Google Bets on Agents
- Google is positioning AI agents—task-oriented digital assistants—as central to its enterprise monetisation strategy. - It rolled out a Gemini Enterprise Agent Platform and unveiled an eighth-gen TPU architecture split for training and inference. - The company also signed major infrastructure deals and is building chips and cloud capacity to scale agent fleets and cut inference costs. (reuters.com) (techcrunch.com) (benzinga.com)
Google used its Cloud Next conference on April 22 to make AI agents the center of its enterprise pitch. Reuters reported the company is tying its artificial-intelligence sales push to software that can take actions for workers, not just answer questions. (reuters.com) Google said its new Gemini Enterprise Agent Platform is a system for building, scaling, governing and optimizing agents inside companies. The platform folds together Vertex AI model tools with new controls for integration, orchestration, security and DevOps. (blog.google) (cloud.google.com) At the same event, Google introduced a Gemini Enterprise app that lets employees discover, create, share and run agents in one place, with connectors to company data and third-party systems. Reuters said Google also showed a dedicated inbox where agents can post updates and progress reports for workers. (cloud.google.com) (reuters.com) An AI agent is software that can carry out a job across several steps, like gathering data, making a plan and sending back a result. Google’s pitch is that companies now need tools to manage “thousands” of those agents, not just a single chat window for one employee. (blog.google 1) (blog.google 2) That helps explain the chip announcements. Google said its eighth-generation Tensor Processing Units split into TPU 8t for training models and TPU 8i for low-latency inference, the stage where a trained model generates answers fast enough for live use. (blog.google) Google said TPU 8i is built for “fast, collaborative AI agents,” while TPU 8t is aimed at large training runs with a single massive memory pool. The company said both chips are available for customer requests now ahead of general availability later in 2026. (blog.google 1) (blog.google 2) The spending behind that push is rising fast. Sundar Pichai said Google’s first-party models now process more than 16 billion tokens per minute through direct customer application-programming-interface use, up from 10 billion in the prior quarter, and that just over half of Google’s 2026 machine-learning compute investment is expected to go to the Cloud business. (blog.google) Google is also using outside demand to fill that infrastructure. TechCrunch reported that Mira Murati’s Thinking Machines Lab signed a new multi-billion-dollar agreement to expand its use of Google Cloud systems powered by Nvidia’s GB300 graphics processors. (techcrunch.com) The company is making that case as it tries to turn heavy AI capital spending into enterprise revenue. Reuters said Google is pitching agents as a way to overhaul daily office work while competing with OpenAI and Anthropic for corporate customers. (reuters.com) (mercurynews.com) Google’s message in Las Vegas was that the next sale is not one chatbot license but a managed fleet of software workers. The harder part now is proving companies will buy enough of them to cover the chips, cloud capacity and power needed to run them. (reuters.com) (blog.google)
