Electricity, water, emissions and electronic waste: researchers extend the notions of green AI and red AI to the entire IA-data-telecom pile and classify remedies into four levels.
Data centers could absorb 6.7 to 12% of U.S. electricity in 2028, up from 4.4 percent in 2023, and AI-related water demand could reach 4 to 6 billion cubic meters per year by 2027. Based on these orders of magnitude, one team believes that the debates on "green AI" and "red AI" remain too focused on computational efficiency and carbon, and do not shed light on lifecycle trade-offs.
An environmental, economic, social and technical grid
The authors propose generalized versions of these concepts, extended to big data, information and communication technologies and more broadly to electronics and digital technology. Their framework crosses four dimensions — environmental, economic, social and technical/operational — and monitors, based on indicators such as electricity, carbon, water and electronic waste, the way in which risks spread from one layer to another and between the service provided and its use.
Countermeasures are categorized into four levels: reforming metrics to make actors accountable, reviewing design principles and driving demand, adopting more sober LLM development practices (data, architecture, training, scaling), and accelerating the transition of energy infrastructure.
Read the full article on the original source
Go to the source →