Amazon AWS recently introduced a new service called Amazon EC2 Capacity Blocks, which allows users to reserve Nvidia H100 GPUs for future periods, similar to booking a hotel room, with each reservation lasting up to 14 days. This service effectively alleviates the demand pressure for high-performance GPUs in machine learning and deep learning projects. Users can flexibly allocate between 1 and 64 GPU instances based on their project plans and resource needs, with each instance containing 8 Nvidia H100 GPUs. During the reservation period, users only pay for the actual usage time of the GPU resources. The price of GPU instances will fluctuate dynamically based on supply and demand. The service is initially launched in the AWS US East region and will expand to more areas in the future. Industry insiders believe that this innovative on-demand GPU purchasing method can effectively improve the utilization efficiency of expensive GPU resources and reduce the computational costs of AI projects.