Google's TPU v8 + Agent Platform

Google used its Cloud Next event on April 22 to pair new eighth-generation Tensor Processing Units with a new Gemini Enterprise Agent Platform, tying chips and software into one pitch for building and running artificial intelligence agents at scale. (blog.google) The new chip line comes in two versions: TPU 8t for training models and TPU 8i for inference, the stage when a trained model answers prompts or takes actions. Google said both are part of its eighth-generation custom Tensor Processing Unit family and are “coming soon.” (blog.google) Google also said Gemini Enterprise Agent Platform is the “evolution of Vertex AI,” combining model selection, model building and agent building with new tools for integration, DevOps, orchestration and security. The company announced it the same day at Cloud Next ’26. (cloud.google.com) A Tensor Processing Unit is Google’s in-house artificial intelligence chip, built to handle the large volumes of math behind model training and inference. An agent platform is the software layer that lets companies build systems that can call tools, connect to business data and carry out multi-step tasks with controls around them. (blog.google; docs.cloud.google.com) Google’s argument is that customers want those layers bundled together. Its Agent Platform documentation says the service includes access to more than 200 models through Model Garden, plus underlying TPU and graphics processing unit infrastructure, MLOps tooling, and governance, security and compliance controls. (docs.cloud.google.com) The company framed the launch around enterprise demand for larger numbers of agents. In a Cloud Next post, Chief Executive Sundar Pichai said Google is helping organizations “build and manage thousands of AI agents,” and said Google’s first-party models now process more than 16 billion tokens per minute through direct customer API use, up from 10 billion the prior quarter. (blog.google) Google tied the new platform directly to its Gemini model family. The company said Agent Platform provides access to Gemini 3.1 Pro, Gemini 3.1 Flash Image, which Google also calls Nano Banana 2, and Lyria 3. (blog.google) The product also folds agent management into Google’s broader business software push. Google’s Gemini Enterprise app says companies can create, deploy and govern Google-made, custom-built and third-party agents in one place, and that standard and plus editions can upload and govern agents built on external platforms. (cloud.google.com) That bundling follows Google’s earlier chip cadence. In April 2025, the company introduced Ironwood as its seventh-generation Tensor Processing Unit, aimed at inference-heavy “thinking” models, and now it is moving to an eighth-generation family split between training and inference workloads. (blog.google; blog.google) The immediate change for Google Cloud customers is less about a single chip spec than about where new artificial intelligence work will live. Google said Vertex AI services and future roadmap items are being delivered through Agent Platform, while the new TPU line supplies the computing layer underneath that stack. (cloud.google.com; docs.cloud.google.com)