Workato AI Research Lab is the leader in AI-driven enterprise automation, helping over 12,000 global customers unify data, processes, and applications. In 2025, the company launched Workato One, a unified platform that brings together everything organizations need to build and deploy enterprise-ready agents across business functions.

At the heart of Workato’s AI innovation is Workato AI Research Lab, a San Francisco-based research hub dedicated to advancing the science of autonomous agents through synthetic evaluation, automated reinforcement learning, and custom model optimization. To power this frontier research, Workato partnered with DigitalOcean to build their high-performance infrastructure using NVIDIA H200 GPU Droplets and DigitalOcean Managed Kubernetes, achieving a 2-3x acceleration in time-to-value and a 67% reduction in inference costs.

Workato’s Inference and Support Needs

For the lean engineering team at Workato AI Research Lab, the primary obstacle wasn’t just finding GPUs, but also the operational overhead required to manage them. As a research-focused unit, every hour spent debugging network configurations or container orchestration was an hour lost to core AI innovation. The team needed an infrastructure partner that functioned as an extension of their own engineering staff, providing both rapid deployment and proactive optimization.

The AI Labs required infrastructure capable of handling distributed training and sustained, reasoning-heavy inference under real production load. Not only was DigitalOcean able to provide high-performance NVIDIA HGX™ H200 GPUs faster than any other provider for the team to begin their work, the Workato team quickly discovered the significant performance boost and TCO improvement as a result of DigitalOcean’s inference-optimized architecture and simplified experience.

“Before DigitalOcean, we didn’t have a dedicated solution for in-house training and multi-node serving, and that was a major blocker for AI research,” said Oscar Wu, AI Research Scientist at Workato AI Research Lab.

“DigitalOcean was the fastest provider to get us up and running, enabling us to advance our AI programs. Beyond access to GPUs, the collaboration on performance optimization coupled with the support from their team of solutions architects, accelerated our progress by roughly two to three times.”

Beyond procurement of GPUs, Workato was also hesitant about the lack of high-touch support from hyperscale providers. For a small team, the black box nature of legacy cloud support was a major risk. They required a partner who could provide deep technical root-cause analysis rather than generic documentation.

From Evaluation to Impact: A High-Performance Backbone for Frontier Research

To move beyond these blockers, Workato conducted a rigorous evaluation of multiple providers, including both hyperscale cloud providers such as AWS and Google Cloud Platform and AI neoclouds including Baseten.

Workato ultimately selected DigitalOcean for its superior execution speed and the specialized expertise of its engineering team, which worked hand in hand with Workato to optimize GPU performance and costs. This partnership enabled Workato to build a dedicated infrastructure capable of handling both intensive model training and high-throughput inference, with a level of agility the hyperscalers couldn’t match.

The transition to an optimized DigitalOcean environment fundamentally transformed Workato’s research economics. By controlling the full hardware stack—centered on interconnected NVIDIA H200 GPU clusters—the Labs team was able to reduce the cost of serving specialized agents while dramatically increasing the speed at which those agents respond to business events. These H200 clusters provide a 33% hardware cost advantage over the A100 series while delivering significantly higher memory bandwidth, a critical factor for Workato’s high-stakes agentic workloads.

Precision Engineering and Quantified Results

To support frontier models like Llama-3.3-70B and Llama-3.1-8B, the Labs team implemented a highly tuned software and hardware stack. By serving models in FP8 and FP4 numerical precision, they maximized throughput without sacrificing accuracy.

The team utilized NVIDIA Dynamo v0.4.1 combined with vLLM (virtual Large Language Model), leveraging PagedAttention to manage memory for long-context agentic reasoning. To facilitate multi-node communication for larger models, DigitalOcean configured the nodes with NVLink. As Huang explains, “We use them for just doing multi-node stuff, like training bigger models that are not trainable on one node.”

This technical precision led to immediate, measurable gains:

• 67% Lower Cost per 1M Tokens: Infrastructure costs dropped to just $0.77 per million tokens, significantly outperforming high-tier API providers.
• 67% Throughput Increase: The optimized H200 setup reached 13,561 tokens/sec per GPU, a massive leap from the baseline of 8,111 tokens/sec.
• Reliable Low Latency: For real-time “Genie” interactions, the team achieved a P50 latency of 13 seconds at 3.38 requests per second even on heavy 70B parameter models.

Deep Engineering Partnership as a Research Multiplier

Beyond raw metrics, the deciding factor for Workato was the deep AI engineering partnership that turned DigitalOcean into a true extension of the Workato team. During the critical GPU setup and testing phase, the team hit a roadblock with a broken etcd image reference in a standard Docker configuration. Rather than providing a generic support ticket response, DigitalOcean’s Solutions Architect, Rithish, worked directly with Workato to identify the root cause and stabilize the cluster. This proactive intervention saved the team an estimated one week of troubleshooting.

This level of collaborative engineering, which encompasses setup, rigorous benchmarking, and ongoing maintenance, gave Workato the confidence to move at research speed. “DigitalOcean is the only provider that lets me focus on research instead of managing infrastructure,” said Kevin Huang, Infrastructure Engineer at Workato AI Labs. “We can provision GPUs quickly, deploy inference workloads in production, and iterate on real customer traffic without getting bogged down in platform complexity. That speed has been critical to maintaining our momentum.”

This collaborative synergy turned an aggressive sprint into a success. “Being able to collaborate on an aggressive sprint and achieve results that were impressive was a major factor in our decision to use DigitalOcean,” says Wu. The result was a “two to three X speed up in time to value.”

What’s Next for Workato AI Research Lab: Scaling the Autonomous Enterprise

With a stable infrastructure foundation, Workato AI Labs is now focused on the next generation of agentic capabilities. This includes developing Self-Improving Agents that can observe business operations, identify inefficiencies, and autonomously design new process “recipes.”

To support this, Workato is scaling its Agent Knowledge Base—a continuously evolving memory layer—and its support for the Model Context Protocol (MCP), which allows agents to share skills across different platforms.

“We’re building towards a future where agents can observe business operations, and optimize existing ones in this continuous self-improving loop,” says Wu. “DigitalOcean and NVIDIA’s infrastructure is foundational to that work, as we need reliable, high-performance compute to keep on pushing the frontier.”