Google Cloud Next push

Google used its Cloud Next conference on April 22 to pitch a fuller enterprise AI stack, from custom chips to software that builds and runs workplace agents. (blog.google) At the center of the launch were two eighth-generation Tensor Processing Units, or TPUs, Google’s in-house AI chips. TPU 8t is aimed at training models, while TPU 8i is built for inference, the step where a trained model answers prompts and executes tasks. (blog.google, techcrunch.com) Google said its first-party models are now processing more than 16 billion tokens per minute through direct customer API use, up from 10 billion last quarter. Sundar Pichai also said just over half of Google’s machine-learning compute investment in 2026 is expected to go to the Cloud business. (blog.google) The company tied those chips to a broader “agentic enterprise” push, its term for businesses using artificial intelligence systems that can reason through multi-step work and take actions across software. Google Cloud said the new Gemini Enterprise Agent Platform is meant to build, scale, govern and optimize those agents on one stack. (cloud.google.com, cloud.google.com) Google also pushed the same idea into office software with Workspace Intelligence, a system that Google said can understand relationships across Gmail, Docs, Slides and Drive, along with a company’s projects, collaborators and internal knowledge. Google Workspace said the goal is to support “agentic work” inside everyday productivity tools rather than only in developer platforms. (workspace.google.com) In Workspace, Google announced “skills” that let teams turn standard operating procedures into reusable automations inside Workspace Studio. Google’s example was invoice review: the system compares a new invoice against prior ones in a user’s inbox and flags discrepancies. (workspace.google.com) The split-chip design shows how AI infrastructure is being reorganized around different jobs. Google’s technical blog said training, post-training and real-time serving now have different hardware needs, so TPU 8t and TPU 8i were designed as separate systems inside its AI Hypercomputer architecture. (cloud.google.com) That puts Google more directly against Nvidia in the cloud market, where access to training and inference hardware shapes which platform companies choose for AI workloads. TechCrunch reported that Google is positioning the new TPUs as custom alternatives as enterprises decide where to build and run large-scale AI systems. (techcrunch.com) Google’s pitch is not only faster hardware or newer models. It is that chips, models, agents, security and workplace software should be bought as one managed control layer for business processes. (cloud.google.com, blog.google) The next test is whether customers treat those agents like software features or like workers with rules, logs and approvals. Google spent Cloud Next arguing that more of that stack now belongs inside Google Cloud. (blog.google, cloud.google.com)