Within the CAIPM framework, infrastructure strategy for AI workloads must balance performance, cost efficiency, scalability, and flexibility. For workloads such as large-scale model training that are intermittent but computationally intensive, organizations benefit from on-demand access to high-performance compute rather than investing in permanent infrastructure.

The scenario clearly highlights key constraints: training workloads are short-lived but require powerful accelerators, and owning such hardware would result in underutilization and long procurement cycles. Cloud-based GPU resources directly address these challenges by offering scalable, on-demand access to high-performance accelerators without capital expenditure or long-term commitment. This enables organizations to provision resources quickly when needed and release them afterward, optimizing both cost and operational agility.

Option A, hybrid infrastructure, may still involve ownership and does not fully eliminate underutilization concerns. Option B, spot or preemptible instances, can reduce cost but introduce reliability risks, making them less suitable for critical training jobs requiring stability. Option D contradicts the requirement to avoid long-term hardware ownership.

CAIPM emphasizes leveraging cloud-native capabilities for elastic scaling and efficient resource utilization in AI programs. Therefore, cloud-based GPU resources are the most appropriate solution for flexible, high-performance compute access.