KI: This is how much electronic waste is caused by new technology – how can it be prevented?

According to a study, if AI spreads quickly, 1,000 times more electronic waste could be generated in 2030 than in 2023. However, the amount of waste can be significantly reduced through various measures, wrote a group led by Peng Wang of Chinese Academy of Sciences in Xiamen in the journal “Computational natural sciencesThe numbers are based on typical calculations based on changing newer computer systems every three years.

Large language models are used for AI applications such as ChatGPT. “Large language models require significant computational resources for training, and require extensive computing hardware and infrastructure,” the authors write.

Studies on sustainability to date have mainly focused on the energy consumption and CO2 emissions of AI models. On the other hand, Wang and his colleagues wanted to know what amounts of electronic waste are generated when computationally intensive AI applications are used in more and more fields.

Scientists use the scenario as a basis on which they assume that large language models will also be adopted for everyday use, as can already be seen in some search engines and social platforms. With such a broad application of AI, data centers dedicated to training and deploying AI models must grow very quickly.

As a result, the amount of electronic waste generated by discarded servers and other devices could increase from about 2,550 tons in 2023 to up to 2.5 million tons in 2030. In scenarios with less use of AI, the amount of scrap that year could still be limited Ranging from 400 thousand to 1.5 million tons.

The researchers also calculated how much different measures could reduce the amount of scrap. So the most effective solution would not be to throw away servers and other hardware after three years, but rather use them for an additional year for simpler AI tasks or for completely different purposes. This would reduce the amount of waste by 62 percent compared to the basic scenario.

Desirable circular economy

If individual units of systems, such as processors and memory, were processed and reused, this could result in a savings of 42 percent. Additionally, improved algorithms deliver 50 percent savings, and more efficient chips deliver 16 percent savings.

Wang's team also points to the latest edition of the Global E-Waste Monitor. Accordingly, the total amount of scrap generated by smaller electronic devices – such as smartphones or personal computers – is expected to reach 43 million tons by 2030.

The total waste generated by AI servers and devices over the years, calculated by the authors, could reach five million tons by 2030 in the base scenario, roughly twelve percent of this amount. Based on the study's most conservative scenario, the cumulative e-waste generated by AI would represent a good three percent of the e-waste generated by smaller electronic devices.

Christiane Plosnick of the German Research Center for Artificial Intelligence (DFKI) in Kaiserslautern noted that there is only a small database of the assumptions made by the authors in the base scenario. But even the most conservative scenario with much lower amounts of waste and the forecasts of the Global E-Waste Observatory provided important reasons for creating a circular economy in IT.

“We need to create awareness in society that behind the cloud or AI application there are data centers with high resource consumption,” Plosinic emphasized. Reusing IT devices is better than recycling.