Google bets on AI agents

Google used its Cloud Next conference on April 22 to sell companies on AI agents that can carry out office tasks, not just chatbots that answer questions. (reuters.com) Alphabet Chief Executive Sundar Pichai said at the Las Vegas event that Google is helping organizations build and manage “thousands of agents,” and the company introduced a Gemini Enterprise Agent Platform alongside new agent tools in Google Workspace. (blog.google) Reuters reported that Google spotlighted customers including Wendy’s and Mercedes-Benz using agent-style systems, part of a broader enterprise software push as it tries to turn artificial intelligence into recurring corporate revenue. (reuters.com) An AI agent is software that can plan and complete a multistep job — like pulling data, writing a draft, and sending it for approval — with less human prompting than a standard chatbot. Google said its new TPU 8i chip is built for that kind of fast-response work, while TPU 8t is aimed at training larger models. (blog.google) Google paired that product pitch with infrastructure news. TechCrunch reported on April 22 that Mira Murati’s Thinking Machines Lab signed a new multibillion-dollar agreement to expand its use of Google Cloud systems powered by Nvidia’s latest GB300 graphics chips. (techcrunch.com) That matters because enterprise buyers are now weighing two costs at once: the price of the model and the price of running it all day inside real workflows. Google’s April 22 announcements bundled both sides — software to automate work and custom chips meant to lower the cost and delay of serving those systems. (blog.google) Google has been building Tensor Processing Units, or TPUs, for more than a decade as in-house chips for AI workloads, and it has increasingly turned them into a cloud product to compete with Nvidia-based offerings from rivals. In 2025, Google introduced Ironwood as its seventh-generation TPU and said it was designed specifically for inference, the step where a trained model generates answers. (blog.google) The 2026 update split that strategy in two. Google said TPU 8i is optimized for low-latency inference for “highly responsive agentic AI,” while TPU 8t is designed for training on a large shared memory pool, a setup aimed at bigger and more complex models. (blog.google) Google is making that case while Microsoft, Amazon and OpenAI partners are all racing for the same enterprise budgets. Reuters said Google’s message to investors was that agents are becoming a linchpin of how it plans to monetize artificial intelligence inside its cloud business. (reuters.com) The pitch from Las Vegas was straightforward: Google wants companies to buy a full stack, from the chips in the data center to the software worker on the screen. (blog.google)