Google unveils TPU v8 and controls

A Tensor Processing Unit is Google’s in-house artificial intelligence chip, built to train models and answer prompts at scale. On April 22, Google used Cloud Next 2026 in Las Vegas to unveil eighth-generation TPUs and a new control layer for managing AI agents. (blog.google 1) (blog.google 2) Google split the new hardware into two products: TPU 8t for training large models and TPU 8i for running them in production. Google said both systems are coming later in 2026 and are designed for “training and inference” as separate jobs rather than one compromise chip. (blog.google) (cloud.google.com) That design choice reflects how AI workloads have changed. Google said newer systems need longer context windows, multi-step reasoning and continuous feedback loops, which put different pressure on training clusters and live-serving infrastructure. (cloud.google.com) Google paired the chips with what it called the Gemini Enterprise Agent Platform, a system to build, scale, govern and optimize large fleets of software agents. Chief executive Sundar Pichai said the question has shifted from building one agent to managing “thousands of them.” (blog.google) The conference message was broader than faster chips. Google’s official recap put Agentic Defense, an Agentic Data Cloud and the agent platform alongside the TPU launch as the core of the event. (cloud.google.com) (blog.google) On security, Google announced three new agents in Google Security Operations: a Threat Hunting agent and Detection Engineering agent, both in preview, plus a Third-Party Context agent coming soon to preview. The company said the products combine Google threat intelligence with Wiz’s cloud and artificial intelligence security tools. (cloud.google.com) On data, Google and SAP said SAP BDC Connect for BigQuery will offer bidirectional, zero-copy data sharing, which means companies can analyze shared data without moving or duplicating the underlying files. Google said the goal is to keep business data current enough for “mission-critical AI workloads.” (cloud.google.com) Google also used the event to show how much demand it is seeing. Pichai said Google’s first-party models were processing more than 16 billion tokens per minute through direct customer API use, up from 10 billion in the prior quarter, and said just over half of Google’s machine-learning compute investment in 2026 is expected to go to Cloud. (blog.google) For cloud customers, the announcements point to a stack built around control as much as raw speed: separate chips for separate jobs, a platform for supervising agents, and security and data products wrapped around both. That was Google’s pitch in Las Vegas, and it is the frame the company used to close out Next 2026. (cloud.google.com)