Google ties agents to hardware

Google used its April 22 Cloud Next keynote to sell AI agents as an enterprise product — and the pitch came bundled with new chips and cloud infrastructure. (cloud.google.com) (usnews.com) At the Las Vegas event, Google introduced Gemini Enterprise Agent Platform as the successor to Vertex AI for building, scaling, governing, and optimizing agents. Google said nearly 75% of Google Cloud customers already use its AI products, and 330 customers processed more than 1 trillion tokens each over the past 12 months. (cloud.google.com 1) (cloud.google.com 2) 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) An AI agent is software that can take a goal, call tools, and complete a multistep task with limited human input. Google’s new platform wraps those agents with model access, orchestration, security controls, and operations tools inside one managed service. (cloud.google.com) (blog.google) The hardware piece matters because those agents do not run in the abstract; they run on accelerators, networking, and data-center layouts chosen by the cloud provider. Google paired its agent push with seventh-generation Ironwood TPUs now generally available and a roadmap that extends to eighth-generation TPUs. (cloud.google.com) (blog.google) NVIDIA used the same conference to tie its roadmap to Google Cloud’s. The company said Google Cloud will offer Vera Rubin-powered A5X bare-metal instances, confidential virtual machines with Blackwell GPUs, and Gemini on Google Distributed Cloud running on Blackwell and Blackwell Ultra GPUs in preview. (blogs.nvidia.com) NVIDIA said A5X systems are built on Rubin NVL72 racks and are designed to scale to 80,000 Rubin GPUs in one site and 960,000 across multiple sites. It also said the new setup can deliver up to 10 times lower inference cost per token and 10 times higher token throughput per megawatt than the prior generation. (blogs.nvidia.com) Google had already been moving in this direction with Blackwell-based A4X virtual machines, which package 72 Blackwell GPUs and 36 Grace central processors into one system for large model training and inference. Google said those A4X machines became generally available on May 29, 2025. (cloud.google.com) Google says its agent platform supports first-party Gemini models and is integrated with Google’s data and security stack. NVIDIA says the same platform will also work with its Nemotron open models and NeMo software, which gives buyers more than one model family but still keeps orchestration, security, and infrastructure inside the Google Cloud environment. (blog.google) (blogs.nvidia.com) Reuters reported that Google is using agents as a central part of its enterprise revenue push as it competes with Microsoft, Amazon, OpenAI, and Anthropic for corporate AI spending. The result is that buying an “agent platform” increasingly also means buying a cloud architecture, a chip roadmap, and the switching costs that come with both. (usnews.com)