Google Cloud: new TPUs and agents

Google Cloud used its Next conference on April 22 to launch two new artificial intelligence chips and a new platform for building and managing AI agents. (techcrunch.com) The chips are Google’s eighth-generation tensor processing units, or TPUs: TPU 8t for training models and TPU 8i for inference, the work models do after a prompt arrives. Google said TPU 8t can deliver up to 3x faster training, 80% better performance per dollar, and clusters of more than 1 million TPUs. (techcrunch.com) Google said TPU 8i is designed for AI agents that need to reason through multi-step tasks quickly, while TPU 8t is built for large training jobs with a single, massive pool of memory. The company announced both chips on April 22 in Las Vegas at Google Cloud Next ’26. (blog.google) An AI agent is software that can plan, call tools, and carry out a sequence of actions instead of answering one prompt at a time. Google’s new Gemini Enterprise Agent Platform is meant to give companies one place to build, run, govern, and monitor those systems. (cloud.google.com) Google said the platform is the evolution of Vertex AI and adds agent integration, orchestration, security, and DevOps controls on top of its existing model-building tools. It also includes Agent Studio for low-code work, an upgraded Agent Development Kit for code-first teams, and a re-engineered runtime for long-running agents that can keep state for days. (cloud.google.com) The management layer is as important as the model layer in Google’s pitch. The company said customers are moving from asking whether they can build an agent to asking how to manage thousands of them, and it positioned the new platform as “mission control” for that job. (blog.google) Google is also trying to make outside agents part of the same system. The company said partner-built agents can run natively inside Gemini Enterprise Agent Platform, so internal agents and third-party tools can share the same governance and identity controls. (cloud.google.com) The chip push does not replace Nvidia in Google’s cloud. TechCrunch reported that Google still plans to offer Nvidia’s Vera Rubin chips later this year and is working with Nvidia on networking software called Falcon to improve performance in its cloud. (techcrunch.com) Google tied the announcements to rising usage across its cloud AI business. Sundar Pichai said Google’s first-party models now process more than 16 billion tokens per minute through direct customer API use, up from 10 billion last quarter, and said just over half of Google’s machine learning compute investment in 2026 is expected to go to the Cloud business. (blog.google) The package Google showed on April 22 is a full-stack sales pitch: custom chips for training and inference, Gemini models, and a control plane for fleets of agents. The next test is whether enterprises buy the whole stack or keep mixing Google’s tools with Nvidia hardware and rival software. (cloud.google.com)