Generative AI's Environmental Impact: What Our Experts Learned

A few years into the generative AI revolution, it's no secret that this technology requires significant computing power. Indeed, much like humans need more energy to write than to read, generating content consumes far more resources than the traditional machine learning tasks of the past (analysis, pattern-finding, etc.) did.  And to power GenAI, additional data centres are being built across the globe, sometimes prompting worries of potential strains on local energy grids or water supplies.

While the environmental footprint of GenAI stems from both machine manufacturing and electricity consumption, the latter accounts for over 75% of overall GenAI emissions according to a study by The Shift Project. As a result, more and more organizations are wondering how to embrace the AI revolution without sacrificing their CSR goals.

Why Generative AI Consumes So Much Energy

According to fifty-five's research, over a year of use, a single model family has an estimated carbon impact of 240,000 tCO₂e, comparable to Belize's annual emissions. Considering the number of models currently on the market, the additional carbon cost of their training, and the number of future models under development (more than 10,000 models are published on Hugging Face every month), the global carbon cost of Generative AI becomes staggering. At the heart of this exponential consumption: data centres, their infrastructure, and their GPUs.

While data centres are needed for many digital services, they are now mostly known for their essential role in the development of Generative AI. They are required to train GenAI models by feeding them hundreds of terabytes of data, a process with a high energy cost in itself, and for inference as well. Every response generated by a model activates powerful processors, which also require high-level cooling, further increasing data centres' energy consumption. And beyond energy use, data centres also require large quantities of water, rare metals for components, and land to be built on.

The Electricity Consumption of Data Centres

Data centres accounted for 2% of France's electricity consumption in 2023, and projections from the Shift Project suggest that this percentage will climb to 7.5% by 2035. For the UK, the House of Commons reports a 2.5% of energy consumption for data centres, with a projected fourfold increase by 2030, while in Ireland, they are reportedly already consuming more electricity than all urban homes combined – a fifth of the country's energy demand in 2023, a proportion still rising as demand for data processing power increases. To this day, however, the majority of data centres are located in the US. In 2025, they accounted for approximately 4.4% of the country's annual electricity consumption, with projections indicating this could reach 12% by 2028.

While this growing consumption is less carbon-intensive in green energy-reliant countries, additional pressure on power grids from new data centres could force the use of more carbon-intensive electricity to avoid shortages. Furthermore, the electricity consumed by data centers cannot be allotted to processes essential to the green transition, such as recharging EVs or electric heating.

Why Measuring the Carbon Impact of AI Remains So Difficult

As previously mentioned, the environmental impact of GenAI is multifactorial (carbon emissions, water use, metals, land use, etc.), making it more challenging to measure. In our research so far, fifty-five has focused on the carbon footprint of GenAI. While this analysis is far from simple, it is where methodology is the most advanced at this stage.

First, it is essential to take into account that both training and inference contribute to each model's footprint. The carbon cost of Generative AI inference also varies greatly based on machine type, the energy mix of data centres' location, the model's size, the number of generations involved, the type of task, and the size of the output. A text-based output will require significantly less energy to produce than an image or a video.

These factors would be challenging on their own, but most AI providers also share minimal data and use inconsistent reporting methods, creating significant transparency gaps. To address these gaps, fifty-five, the Brandtech Group and Scope3 released an exhaustive study estimating the carbon cost of GenAI as well as an open-source carbon calculator to help calculate the footprint of generative AI tasks.

How to Reduce the Environmental Cost of Generative AI

To reduce the environmental cost of GenAI, fifty-five identified two high-potential levers:

1. Technical design choices

Choosing the right model for the right task (e.g., translation vs. video generation), and selecting low-carbon locations for computation, such as Canada, France, Scandinavia, etc.

2. Usage optimisation

Reducing the number of generations from a model, notably by avoiding unnecessary prompts, and optimising the size and format of outputs. This optimisation can be achieved through specialised training courses, which our teams have led for several of our clients.

With thoughtful design and better usage patterns, our research shows it is possible to divide GenAI's carbon impact by four. But the foremost question should not be how to use AI, but why. Per a striking report by MIT, only 5% of AI POCs actually generate value. And to this day, many brainstorming workshops aim to "find AI use cases" rather than centering on business needs first.

What Companies Are Doing Today

Very few organisations currently measure the environmental impact of AI. Not only is accurate measurement complex for this technology, but the topic is often avoided out of fear of slowing AI adoption, which many believe is strategic for growth. Still few and far between, existing GenAI initiatives tend to focus on charters, training, and governance but are rarely connected to operational efficiency, even though reducing GenAI's footprint would also save money, energy, and time.

At fifty-five, we recommend first bringing business needs back to the centre. Start with a problem to solve, then identify the best technical response, keeping in mind that it might not be AI, generative or not. And if AI is indeed required, deploy it efficiently and support teams with governance and training.

Towards More Purpose-Driven, Responsible AI

CSR objectives and, more broadly, global climate goals should not be set aside in the name of technology. For a deeper dive into this topic and our complete list of best practices, our full study on the carbon footprint of AI is available for free, as is our carbon footprint calculator. And if you'd like to see how your organization can achieve a more responsible transition towards AI, don’t hesitate to contact us!