主题演讲 - 人工智能的肮脏秘密：解决人工智能硬件制造的隐藏环境成本.pdf

1、AIs dirty secretAddressing the hidden environmental cost of AI hardware manufacturingProvided by DigiconomistAlex de Vries-Gao in collaboration with Greenpeace East AsiaCurrent discussions have largely centered on the electricity demand of data centers.These facilities,which support the AI training

2、and inference processes,are placing increasing stress on electricity grids and sustainability due to their surging electricity capacity demand.AI is the main driver of data center expansion(IEA,2025),with AI already being responsible for 20%of global data center power demand at the end of 2024,while

3、 this share may increase to half of data center power demand by the end of this year(de Vries-Gao,2025).However,the assessment of AI energy consumption should extend beyond data centers to encompass the broader landscape of the AI industry,notably the energy-intensive process of upstream chip manufa

4、cturing.The Growing AI IndustryEnergy Demand from AI ChipmakingThe exact energy demand from AI chipmaking is not disclosed,but it is possible to make an assessment by considering the chip manufacturing process and the available capacity for packaging AI chips,which has been a key bottleneck for AI h

5、ardware manufacturing in recent years.Estimating wafer and energy demandUsing the specifications of AI chip components,with the consideration of production yields,market share and packaging capacity data,we estimated the wafer demand related to device production.Based on our estimated wafer demand o

6、f AI hardware,we calculated the electricity demand for the wafer production and found the energy needs increased significantly from 218 GWh in 2023 to 983.9 GWh in 2024,representing a 351%rise.Note:Over the past two years,Nvidias market dominance in the data-center GPU sector and its H100 alone occu