Artificial Intelligence is taking the world by storm. Recently, we have witnessed what can be called a generative AI boom and mass adoption of large language model-based products like ChatGPT and DALL-E, from the second half of 2022. It has instilled both excitement in terms of immense possibilities and applications in diverse fields and fear that it may lead to a massive number of job losses, affecting a significant portion of the workforce around the world.
In such an environment, there is another critical question we should be asking: what is the impact of AI on the environment? Is it really sustainable in the long run? While there is some research on the potential of AI to contribute towards safeguarding our natural resources, figures from some big tech companies’ force us to rethink the whole scenario. For instance, Microsoft has reported a significant increase of 29% in carbon emissions since 2020, mostly due to its data centre expansions that were designed to support AI workloads. Another tech giant, Google, has also been facing challenges due to increased electricity demand driven by AI. It saw a rise of 48% in its carbon emissions since 2019, driven by data centres and supply chain emissions.
This blog expands upon these aforementioned premises and primarily focuses on the impact of Artificial Intelligence on the environment, in terms of electricity and water consumption, along with carbon dioxide emissions, before explaining how AI is affecting the regional power distribution in countries where data centres are operated and then charts some countries and regions that are coming up with more robust and complete regulations, to be the guiding star for AI applications.
This section of the blog examines the environmental impact of AI, with a focus on electricity and water consumption, along with carbon dioxide emissions.
AI models impact the environment in two key phases: the training phase and the inference phase. The latter accounts for 80% of the environmental footprint. While training is usually done multiple times to keep AI models updated and optimised, inference is done very many times to serve millions of users and specific scenarios and therefore, the energy cost of a single inference despite being small, results in a significant overall energy cost considering the number of queries a model like ChatGPT responds to per day. Google also reports that out of their total energy usage over the last three years, machine learning (ML, another term for AI) training and inference accounted for 10-15% of their energy use.
Building AI models demands significant water usage for data centre cooling to prevent overheating. The growing usage of AI has increased the water footprint of major tech companies. For example, water consumption at Google's data centres has increased by 17% since 2023, due to the expansion of AI products and services. Similarly, the water usage effectiveness (WUE) of Meta grew from 0.24 in 2017 to 0.30 in 2020, although it has come down to 0.20 in 2022.
Water consumption in data centres also depends on location. For example, running a GPT-3 inference uses 48 ML of water in Washington but only 7 ML in Ireland, since cooler climates reduce cooling needs.
Natural Language Processing (NLP) is a rapidly growing AI application, (according to a recent Transforma Insights study, 27% of AI-related acquisitions between 2021 and 2023 were focused on companies supporting NLP. Large language models (LLMs) like ChatGPT, Cohere, and Claude are particularly compute-intensive and training these models has significant environmental impacts. For instance, training GPT-3 with 175 billion parameters emits 502 metric tonnes of CO2, comparable to the annual emissions of 112 gasoline-powered cars. Content generation tasks, such as image creation or text summarisation, are more carbon-intensive than tasks like movie ranking, with AI-generated images having a carbon footprint similar to charging a smartphone.
Moreover, a University of Massachusetts study found that training one AI model with the US energy mix emits over 626,000lbs (284,000kg) of CO2. This is equivalent to five times the emissions of an average American car over its lifetime, forty times the annual emissions of a typical US household, or fifty-seven times the yearly emissions of an individual.
It is well-known now that artificial intelligence (AI) relies heavily on data storage and processing, so the expansion of AI has significantly increased the demand for data centre capacity and the energy demand is expected to soar further. Currently, AI applications account for 10-20% of data centre electricity consumption, but this percentage is expected to grow rapidly in the coming years. According to the International Energy Agency (IEA), electricity demand from data centres will increase from 460 TWh in 2022 to 800 TWh by 2026.
The geographic distribution of data centres is uneven across countries, significantly affecting national energy consumption, particularly in smaller nations. For example, although data centres account for 1-2% of global electricity consumption, they represent 18% of Ireland’s total electricity use. In Denmark, data centre electricity consumption is expected to increase sixfold by 2030, potentially comprising 15% of the country’s overall electricity usage.
Currently, while some data centre regulations exist, governments around the world are gradually coming up with more well-defined and robust regulations for the development of more sustainable AI solutions. This section of the blog addresses some of the emerging regulations from around the world.
Under Article 12 of the revised Energy Efficiency Directive, EU data centres with a capacity of over 500 kW must disclose their energy performance and sustainability metrics. Operators were required to submit key performance indicators to a European database by 15 September 2024, and subsequently by 15 May each year.
The government has launched the Special Action Plan for Green and Low-carbon Development of Data Centers to ensure sustainability and improve energy efficiency in data centres. It aims to optimise the layout of data centres and reduce the average PUE to 1.5 or lower.
Starting January 1, 2024, the Netherlands has imposed temporary restrictions on the construction of new data centres. Hyperscale data centres will be prohibited nationwide, with exceptions at select locations.
Germany's Energy Efficiency Act, passed in October 2023, mandates that data centres operational before 1 July 2026, should achieve a Power Usage Effectiveness (PUE) of 1.5 by July 2027 and 1.3 by July 2030. Data centres commencing operations after 1 July 2026, must meet a PUE of 1.2. The Act also requires these facilities to use 50% renewable electricity starting in 2024, rising to 100% by 2027.
Singapore's Infocomm Media Development Authority (IMDA) has introduced new standards under its Digital Sustainability Blueprint to gradually raise data centre operating temperatures. This change could save 2-5% in cooling costs for each degree increase in temperature.
Although the fast-paced advances of AI are crucial for possible positive outcomes, this should not continue without first addressing the issue of increasing energy needs for the development of AI applications. In this context, it should be added that energy-efficient GPUs and efficient data centre cooling infrastructure must play a part in stabilising the growing electricity demand of AI.
Additionally, it should also be noted that while currently there are no specific laws governing AI energy consumption, several countries and regions, including the EU, China, Germany, the Netherlands, and Singapore, have regulations on data centre power usage that indirectly support AI sustainability. To enhance transparency and accountability in AI’s environmental impact, regulatory measures are likely necessary.
A more in-depth discussion on this topic is provided in the Transforma Insights report, Growth in Artificial Intelligence forces vendors to refocus on sustainability.
