AWS has introduced an innovative solution for machine learning developers in need of reliable computing resources. The launch of Amazon EC2 Capacity Blocks for machine learning allows users to reserve GPU capacity specifically for training and deploying generative AI and large language models. This service functions like a hotel reservation system, enabling customers to specify their requirements, such as the number of GPU instances and the duration needed. This approach ensures that machine learning projects can proceed without interruption, no longer constrained by a shortage of computational power.
With this new offering, businesses can efficiently train or fine-tune their models, conduct experiments, or utilize the service on-demand for disaster recovery scenarios. Channy Yun, a principal developer advocate at AWS, emphasized, “You can use EC2 Capacity Blocks when you need capacity assurance to train or fine-tune machine learning models, run experiments, or plan for future surges in demand for machine learning applications.”
Currently, the service is accessible for Amazon EC2 P5 instances, which are powered by Nvidia H100 Tensor Core GPUs. Customers located in the AWS U.S. East Region can take advantage of this service. The pricing structure is dynamic and influenced by supply and demand, offering users the flexibility to purchase GPU instance blocks ranging from one to 14 days and, at times, up to eight weeks in advance.
This move into the GPU rental market reflects a broader trend, as various companies are seeking to leverage the increasing demand for high-performance computing solutions. Notably, NexGen Cloud is planning to launch an 'AI Supercloud' service, enabling developers to rent resources for model training. Additionally, Hugging Face has introduced a Training Cluster as a Service solution earlier this year, and the U.S. government has significantly reduced the rental price for its Perlmutter supercomputer.
In the competitive landscape of AI chip production, Nvidia continues to lead the charge. In the second quarter alone, Nvidia shipped approximately 900 tons of H100 GPUs, yet competitors such as AMD, IBM, and SambaNova are actively working to gain ground in this rapidly evolving market.