expected in 2027 (European Commission).

3. Measure before agents scale

Start recording AI usage before it becomes difficult to trace. Track which models and tools are used, for which tasks, and how often agents run. Ask vendors for energy, emissions and water estimates, including what those estimates cover. Even an imperfect baseline is more useful than trying to reconstruct consumption after a dozen automations are in production.

4. Point AI at work that reduces your footprint

The IEA estimates that wider adoption of existing AI applications could avoid more emissions than data centres produce. It also warns that this adoption is not guaranteed (IEA). Companies can look for concrete uses such as route planning, energy management in buildings, reducing scrap in production and avoiding unnecessary journeys. The benefit needs to be measured against what would otherwise happen, including the footprint of running the AI and any extra activity it creates.

Our answer

The tension remains. AI adds electricity demand, and in places such as Belgium that can put more pressure on an already constrained grid. Useful work can justify that demand, but it does not remove the environmental cost. Companies still need to understand it and reduce unnecessary consumption.

So our answer in that workshop would be this: a simple text prompt can have a small footprint, while the infrastructure serving billions of tasks has a substantial one. Both matter. For companies, the most useful decisions concern what they run, how often, where, and at what scale. Use AI where it helps, choose tools that fit the task, and measure the footprint as usage grows.

Want to understand the footprint of your AI use?

Tell us where you are with AI. We’ll tell you what we’d do next.