Clouds target 'physical AI' workloads

Google Cloud and Nvidia used Google Cloud Next this week to pitch robots, factory software and AI agents as a single cloud workload, not separate bets. (blogs.nvidia.com) The partnership update, announced April 22 in Las Vegas, includes Google Cloud A5X bare-metal instances built on Nvidia Vera Rubin NVL72 systems. Nvidia said those clusters can scale to 80,000 Rubin graphics processors in one site and 960,000 across multiple sites. (blogs.nvidia.com) Google also launched the Gemini Enterprise Agent Platform on April 22 as an expanded version of Vertex AI for building, running and governing AI agents. Google said the platform adds tools such as Agent Designer, an activity Inbox, long-running agents, Skills and Projects. (cloud.google.com, cloud.google.com) An AI agent is software that can plan and carry out a multistep task; physical AI is the same idea pointed at machines that move in the world, such as robots and factory systems. Google said its new eighth-generation chips split that work in two, with TPU 8i for fast agent responses and TPU 8t for training larger models. (blog.google) That hardware push is tied to simulation software, the digital rehearsal before a robot touches a real floor. Google Cloud Next sessions and company posts said teams are running Nvidia Omniverse and Isaac Sim on Google Cloud G4 instances to build digital twins, test autonomous systems and train models before deployment. (www.googlecloudevents.com, cloud.google.com) Google is also selling governance as part of the package, not an add-on after deployment. Its documentation and product launch posts say Gemini Enterprise Agent Platform is meant to build, scale, govern and optimize agents, and outside coverage from the event reported each agent can get a cryptographic identity tied to auditable authorization policies. (docs.cloud.google.com, cloud.google.com, www.infosecurity-magazine.com) The cloud vendors have been moving toward this stack for months. Google said at Nvidia GTC on March 16 that agentic AI was already changing enterprise infrastructure needs, and in October 2025 Nvidia and Google had already put Omniverse and Isaac Sim images on Google Cloud for industrial digitalization. (cloud.google.com, blogs.nvidia.com) Google framed the demand in large numbers. Thomas Kurian said nearly 75% of Google Cloud customers use its artificial intelligence products, 330 customers processed more than 1 trillion tokens over the past 12 months, and 35 topped 10 trillion. (cloud.google.com) The pitch now is that the same cloud account should handle model training, agent orchestration, simulation and fleet operations. That makes “physical AI” look less like a robotics side project and more like a line item in mainstream cloud infrastructure. (blogs.nvidia.com, cloud.google.com)