Google bets on AI agents

Google spent this week telling corporate customers that AI agents — software that can plan and carry out multistep tasks — are becoming the center of its cloud business. (msn.com) At Google Cloud Next in Las Vegas on April 22, Chief Executive 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. He also said just over half of Google’s machine-learning compute investment in 2026 is expected to go to the cloud business. (blog.google) Google used customer examples to show what it means by an “agentic enterprise.” In its roundup, Google said Capcom is using agents for game-testing work, Citi for financial-advice workflows, and other large companies for customer service, software operations, and internal productivity tasks. (blog.google) An AI agent is a model hooked up to tools, memory, and business software so it can do a chain of actions instead of answering one prompt. Google’s pitch is that companies will buy those agents through Google Cloud, then pay for the models, data systems, security tools, and computing power behind them. (msn.com) (blog.google) That sales pitch comes as cloud providers race to turn generative artificial intelligence from a chatbot feature into a larger software business. Reuters reported that Google is pushing deeper into enterprise software as investors press big tech companies to show how artificial-intelligence spending turns into revenue. (msn.com) Google tied that strategy to new hardware. On April 22 it introduced two eighth-generation Tensor Processing Units, or TPUs: TPU 8t for training large models and TPU 8i for inference, the stage where a trained model answers requests. (blog.google) Google said TPU 8i is designed for the fast response times AI agents need when they plan and execute tasks, while TPU 8t is tuned for training and can run larger models in one shared memory pool. CNBC reported Google said the training chip delivers 2.8 times the performance of the seventh-generation Ironwood TPU at the same price, and the inference chip improves performance by 80%. (blog.google) (cnbc.com) Google is also pairing its own chips with Nvidia systems when customers want them. TechCrunch reported that Mira Murati’s Thinking Machines Lab signed a multibillion-dollar Google Cloud deal for infrastructure built on Nvidia’s GB300 chips, and Google separately announced an expanded agreement with the startup at Cloud Next. (techcrunch.com) (marketwatch.com) That mix of agents, chips, and startup capacity reflects the economics of the market Google is chasing. The more often companies deploy agents into customer support, marketing, coding, and back-office work, the more pressure they put on cloud providers to cut inference costs and keep enough computing capacity online. (blog.google) (msn.com) Google’s message in Las Vegas was that the next cloud contract may not be about renting raw servers alone. It may be about selling a company the worker, the manager, and the factory floor in one bundle. (blog.google) (msn.com)