Google bets on agentic infrastructure

Google Cloud spent Next ’26 arguing that enterprise AI now depends less on a single model and more on the systems that let software agents work safely at scale. (cloud.google.com) (blog.google) An AI agent is a program that can plan, call tools, fetch data and complete a task in several steps instead of answering with one block of text. Google’s new Gemini Enterprise Agent Platform is built to manage that full cycle: build, deploy, govern and optimize. (cloud.google.com) (docs.cloud.google.com) Google launched the platform on April 22 as an evolution of Vertex AI, adding agent integration, DevOps, orchestration and security features for enterprise deployments. The company also tied it to the Gemini Enterprise app, which now includes Agent Designer, an Inbox for agent activity and support for long-running agents. (cloud.google.com 1) (cloud.google.com 2) The hardware pitch was just as explicit. Google said its eighth-generation Tensor Processing Unit family now splits into TPU 8t for training and TPU 8i for inference, reflecting the idea that teaching a model and running millions of agent requests are now different engineering problems. (cloud.google.com 1) (cloud.google.com 2) Google said TPU 8t scales to 9,600 chips in one superpod with 2 petabytes of shared high-bandwidth memory and delivers up to 2.7 times better training performance per dollar than Ironwood. It said TPU 8i targets serving and reasoning workloads and can deliver up to 80% better inference performance per dollar than Ironwood. (cloud.google.com) (forbes.com) Google paired that compute story with a data story. Its Agentic Data Cloud pitch centers on a cross-cloud lakehouse built around Apache Iceberg, with Google saying agents need always-on business context and direct access to operational data, not just a model endpoint. (cloud.google.com 1) (cloud.google.com 2) The company also leaned on interoperability. Google’s Agent2Agent, or A2A, protocol is meant to let different agents discover each other and work together across tools and services, while Cloud Run and Agent Builder documentation now describe how to host and test those systems. (cloud.google.com) (docs.cloud.google.com) Google used customer-scale numbers to show the bet is already commercial, not experimental. Thomas Kurian said nearly 75% of Google Cloud customers use its AI products, 330 customers processed more than 1 trillion tokens over the past 12 months, and 35 reached the 10-trillion-token mark. (cloud.google.com) (blog.google) Partners are being pulled into the same strategy. Google announced a $750 million partner fund for agent development and deployment, and said Deloitte is forming a dedicated Google Cloud Agentic Transformation practice focused on Gemini Enterprise while rolling Gemini Enterprise out to more than 100,000 of its own teams. (cloud.google.com) (googlecloudpresscorner.com) That makes Google’s latest AI pitch less about a chatbot and more about a control stack: chips for different workloads, data systems for live context, protocols for agents to coordinate, and governance layers for enterprises that need those agents to keep running. (cloud.google.com) (cloud.google.com)