UiPath expands Google Cloud use for shared GPU fleet

UiPath has expanded its use of Google Cloud to run all training workloads and most intelligent document processing model inference. The move centres on a shared GPU fleet for the company's AI systems.

It rebuilt its infrastructure to shift from isolated GPU clusters to a common pool managed across training and inference tasks. UiPath uses Google Cloud A3 virtual machines with Nvidia H100 chips for model training and fine-tuning, while G4 virtual machines with Nvidia RTX Pro 6000 chips handle lighter inference work.

The change comes as UiPath pushes further into agentic AI, in which software agents are designed to reason, make decisions and carry out business processes across multiple systems. That work has increased demand for computing resources, particularly for intelligent document processing, computer vision and large language model-based reasoning.

UiPath has long run its full-stack automation platform on Google Cloud, but its earlier method of provisioning GPU nodes on demand became harder to sustain as workloads grew. Teams could scale resources up and down as demand changed, but the model left idle capacity during quieter periods and made it harder to secure enough top-end chips for large training runs.

Supply constraints also played a part. Demand for A3 instances with eight H100 GPUs made it difficult to expand training simply by adding more nodes, while serving customers in different regions required dedicated inference clusters that increased operational overhead.

Shared fleet

In response, UiPath reorganised GPUs as a shared resource managed by its machine learning services platform. Under that model, the fleet handles real-time inference and latency-sensitive workloads during busier periods, then switches to batch training and longer-running jobs when demand eases.

This gives engineering teams a way to plan capacity in advance rather than rely on per-instance elasticity. It also lets researchers and product teams draw on the same pool of hardware without tying specific machines to particular products.

Google Cloud's Dynamic Workload Scheduler is part of that setup. The scheduling tool lets UiPath reserve GPU capacity ahead of time for short bursts or larger training runs, improving predictability as demand for advanced chips remains tight.

"Realizing the full potential of enterprise agentic AI requires an infrastructure that matches our ambition. Google Cloud provides the scale and flexibility we need to train specialized models and deploy them globally. This partnership allows us to deliver high-precision intelligent document processing and autonomous agents that don't just chat, but actively drive business outcomes for our customers," said Raghu Malpani, Chief Technology Officer, UiPath.

UiPath's engineering teams are using large language model grounding to help software robots interpret interfaces, alongside document models built on the Qwen architecture to extract data from unstructured paperwork. Those workloads run on UiPath's cloud infrastructure, where latency is a key concern for production systems.

Production focus

The revised setup allows UiPath to schedule large training jobs without interrupting production inference. That separation is important for systems that must support live customer workloads while still giving researchers access to compute for model updates and experiments.

UiPath pointed to customer examples to show the output of its intelligent document processing work. Omega Healthcare uses UiPath to automate more than 100 million transactions with 99.5% accuracy, a 40% reduction in processing time and 15,000 fewer hours of repetitive tasks each month. Thermo Fisher Scientific uses the software to extract data from documents such as invoices and purchase orders, and now processes 53% of its invoices without human involvement while cutting processing time by 70%.

Those figures show why infrastructure decisions are becoming more important for software groups building AI products. For providers such as UiPath, the challenge is no longer only access to GPUs, but how to use scarce and expensive hardware across workloads that vary by region, time of day and business priority.

UiPath said its most immediate gains from the new arrangement have been higher availability and improved reliability. It added that further cost improvements are likely as it continues to retire older GPU resources and completes the transition to the new operating model.

"The shift to a shared fleet on Google Cloud transformed our operational model. We moved from reactive provisioning to a predictable, high-performance engine that powers our most advanced IDP and agentic AI workloads. With tools like Dynamic Workload Scheduler and a mix of A3 and G4 instances, we have the flexibility to optimize for both cost and speed. This ensures our engineers spend their time innovating rather than waiting for compute," said Wilcke.