《富国银行：能源行业研究：AI算力需求激增推动电力需求预测上修-资本开支与ASIC芯片应用双轮驱动-250922（英文版）（20页）.pdf》由会员分享，可在线阅读，更多相关《富国银行：能源行业研究：AI算力需求激增推动电力需求预测上修-资本开支与ASIC芯片应用双轮驱动-250922（英文版）（20页）.pdf（20页珍藏版）》请在三个皮匠报告上搜索。

1、Equity ResearchIndustry Update September 22,2025EnergyAI Power Surge:Capex&ASIC Adoption Push Our Power Demand Forecast HigherOur CallWere updating our AI power model to reflect higher hyperscaler capex post-Q2 EPS&rising ASIC adoption.We project 2030 AI power demand of 90 GW(vs 75 GW prior)&gas dem

2、and of+11 Bcf/d(vs+9 Bcf/d).Best positioned in our covg:BE,GEV,WMB.Key Changes To Our Power Model.Were revising our capex-based power model to reflect:(1)Forward year hyperscaler capex of$437B(from$358B last Q)following Q2 earnings,reflecting increased budgets across AMZN AWS,GOOG,MSFT,ORCL,&META.(2

3、)We now discretely model spend for ORCL&CRWV given their expanding roles in data center build-out.(3)Our power gen forecasts now incorporate the impact of rising ASIC adoption,starting with Google&Amazon.Raising Our 2030 AI Power Forecast To 90 GW.Were increasing our 2030 AI power demand estimate to

4、 90 GW,up from 75 GW previously.Roughly 10 GW of the uplift stems from higher hyperscaler capex assumptions,while 5 GW reflects the shift to ASICs.This represents our unconstrained growth view.Actual deployment will depend on whether AI firms can scale behind the meter(BTM)power gen,which is now gai

5、ning traction as grid&heavy duty gas turbine limitations become more apparent.ASICs Raise Our 2030 Power Forecast By 5 GW.ASICs(application specific integrated circuits)are custom chips built for specific tasks,unlike general purpose GPUs.Examples include Googles TPUs&AMZNs Tranium&Inferentia chips.

6、Hyperscalers deploying ASICs benefit from meaningful cost savings vs NVDA GPUs.We est a cost/watt of$40 for leading NVDA GPUs vs a cost/watt of only$10 for leading ASICs.With a fixed capex budget,a higher mix of ASICs therefore drives higher power demand.Raising Natural Gas Consumption Forecast To 1