Google Announces Eighth-Generation TPUs at Cloud Next '26
Google presented its latest hardware and software during the Cloud Next '26 event. The highlights include the eighth-generation Tensor Processing Units, a new agent platform, and an AI layer for Workspace. Officials described the lineup as part of an "Agentic Enterprise" approach.
For the first time, the company divided its TPUs into distinct versions: TPU 8t for training and TPU 8i for inference. Amin Vahdat, senior vice president and chief technologist for AI and infrastructure, explained that this change addresses growing needs from AI agents that plan, act, and learn in cycles.
Google prioritizes large-scale operations over maximum performance from individual chips. Reports from The Register point out that Nvidia's Rubin GPUs provide higher compute and memory bandwidth per chip than TPU 8t. However, success in training advanced models depends on connecting many chips effectively.
Scaling TPUs with Advanced Networking
Google claims advantages in cluster size. Nvidia connects up to 576 GPUs in one NVLink domain, then relies on Ethernet or InfiniBand. Google links 9,600 TPUs in a pod using optical circuit switches. The Virgo Network connects multiple data centers into groups of up to one million TPUs. A managed Lustre storage system delivers data directly to accelerator memory. The setup aims for 97 percent goodput, the portion of time chips use for actual training instead of handling checkpoints or errors.
The TPU 8i focuses on inference with extra on-chip SRAM and quicker HBM access. Increased SRAM holds more of the key-value cache, the model's record of past responses, keeping cores active. A Collective Acceleration Engine boosts mixture-of-experts models. Boardfly, a new network design, reduces delays between chips.
Both TPU types now pair with Google's Arm-based Axion CPUs, a first for the hardware.
Google has developed TPUs since 2016 as custom accelerators for machine learning workloads. They power services like Search and Translate, and Cloud Next serves as the annual showcase for Google Cloud updates since 2015.
Gemini Enterprise Agent Platform for Secure Agents
Google combined its AI tools into the Gemini Enterprise Agent Platform, extending Vertex AI. Developers use a flowchart tool to outline agent interactions and Agent Studio to build agents via natural language prompts. A central registry avoids duplicate agents across teams.
For deployment, agents manage multi-step tasks without constant human checks. Sandboxed environments allow safe code execution or browser tasks. The Memory Bank provides persistent recall across sessions.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Security features counter risks from autonomous agents: cryptographic identities per agent, filters for prompt injection, and detection for odd actions like improper data access or endless loops. Simulation tools test agents with fake user scenarios.
Supported models cover Gemini 3.1 Pro, Nano Banana 2, Lyria 3, plus Anthropic's Claude Opus, Sonnet, Haiku, and the recent Claude Opus 4.7.
The Gemini Enterprise app lets workers build agents from components, monitor tasks in a dashboard like an inbox, and edit files within the app.
Vertex AI, launched in 2021, offers managed ML services, making this platform a natural evolution for enterprise agent development.
Workspace Intelligence Connects Apps
Workspace Intelligence forms a unified layer across Gmail, Docs, Drive, Meet, and Chat. It enables Gemini and agents to grasp connections between emails, meetings, messages, and documents, rather than searching apps separately.
Specific uses include Gemini sorting and summarizing Gmail messages, creating events or docs from Chat threads, drafting Docs from emails and files, generating Sheets dashboards, assembling Slides presentations, and organizing Drive Projects by topics. Google provides quicker migration options from Microsoft 365 for switchers.
Google Workspace, formerly G Suite since 2020, supports over 3 billion users daily, and this AI integration builds on prior Gemini additions.