GPU 있어도 에너지가 없다... SemiAnalysis AI 인프라 전격 진단 | AI 데이터센터가 발전소를 짓는 이유 | 전력 병목의 모든 것

Situations are unfolding where data centers cannot open simply because they cannot set a date for power to be supplied, despite having secured GPUs and HBM and even constructed the buildings. The standard for competition in AI infrastructure is shifting from "how many chips were purchased" to "when power is plugged into those chips." Based on analysis by SemiAnalysis, this video examines the limitations facing the U.S. power grid and the BTM (Bottom-to-Measure) structure, where data centers operate directly alongside power plants. We have summarized where the real bottlenecks lie in the process of generating, transmitting, and turning on electricity, as well as which companies benefit, ranging from gas turbines to power equipment and completion capabilities. We also cover the trend of connecting multiple small generators instead of one large power plant, delivery issues for power equipment like transformers, and the stories of Korean power equipment companies that are precisely positioned at these bottlenecks. We also highlighted the transformation of data centers from mere customers receiving electricity into small power grids that regulate power themselves. We examine how the roles of utilities and private power generators are changing, and why energy companies like Chevron are entering this market. Finally, we analyze how all this power discussion ultimately returns to the demand for GPUs and HBM. We trace the reasons why power bottlenecks do not block semiconductor demand, but rather open up alternative routes. Written by Patch Edited by Lee Jae-hyun [email protected]