エージェントAIのスケーリング：UiPathがAI HypercomputerでGPUプラットフォームを構築した方法Scaling agentic AI: How UiPath built its high-performance GPU platform on AI Hypercomputer

UiPath, a company known for enterprise automation, has built a high-performance GPU platform on Google Cloud's AI Hypercomputer to support the large-scale deployment of agentic AI, according to a Google Cloud blog post. The move matters because it illustrates a broader shift underway across enterprise software: the transition from software that automates predefined tasks to autonomous agents that can reason, make decisions, and execute complex, multi-step business processes on their own. That evolution demands far more computational power and far more reliable infrastructure than earlier generations of automation tools required.

UiPath describes itself as a market leader in enterprise agentic automation and business orchestration, and it frames its work as helping to pioneer an industry-wide move toward agentic AI. In practice, the company is deploying autonomous agents that actively reason, make decisions, and carry out business processes across disparate systems. This is a meaningful departure from the company's roots in robotic process automation, or RPA, where software "robots" typically follow scripted rules to handle repetitive tasks such as data entry or moving information between applications. Cognitive, decision-making agents are considerably more computationally intensive because they rely on large models that must be trained and then served in production.

The core technical challenge, as the company presents it, is scale and reliability. Being able to orchestrate hundreds of GPUs in harmony, UiPath argues, is what separates a research experiment from a global AI platform capable of serving the world's largest enterprises. That orchestration involves balancing large, resource-hungry training jobs against real-time inference, where agents must respond quickly to live requests without interruption. Training and inference place different demands on hardware and scheduling, and running both on shared infrastructure requires careful coordination to avoid contention and to keep latency low for production workloads.

AI Hypercomputer is Google Cloud's integrated approach to this kind of workload. Rather than treating chips, networking, storage, and software as separate pieces, it packages them as a single supercomputing architecture designed to run demanding AI training and serving jobs. The system supports both GPUs and Google's own Tensor Processing Units, or TPUs, and emphasizes high-bandwidth networking between accelerators, which is important when a job is spread across many chips that must exchange data frequently. For a company scaling agentic workloads, that tight integration is intended to reduce the engineering burden of stitching together infrastructure and to make large clusters of accelerators behave more predictably.

The emphasis on orchestrating hundreds of GPUs points to the practical difficulty of operating at this level. Coordinating large fleets of accelerators typically depends on schedulers and cluster management systems that decide how workloads are distributed, how failures are handled, and how capacity is shared. In the broader ecosystem, tools such as Kubernetes and specialized job schedulers are commonly used to manage these clusters, while frameworks for distributed training help split models and data across many devices. The reliability requirement UiPath highlights is not incidental: enterprise customers running critical business processes generally expect consistent uptime and performance, which raises the bar beyond what an isolated experiment would need.

The announcement fits into a wider industry pattern in which established software vendors are re-architecting their platforms around large models and autonomous agents. Cloud providers, including Google Cloud with its Gemini models and Vertex AI platform, Microsoft with Azure, and Amazon Web Services, have each been positioning integrated infrastructure and model offerings to capture this demand. Agentic AI, in which systems plan and act across multiple steps and tools rather than answering a single prompt, has become a prominent theme in enterprise technology, though the term is used loosely and real-world reliability remains a key concern for adopters.

For UiPath, building on AI Hypercomputer appears to be a bet that owning a robust, scalable GPU platform is a prerequisite for delivering agentic automation to large organizations. The blog post is published from Google Cloud's perspective, so it is likely to foreground the strengths of that infrastructure. Even so, the underlying message is consistent with what many enterprises are weighing: moving from task automation to decision-making agents is as much an infrastructure problem as it is a modeling one, and the ability to run training and inference reliably at scale is central to making these systems production-ready.