UiPath’s GPU Fleet: A Capybara’s Guide to Scaling Agentic AI

UiPath, the enterprise automation heavyweight, is diving deep into agentic AI—where autonomous agents don’t just follow scripts but actually reason, decide, and execute complex business processes. That’s a big leap from simple task automation, and it demands serious computational muscle. But orchestrating hundreds of GPUs in perfect harmony? That’s like herding capybaras through a swamp—tricky, but with the right approach, it’s a smooth glide.

UiPath’s answer? A shared GPU fleet on Google Cloud, balancing heavy-duty training with real-time inference, all while keeping costs and latency in check. Let’s wade into how they did it.

The Heavy-Duty Math Problem

UiPath’s platform handles intelligent document processing (IDP), computer vision, and LLM-powered reasoning—all requiring serious number-crunching. Their models, built on LLAMA and Qwen architectures, need to extract data from messy, real-world paperwork with human-like clarity. Every millisecond of latency matters, and moving from a cool demo to a reliable production tool without exploding costs meant rethinking their silicon strategy.

The Swamp of Challenges

UiPath’s old approach—provisioning GPU nodes on demand—worked when cloud capacity was cheap and elastic. But as AI ambitions grew, they hit three snags:

• Spiky workloads: Buying extra capacity for peak demand meant paying for idle GPUs during quiet times.
• Supply bottlenecks: High-end A3 VMs with H100s are gold for training, but global demand outstrips supply, making scaling a slow crawl.
• Operational overhead: Maintaining dedicated clusters in multiple regions for low latency added layers of inefficiency.

The Capybara Solution: A Shared Fleet

Instead of treating GPUs as product-specific elastic resources, UiPath made them a shared strategic asset. Their machine learning services (MLS) platform now prioritizes workloads across teams and time windows. During the day, the fleet handles real-time inference; at night, it switches to batch training. This fleet-level coordination maximizes utilization and reduces contention, all without per-instance elasticity.

Why Google Cloud’s AI Hypercomputer?

UiPath chose to expand on Google Cloud, leveraging the AI Hypercomputer—a system-level approach that integrates performance-optimized hardware, open software, and flexible consumption. It minimizes friction between hardware and software, letting engineers focus on model performance, not infrastructure.

Key tools include Google Kubernetes Engine and the Dynamic Workload Scheduler (DWS), which lets UiPath book GPU capacity in advance, solving supply bottlenecks. They use A3 VMs (H100s) for training and G4 VMs (RTX Pro 6000) for inference—right-sizing silicon for the job.

Real-World Wins

With consistent GPU access, UiPath delivers advanced IDP at scale. For example, Omega Healthcare automates over 100 million transactions with 99.5% accuracy, cutting processing time by 40%. Thermo Fisher Scientific processes 53% of invoices without human help, slashing processing time by 70%.

Lessons from the Swamp

UiPath’s journey offers three takeaways for building similar platforms:

1. Decouple capacity: Pool resources to smooth out spikes.
2. Schedule, don’t react: Use tools like DWS to book compute in advance.
3. Right-size the silicon: Use A3 for training, G4 for inference.

For more on cloud infrastructure, check out our reviews on Google Cloud, Cloudflare Workers, and Neon Database.

Original announcement published on Google Cloud.