The editor of Downcodes learned that researchers from the University of Cambridge and the Chinese Academy of Sciences published a paper in Nature magazine predicting that by 2030, generative AI may produce more than 1 billion iPhone-equivalent electronic waste every year. This research is not intended to limit the development of AI, but to assess its environmental impact in advance and explore sustainable solutions. Through different growth models, the research team predicts that the amount of e-waste may increase to 400,000 to 2.5 million tons in 2030, an increase of up to a thousand times. While the 2023 baseline data may be slightly off, it still reflects the impact of the generative AI wave on e-waste.
Picture source note: The picture is generated by AI, and the picture authorization service provider Midjourney
In the paper, the research team noted that while energy consumption has long been a focus of attention, the physical materials associated with this process and the waste stream of obsolete electronic devices have not received enough attention. Their study does not aim to accurately predict the number of AI servers and resulting e-waste, but rather to provide a first rough estimate to highlight the scale of future challenges and explore possible circular economy solutions.
The researchers used different growth scenario models, including low, medium, and high growth models, to analyze the required computing resources and their service life. The results show that from 2,600 tons of e-waste in 2023, the amount of waste may grow to between 400,000 and 2.5 million tons by 2030, an increase that may be as high as a thousand times.
It should be noted that the figure of 2,600 tons in 2023 may be slightly misleading, because a lot of computing infrastructure has been deployed in the past two years, and this has not yet been counted as waste. However, this data can indeed be used as a reference standard for changes in electronic waste before and after the wave of generative AI.
Researchers have proposed some possible ways to slow down the growth of e-waste, such as downgrading servers instead of discarding them when they reach the end of their useful life, or reusing their communications and power components. In addition, software and efficiency improvements can also extend the effective use time of a specific chip or GPU. The study mentioned that quickly updating to the latest chips may be beneficial, because if not upgraded in time, enterprises may need to purchase two lower-performance GPUs to complete the work of one high-end GPU, which will exacerbate the generation of electronic waste.
By taking these mitigation measures, researchers estimate that e-waste generation can be reduced by 16% to 86%. However, whether this reduction can be achieved depends more on whether these measures will be adopted and how well they are implemented. If every H100 chip can continue to be used in the university's low-cost inference servers, the pressure on electronic waste in the future will be greatly reduced; on the contrary, if only one-tenth of the chips are reused, the electronic waste problem will remain severe.
Highlight:
? It is estimated that by 2030, generative AI may produce more than 1 billion iPhone-equivalent electronic waste every year.
♻️ Researchers suggest reducing the generation of electronic waste through downcycling and reusing components.
? The generation of electronic waste can be reduced by 16% to 86%. The key lies in the adoption and implementation of measures.
This research sounds a wake-up call for us, calling on the industry and government to pay attention to the environmental challenges brought by generative AI, and actively explore and implement sustainable solutions to reduce electronic waste and protect the environment. The editor of Downcodes will continue to pay attention to the latest developments in this field.