AI industry set to become one of the biggest carbon emissions contributor, says IT sustainability expert

An article published in the MIT Sloan Management Review has revealed that the AI industry is likely to become one of the biggest contributors to global carbon emissions, as well as have other kinds of environmental costs such as generation of e-waste and excessive water use. AI’s carbon footprint is much higher than conventional computing systems: a single ChatGPT query can generate 100 times more carbon than a regular Google search. That footprint is also growing fast, with computing power required for AI training runs doubling every 3.4 months, on average.

Niklas Sundberg, the author, points out several areas where this new industry’s impact is being felt.

Energy consumption by the IT industry is already massive. Data centres and transmission networks make up 1% to 1.5% of global electricity use, and 0.6% of global carbon emissions. An average data centre consumes enough power annually to heat 50,000 homes. Large language models (LLMs) like ChatGPT require considerably greater computational power, both for initial training and for answering queries. For instance, ChatGPT’s training alone used around 1.3 gigawatt-hours of energy (equivalent to 120 average U.S. households’ yearly consumption) and generated 552 tonnes in carbon emissions (equivalent to the yearly emissions of 120 U.S. cars). GPT-4 is expected to be around 10 times larger, and there are several competitors entering the market.

Water consumption in data centres is another concern. Data centres require large amounts of water for cooling, and those requirements will increase with the greater storage and computational demands of complex AI models. Microsoft and Google reported 34% and 20% increase in water consumption respectively from 2021 to 2022, which has been linked with their development of AI technologies.

Apart from power consumption, the production and disposal of the hardware that AI systems run on also poses a problem. By 2050, the volume of electronic waste annually is expected to double to 120 million tons. A mere 20% of that gets recycled.

The solution, naturally, is not to abandon AI altogether but, rather, to follow certain best practices for sustainable AI, which Sundberg refers to as the three R’s – relocate, rightsize, and re-architect.

• Relocate addresses the fact that some locations have easier and cheaper access to renewable energy sources, and a strategic relocation can therefore reduce carbon emissions by 1.4x to 2x.
• Rightsize refers to using processors and systems designed specifically for machine learning training and AI workloads, rather than general-purpose servers. This can increase performance and efficiency by 2x to 5x.
• Re-architect means selecting an appropriate and effective machine learning model architecture, such as a sparse model, which can improve performance while decreasing the computational load by 3x to 10x.

These measures, along with effective data-management, compliance and increased awareness among employees, customers and the general public can help build a more sustainable AI ecosystem. With the issues of climate change and environmental degradation being addressed on many fronts, it is a good idea to start taking them into consideration within the AI industry while it is still growing, rather than having to implement changes retroactively.