Google unveils its 8th generation TPUs and Gemini AI

Google Revolutionizes Its TPUs with the 8th Generation

At the Cloud Next '26 conference, Google introduced its eighth-generation TPUs, marking a significant advancement in the field of artificial intelligence. For the first time, Google has decided to split its Tensor Processing Units into two distinct categories: the TPU 8t, dedicated to training, and the TPU 8i, designed for inference. This separation allows Google to meet the growing demand for inference, particularly for AI agents that require planning, action, and loop learning capabilities.

Unlike Nvidia, which focuses on the raw performance of each chip, Google is betting on scale. Indeed, Google’s TPUs can be connected in massive clusters, reaching up to a million chips through the use of optical switches and the Virgo network. This approach enables Google to overcome the limitations of slower Ethernet or InfiniBand connections, providing a "goodput" rate of about 97%, meaning the chips spend the majority of their time training rather than waiting for data.

The Register highlights that Nvidia's future Rubin GPUs offer more computing power and higher memory bandwidth per chip than the TPU 8t. However, Google’s advantage lies in its ability to efficiently connect a large number of chips. Google uses optical switches to connect 9,600 TPUs in a single pod, and its managed Lustre storage system pushes data directly into the memory of the accelerators.

The TPU 8i, on the other hand, is optimized for inference with increased on-chip SRAM and faster HBM. This allows for more key-value cache memory to be retained directly on the chip, preventing cores from idling while waiting for data. Additionally, a collective acceleration engine has been developed to enhance the performance of mixture of experts models, and a new network topology, called Boardfly, reduces latency between chips.

For the first time, both TPUs now operate on Google’s Arm-based Axion CPUs, representing a major technological integration.

The Gemini Enterprise Platform for Autonomous AI Agents

In parallel, Google launched the Gemini Enterprise Agent Platform, which consolidates its existing AI services under a new framework. This platform is built on Vertex AI and offers innovative tools for creating and executing AI agents. Developers can now use a mapping tool to visualize interactions between multiple agents in a flowchart format, while the Agent Studio allows for the creation of agents using natural language.

To prevent the proliferation of similar agents, a central registry has been established. Furthermore, agents can now autonomously manage complex processes without requiring human intervention at every step. Sandbox testing environments allow agents to execute their own code safely, and a Memory Bank provides long-term memory, preventing agents from starting from scratch in each session.

Security is a priority for Google, which has integrated cryptographic identities for each agent, upstream filters to prevent prompt injection, and anomaly detection systems to identify suspicious behaviors. Simulation tools also allow for testing agents against synthetic user interactions before their actual deployment.

Available models include Gemini 3.1 Pro, Nano Banana 2, and Lyria 3, as well as Claude Opus from Anthropic, Sonnet, Haiku, and the newly added Claude Opus 4.7.

The accompanying Gemini Enterprise application targets end users: employees can assemble their own agents from building blocks, track ongoing tasks in an inbox-style view, and edit documents directly within the application.

Workspace Intelligence: A Centralized AI Integration

In addition to these innovations, Google introduced Workspace Intelligence, a layer that centrally connects information across applications such as Gmail, Docs, Drive, Meet, and Chat. This integration allows AI models to understand the relationships between different types of content, facilitating tasks like creating documents or events from conversations.

In Gmail, for example, Gemini sorts incoming messages and summarizes topics, while in Google Chat, users can create calendar events or documents directly from a conversation. In Docs, Gemini can draft content from emails and files, and in Sheets, it builds dashboards. Slides allows for assembling presentations, and Drive Projects groups files and emails into topic-based workspaces.

For businesses looking to migrate to these new solutions, Google offers a faster migration path from Microsoft 365, thereby strengthening its offering for companies eager to adopt more integrated and secure AI solutions.

All of these innovations are marketed under the name "Agentic Enterprise," highlighting Google’s integrated and advanced approach in the field of artificial intelligence.