Abstract

In recent years, the global surge in applications for Large Language Models (LLMs) and Generative AI (GenAI) has driven major cloud service providers to make substantial investments in graphic processing units (GPUs) to accelerate AI computations. With chip supply constraints expected to persist in the short to medium term, users are increasingly turning to GPU cloud services to support their AI applications. However, given the diversity of access technologies available for these services, users must conduct thorough evaluations to make informed decisions. This report provides an overview of GPU cloud services, examining the development of local GPU cloud access technologies—such as private cloud and consumption-based pricing models—traditional remote GPU cloud access technologies, including virtual machine and bare-metal-as-a-service (BMaaS) technologies, and emerging remote GPU cloud access technologies, such as container and serverless architectures. A comparative analysis of these six GPU cloud access technologies is also presented.