Sustainable AI: How Can AI Reduce its Environmental Footprint?

Calculating the Environmental Impact of AI: Methodological Challenges

Since the launch of ChatGPT in late 2022, the AI industry has grown exponentially. According to Statista, the market for artificial intelligence grew beyond $184 billion U.S. dollars in 2024 and is expected to continue with the market racing past $826 billion U.S. dollars in 2030.

AI market size worldwide from 2020 to 2030. Source: Statista

However, such an expansionist agenda shouldn’t be done without first assessing AI’s planetarian costs. The stakes are higher than ever, as the world is rapidly running out of time to advertise a full-blown climate catastrophe. 

Assessing the environmental footprint of AI is therefore mandatory. However, as we explained in our article on The Environmental Impact of Digital Technologies and Data, this is a challenging task. 

Despite the research on the topic growing, it’s still impossible to make a comprehensive assessment of the environmental footprint of AI. Among the main limitations and challenges are:

Lack of transparency and data gaps

AI companies tend to be reluctant to disclose information about their products, including environmentally-relevant information. This translates into important data gaps and blind spots that make the work of AI researchers. As outlined by Sasha Luccioni, Climate Lead at Hugging Face:

What counts as an environmental footprint?

Most studies focus on energy demand or greenhouse gas emissions (i.e. carbon footprint). However, it’s important to assess other non-energy impacts that add up to the digital environmental footprint too, such as water and mineral consumption.

Looking at the whole AI lifecycle

Most articles have studied the footprint of AI during training. However, as put by Luccioni et al in a 2023 paper, other areas of the AI life cycle should be considered, including material extraction, manufacturing, model deployment, and disposal.

The life cycle assessment approach. Source: Luccioni et al.

The Environmental Footprint of AI

Despite the methodological limitations to estimating the environmental footprint of AI, the reality is that AI comes with substantial costs in terms of natural resources. In the following subsection, we will only cover the carbon and water footprints of AI. However, as already mentioned, new research is needed to assess the impact of AI in other areas, such as mineral consumption or biodiversity loss.

The Carbon footprint of AI

To deliver its magic, AI requires considerable amounts of energy, which often translates into additional carbon emissions. 

The energy consumed by LLMs can be classified into two categories: 

• Operational energy. It’s the energy required for running these tools, either training or inference.
• Hardware manufacturing. It’s the energy required for creating the equipment of AI systems, such as GPUs (graphical processing units).

Most studies have focused on the carbon footprint of AI during the training phase, mainly because it is the phase in the AI’s lifecycle where more data is available, especially in the case of open-source models. 

For example, in a 2019 article, Strubell et al. estimated that training an LLM with 213 million parameters can emit 626,155 pounds of CO2, which is nearly the lifetime emissions of five cars, including fuel. 

However, the energy bill during inference (i.e., the process of using a model to carry out the task it was trained for) can be as much, or even higher than training LLMs. 

In a 2023 study, Luccioni et al. calculated that deploying a 176 Billion parameter BLOOM model over 18 days used an average of 40.32 kWh of energy per day (roughly the equivalent of 1,110 smartphone charges) and emitted approximately 19 kgs of CO2 equivalent per day. 

In a more recent article, Lucioni et al. have also demonstrated that the energy required during the deployment phase is greatly influenced by the task at hand. In particular, they found that image-based tasks are the most energy-consuming. 

Source: Luccioni et al.

Against this backdrop, it’s not surprising that the electricity consumption of data centers –where the computing of AI systems, both training and inference, normally take place–  could double by 2026, according to a recent study by the International Energy Agency.

The spike in the electricity bill due to the widespread adoption of AI is bad news for the net-zero ambitious tech firms. For example, in its latest sustainability report, Google admitted that its emissions grew 13% in 2023 over the year before and 48% since 2019, mainly driven by the increase in energy consumed by its data centers.  The same goes for Microsoft, which reported a 29.1% increase in their emissions since 2020 due mainly to the electricity required to run AI.

The water footprint of AI

An increasing number of researchers are turning to the worrying water consumption of AI. The lion’s share of the water footprint of AI takes place in data centers, which require billions of gallons of water to cool down the servers that compute the calculations to train and answer AI prompts. According to a 2021 study, the average data center uses as much water as three average-sized hospitals. 

The water footprint of data centers could grow dramatically in the coming years as new centers are built to catch up with the demands of AI. According to researchers from UC Riverside, data centers could consume 4.2 – 6.6 billion cubic meters of water in 2027. That is more than half of the total annual water withdrawal in the United Kingdom.

Holistically hidden as a business secret, tech companies are finally starting to release information about their water use, following social and regulatory pressure. For example, in its latest sustainability report, Google announced a 17% increase in its water footprint compared to 2025, accounting for a total of 6.1 billion gallons –roughly the water needed to irrigate 41 golf courses annually. 

Because state-of-the-art cooling systems can only work with clean water, the needs of data centers often collide with the interest of neighbors, especially when data centers are located in areas that are already suffering problems of water scarcity. 

Yet data centers are the primary source of water use, new research is required to shed light on the water footprint in other phases of AI’s lifecycle, particularly during hardware manufacturing, such as the semiconductors that form the GPU chips.  

3 Techniques for Reducing AI Environmental Footprint

Given the increasing environmental cost of developing and deploying AI models, companies must implement strategies and adapt techniques to reduce their footprint. 

In the following subsection, we analyze some of the available solutions to advance a more sustainable AI. 

1. Optimizing model efficiency

One of the most obvious strategies to reduce the carbon footprint of AI models is increasing their efficiency. In AI terms, that means achieving the same level of accuracy of the most advanced models with smaller, less complex models. 

On the one hand, efficiency could be achieved at the hardware level. For example, if engineers and microchip manufacturers can keep up with Moore’s Law –Gordon Moore predicted that the number of transistors on microchips will double roughly every two years–, in the coming years we will see faster, smaller and more efficient GPUs. 

Yet some forecasters predict that the law will end by around 2025 due to physical limits, other cutting-edge technologies, such as quantum computing, new materials, and innovative chip architectures, may be critical in the quest for efficiency. Equally, innovation in the design and operation of data centers is crucial to driving down energy consumption.

On the other hand, there’s software. AI researchers are advancing new techniques to designs to make AI models more efficient during training and inference, such as:

• Model pruning. It refers to the technique of removing unimportant parameters from a neural network to reduce the model size and enable more efficient model inference.
• Quantization. This comprises a set of different techniques to map continuous infinite values to a smaller set of discrete finite values. You can read our guide on LLM quantization to learn more. 
• Knowledge distillation. It’s a technique that aims to transfer the learnings of a large pre-trained model to a smaller model. You can learn more about LLM distillation in our separate article.

2. Utilizing renewable energy

Another solution to address the worrying carbon footprint of AI in data centers is investing in renewable energy. The obvious advantage of solar panels, wind turbines, or hydropower stations is that they create electricity without emitting carbon dioxide.

Big players in the AI industry, like Google, Microsoft, and Meta, are among the biggest investors in clean energy infrastructure. However, as Sasha Luccioni, Climate Lead at Hugging Face, has warned:

As a result, these companies have traditionally relied on other financial instruments to meet their clean energy commitments, such as renewable energy certificates (RECs) and carbon credits.

3. Leveraging cloud-based solutions

Through economies of scale, cloud providers offer AI companies a much more cost-effective solution to train and deploy their models. This is good news for the pockets of companies –for they avoid the up-front costs of buying their own computing infrastructure–, but also for climate change. 

Hyperscale data centers are specifically designed to be efficient. They can run energy-intensive routines faster and at scale, leading to considerable energy savings. At the same time, they are optimized to reduce the water used for cooling purposes.

You can learn more about cloud computing in our Understanding Cloud Computing Course.

Sustainable AI Practices in the Industry

If wisely used, AI can be one of our greatest allies in addressing the environmental crisis. Let’s analyze some compelling use cases where AI has been crucial to advance sustainable goals.

AI for environmental sustainability

The number of AI-powered applications to tackle environmental challenges is rapidly growing, ranging from initiatives to monitor carbon emissions and identify climate-vulnerable regions to improve environmental sustainability in supply chains and implement precision agriculture to limit water use and soil degradation. 

Below you can find a list of some applications worth checking:

Extreme weather modeling

Weather prediction is one of the oldest and most challenging–scientific tasks. AI is helping meteorologists make weather forecasts with unprecedented accuracy. This includes extreme weather event predictions. 

For example, GraphCast, a deep learning model created by Google, can predict the tracks of cyclones with great accuracy further into the future, identify atmospheric rivers associated with flood risk, and predict the onset of extreme temperatures.

Greenhouse gas emissions (GHG) monitoring

Most human economic activities release GHG emissions into the atmosphere. Knowing where, when, who, and how much emissions are emitted is crucial to better understanding climate change, advancing effective climate policies, and identifying emission hotspots. 

A great example of how AI helps in tracking GHG emissions is Climate TRACE, a coalition project from a non-profit coalition of organizations that leverage the power of AI and remote sensing to monitor emissions with unprecedented accuracy.

Optimizing power grids

Modernizing the power grid is a key undertaking to address the transition to clean energy. The rise of renewable energy is rapidly adding complexity to power grids. Compared to power plants that provide a constant flow of energy, solar panels and wind turbines generate variable electricity, for they are greatly influenced by the weather. 

AI can play a pivotal role in managing this complexity and balancing supply with demand. By analyzing vast amounts of data, AI can automatically predict how much electricity will be needed the next day and try to come up with the most cost-effective way to dispatch that energy.

Material discovery

Green technologies, like solar panels and wind turbines, rely on inorganic materials. Finding new materials is critical for technological progress, whether for developing greener technologies or for making the existing ones more efficient. 

AI is accelerating the material discovery process. For example, Google’s AI model GNoME has recently discovered 380,000 stable crystals that hold the potential to develop greener technologies, from better batteries for electric cars to superconductors for more efficient computing.

Corporate responsibility

Businesses from all sectors are also using AI to meet their sustainability objectives. Big Tech companies are among the most prominent examples, given their unbeatable knowledge of AI. For example, in its AI & Sustainability Playbook,  Microsoft identifies five areas where AI can be useful:

Source. Microsoft

DataCamp is also working towards a more sustainable future. We believe in supporting grassroots and global green initiatives with free or highly subsidized access to our world-class education platform. Through our DataCamp Donates program, we’ve worked with a range of organizations with climate change-focused missions to upskill their teams. 

Through our platform, we can support all kinds of organizations and professionals to make a positive impact on the environment. Some of our most notable successes are with NGOs such as CDP and Omenda, or clean energy companies, like Drax and SSE.

Ethical Considerations on the Development of AI

Every industry is compelled to assess its potential impact on the planet and find ways to lower its environmental footprint. The AI industry shouldn’t be an exception. Notwithstanding the promises of AI advance environmental objectives, additional efforts are imperative to reduce the environmental impact of the industry. 

Innovation is required to advance better and more efficient AI models. But this is not enough. Given the increasing environmental cost of AI, a serious and ethical debate about the development and growth of the AI industry is very much needed. 

AI trade-offs

First, we need to assess the implications of prioritizing AI development over other essential societal needs. While AI players are investing considerable resources in clean energy projects to power their data centers, this is energy that we are taking away from other potential uses, such as heating or providing electricity for low-resource families. 

Equally, it’s important to remember that green technologies are highly dependent on critical materials, including minerals, metals, and rare earth elements, that are difficult to find and take an enormous amount of energy to pull out of the earth, with negative implications for the planet. 

While the carbon footprint of AI has a global dimension, this is not the case for its water footprint. Here the trade-off is clear: the water used by a data center is water that is taken out from the neighbors and the biodiversity that lives in the surrounding area. 

Potential rebound effects 

Second, further research is required to assess the potential rebound effects associated with the development of AI. As observed in other sectors, there is the possibility that improvements in the efficiency of the AI models lead to an overall increase of AI use, which can eventually counteract such efficiency gains. 

To illustrate this, let’s take the case of ChatGPT. According to Goldman Sachs, a ChatGPT query needs nearly 10 times as much electricity to process as a Google search. For a tool that receives an average of 10 million queries daily, this translates into a huge energy bill. That’s why you have to pay a fee for most of the ChatGPT versions, and even in the most advanced versions, there is a daily prompt limit.  

Imagine now that OpenAI developed an innovative neural architecture that translates into considerable efficiency gains. Say that, with the new architecture, the latest version of ChatGPT requires only 5 times the electricity of a Google search. It’s very likely that OpenAI will then offer more affordable plans of ChatGPT with less prompt limits, leading to an increase of users. 

The outcome? OpenAI makes additional profits and users are happy, but emissions may be eventually on the rise, as ChatGPT will become even more popular and ubiquitous, thereby counteracting the efficiency gains that followed the new neural architecture.  

AI and the fossil fuel industry 

Another way of looking at the potential rebound effects is the close ties between some of the key AI players, like Microsoft, with fossil fuel companies. By providing expertise and access to AI technologies, Big Tech companies are helping oil and gas companies optimize and increase fossil fuel yield and production, leading to additional GHG emissions and potentially delaying the transition to renewable energies.  

These types of businesses are worrying and are somehow at odds with their global climate commitments. 

As Holly Alpine, former Microsoft’s senior program manager of Datacenter Community Environmental Sustainability, voiced in a recent article:

The Future of Sustainable AI

With all that we’ve covered so far, both the good and the challenging, related to AI, what will we see in the future as AI develops? 

Emerging trends in green AI

Research on the environmental costs of AI is rapidly gaining momentum. New evidence, together with increasing social concerns, will increase the pressure on AI players to reduce the footprint of AI models and advance a greener AI agenda.

Innovation is key to achieving efficiency gains. We are just at the beginning of the AI revolution, there is certainly wide room for improvement. However, as we explained in the previous section, the development of the industry cannot be done without considering the ethical implications of this technology. 

Policy and regulation

In the absence of ambitious, binding commitment by the AI industry, regulators and policy-makers have a say in advancing sustainable AI practice.

This is already happening in the European Union. The recently approved EU AI Act envisages the creation of voluntary industry standards for reducing the consumption of energy and 'other resources consumption’ of high-risk AI systems during its lifecycle and on energy-efficient development of general-purpose AI models, such as OpenAI’s GPT models. 

The Act also establishes that providers of general-purpose AI models should provide information on the known or estimated energy consumption of the model. 

Check out our EU AI Act Fundamentals Skill Track to know more about this influential regulation.

While the environmental requirements in the final version of the Act are less ambitious than the previous text proposed by the European Parliament, this is just the beginning and new regulatory initiatives should follow up to increase environmental transparency and incentivise the AI industry to conduct environmental impact assessments.

Conclusion

The AI revolution is happening alongside the most pressing challenge humanity has ever faced: the climate crisis. AI can be a critical drive to advance sustainable goals, but there is also a lot to do to reduce its worrying environmental impacts. 

At DataCamp, we believe it is important to be aware and mindful of the environmental impact of data and technology, and we actively work with organizations that are committed to environmental causes. Nonprofit organizations are welcome to apply to DataCamp Donates or schedule a demo with our Sales team to discover all the ways DataCamp can support their eco-conscious missions.

• The Environmental Impact of Digital Technologies and Data 
• How DataCamp Helps Drax Accelerate the Transition to Renewable Energy
• Data Storytelling Case Study: Green Businesses Course
• How SSE is Achieving a Net Zero World with Data Upskilling
• AI Fundamentals
• How to Learn AI From Scratch in 2024: A Complete Expert Guide
• AI Fundamentals Certification