UiPath migrated from isolated, on-demand GPU clusters to a shared fleet architecture on Google Cloud to support its agentic AI and intelligent document processing (IDP) workloads. The core challenge was managing spiky workloads, GPU supply bottlenecks, and operational overhead across regions. The solution centers on a machine learning services (MLS) platform that schedules and prioritizes work across teams, using A3 VM instances (NVIDIA H100) for training and G4 VM instances (NVIDIA RTX Pro 6000) for cost-efficient inference. Google Cloud's Dynamic Workload Scheduler (DWS) enables advance capacity booking, eliminating reactive provisioning. Key lessons include decoupling capacity from individual products, scheduling compute in advance, and right-sizing GPU instances per workload type. Customer outcomes include Omega Healthcare automating 100M+ transactions at 99.5% accuracy and Thermo Fisher processing 53% of invoices without human involvement.
