Kaiyuan Securities highlights that, in contrast to Large Language Models (LLMs), multimodal models substantially amplify the demand for AI computing power during both the training and inference stages. Based on Factorial Funds' estimations of Sora's computing power requirements, training Sora necessitates a minimum of 4,200 to 10,500 H100 GPUs, operating continuously for a month. In the inference phase, a single H100 GPU can produce roughly 5 minutes of video content per hour at its peak performance, with inference costs being several orders of magnitude greater than those associated with LLMs. As Sora2 undergoes capability enhancements and its user base expands, it is anticipated to further fuel robust growth in the demand for AI computing power. We maintain a resolutely positive outlook on the developmental trajectories of both the AI industry and the computing power sector in 2025.