Google bets on enterprise agents
- Google pushed AI agents as the core of its enterprise strategy at Cloud Next, not just chat demos. - It also unveiled an eighth‑generation TPU architecture splitting chips for training and inference to lower inference costs. - The move signals engineering work will focus on wiring agents into workflows, cost‑aware inference paths and cloud partnerships ( ).
Google used its Cloud Next conference on April 22 to pitch AI agents as the center of its enterprise business, not as another chatbot add-on. (reuters.com) At the Las Vegas event, Alphabet said it was folding a broader set of business AI products under the “Gemini Enterprise” name and adding governance and security controls for agents. Reuters reported Google executives framed the tools as production infrastructure for companies, while Chief Executive Sundar Pichai and Google Cloud Chief Executive Thomas Kurian tied the push to monetizing artificial intelligence. (reuters.com; cloud.google.com) An AI agent is software that can plan, choose tools and carry out multi-step tasks with limited human input. Kurian told Reuters the main use of Vertex AI had shifted from older machine-learning work toward customers building custom agents, and Google’s keynote said the new Gemini Enterprise app includes Agent Designer, an inbox for agent activity, long-running agents, skills and projects. (reuters.com; cloud.google.com) Google paired that software pitch with new chips because agents are expensive to run after they are built. The company said its eighth-generation Tensor Processing Units now split into TPU 8t for model training and TPU 8i for large-scale inference, the stage where a model answers live requests. (cloud.google.com; blog.google) That split reflects how enterprise AI spending has moved from training a few big models to serving huge volumes of prompts at low delay and lower cost. Google said TPU 8i is built for “cost-effective, near-zero latency inference,” while TPU 8t is aimed at frontier-model training and embedding-heavy workloads. (cloud.google.com; cloud.google.com) Google is making that case as cloud providers and model companies chase the same corporate budgets. Reuters said OpenAI and Anthropic have shifted resources toward business customers in recent months, and Google is trying to compete with both those firms and larger cloud rivals Amazon and Microsoft. (reuters.com) The company also used customer numbers to show demand. Google Cloud said nearly 75% of its customers use its artificial-intelligence products, more than 330 customers processed over 1 trillion tokens in the past 12 months, and 35 customers crossed 10 trillion tokens. (cloud.google.com) Google’s argument to buyers is that the stack works best when the model, agent tools, chips, networking and security controls are sold together. Reuters cited GE Appliances executive Marcia Brey saying her logistics and distribution team deployed AI faster with Google’s tools and data already stored in Google Cloud than with other products the company tested. (reuters.com) The caution around agents has not disappeared. Reuters noted the field has raised concerns about safety, reliability and oversight, which helps explain why Google spent part of the launch on governance, observability and security features rather than only model performance. (reuters.com; cloud.google.com) The immediate test is whether companies buy more than pilots. Google’s message in Las Vegas was that enterprise AI revenue will come from agents wired into everyday work — and from the cheaper infrastructure needed to keep those agents running. (reuters.com; blog.google)
