E-waste not a major challenge with generativeAI

Wang, Chen, Zhang, Tzachor (2024) suggest that the demand for generative AI will create an e-waste problem of 1.2 to 5.0 million tons  for 2020 to 2030. They also suggest this could be made worse by political restriction on access to more advanced  efficient chips and rapid replacement of old hardware. I suggest the problem is not that large and energy use will remain a larger problem than e-waste.

One aspect the authors do not mention is the lack of price signals between server providers and the end user with current generative AI services. This is likely to be self correcting. Currently demand for generative AI is being generated by offering of free services to the public. As the user is not paying for the service and there is therefore no built in fee for responsible disposal of created e-waste, there may be a later problem. Some speculative AI ventures are likely to become bankrupt leaving a toxic legacy (similar to the mountains of scraped e-bikes left by failed startups). However, as users come to rely on Generative AI services, vendors will introduce charges, which can cover e-waste costs.

Currently generative AI server farms are using generic Graphic Processor Unit chips. These are the same chips used for cryptocurrency server farms. The environmental issues are similar with both. There have been articles about AI consuming as much power as small countries (just as there were for crypto). However, there is more of a mainstream use for AI, which will allow for better long term regulation of environmental effects. With its abundance of renewable energy sources and a stable regulatory environment, Australia could provide a popular location for AI centers. This would allow a small query to be sent across the world and answer sent back, effectively embedding the renewable energy in the answer. 

There may also be scope for reuse of older, slower, more energy using AI chips in locations with abundant renewable energy. As more efficient chips were installed close to the user in high energy cost countries, the old hips would be installed further away. Rather than store energy in batteries to run these chips, it may be cheaper to shut them down when the sun isn't shining & the wind isn't blowing. How to do this is something engineers and computer professionals can learn to optimize with specialist training (Worthington, 2012). 

Reference

Wang, P., Zhang, LY., Tzachor, A. et al. E-waste challenges of generative artificial intelligence. Nat Comput Sci (2024). https://doi.org/10.1038/s43588-024-00712-6

Worthington, T. (2012, July). A Green computing professional education course online: Designing and delivering a course in ICT sustainability using Internet and eBooks. In 2012 7th International Conference on Computer Science & Education (ICCSE) (pp. 263-266). IEEE. https://doi.org/10.1109/ICCSE.2012.6295070