UiPath re-architected its infrastructure to handle the shift from simple task automation to agentic AI, which requires far more computational power. The company moved from isolated, product-centric GPU clusters to a shared fleet managed by its machine learning services (MLS) platform on Google Cloud.

Three challenges drove the change: spiky workloads that left expensive capacity idle, supply bottlenecks for high-end H100 GPUs, and operational overhead from maintaining dedicated clusters in multiple regions.

The core architectural decision was to treat GPUs as a shared strategic resource. The MLS platform prioritizes work across teams and time windows—serving real-time inference during the day and switching to batch training at night. This eliminates the need for per-instance elasticity and lets UiPath schedule capacity in advance.

On Google Cloud, UiPath uses AI Hypercomputer for a unified hardware-software stack. For training, it relies on A3 VM instances with 8-cluster H100s. For inference, it adopted G4 VM instances with NVIDIA RTX Pro 6000 as a cost-effective optimization, freeing up the high-end clusters for training. Google Cloud‘s Dynamic Workload Scheduler (DWS) lets UiPath book GPU capacity days in advance, solving the supply bottleneck.

The results include production deployments of intelligent document processing (IDP) models. Omega Healthcare automates over 100 million transactions with 99.5% accuracy and 40% less processing time. Thermo Fisher Scientific processes 53% of invoices without human involvement, cutting processing time by 70%.

UiPath reports that its biggest wins so far are increased availability and reliability. It expects additional cost improvements as it decommissions legacy GPU resources.

Key takeaways for engineering teams building similar platforms: decouple capacity from specific products, use scheduling tools to book compute in advance, and right-size instances—A3 for training, G4 for inference.