Announcing AI Energy Score Ratings

With the increased ubiq­ui­ty of large arti­fi­cial intel­li­gence (AI) mod­els, the envi­ron­men­tal impact of this tech­nol­o­gy has become an urgent con­cern. Recent esti­mates cal­cu­late that AI is respon­si­ble for 100 ter­awatt-hours (TWh) of elec­tric­i­ty glob­al­ly in 2025, with worst-case fore­casts pre­dict­ing a surge to 1,370 TWh by 2035. In the Unit­ed States, data cen­ter elec­tric­i­ty con­sump­tion is expect­ed to rise from 4.4% of total elec­tric­i­ty use in 2023 to 6.7–12% by 2028, dri­ven large­ly by AI. To meet this soar­ing demand, the U.S. is adding 46 gigawatts of nat­ur­al gas capac­i­ty by 2030, equiv­a­lent to the entire elec­tric­i­ty sys­tem of Nor­way. This reliance on fos­sil fuels risks derail­ing the ener­gy tran­si­tion and glob­al cli­mate goals, as AI’s ener­gy con­sump­tion increas­ing­ly com­petes with efforts to decar­bonize the grid. Beyond elec­tric­i­ty use and emis­sions, AI’s growth also rais­es con­cerns about its impact on water con­sump­tion, air pol­lu­tion, elec­tron­ic waste, and crit­i­cal materials.

The grow­ing ener­gy demand stems from both the increas­ing adop­tion of ener­gy-inten­sive AI mod­els, par­tic­u­lar­ly large lan­guage mod­els (LLMs), and the wide­spread adop­tion of AI in user-fac­ing appli­ca­tions. How­ev­er, despite these con­cerns, there is no clear con­sen­sus on what con­sti­tutes “AI” and how to com­pre­hen­sive­ly account for its direct and indi­rect envi­ron­men­tal effects. Most exist­ing stud­ies pri­mar­i­ly exam­ine AI’s ener­gy con­sump­tion and its result­ing emis­sions, while some take a broad­er life-cycle approach, ana­lyz­ing key chal­lenges and oppor­tu­ni­ties, includ­ing AI’s indi­rect impact on oth­er indus­tries. How­ev­er, a crit­i­cal gap remains: the lack of stan­dard­ized meth­ods to assess the envi­ron­men­tal foot­print of indi­vid­ual AI sys­tems and AI’s col­lec­tive con­tri­bu­tion to the cli­mate cri­sis.