Deepseek Suggests Ai’s Power Needs May Be Less Lofty: BI Webinar Replay

Vertiv, GE Vernova, Schneider Electric, Siemens Energy and Constellation Energy may see lower than anticipated demand after China's frontier large language model (LLM), DeepSeek, proved training and inference can be done using a fraction of the GPUs compared to US LLMs. Yet history suggests that, as efficiency improves, total usage increases as new applications are found -- leading to an rise in overall power use. As engineers optimize algorithms, US investment, including Stargate, may continue. Also, tech companies may be reluctant to lower investment given their rivalry to put out ever more capable models, but there is a possibility that more floor space is allocated to IT equipment and the increase in overall power consumption is lower than prior estimates