Back
Technology

Google Announces Workspace Intelligence and Eighth-Generation TPUs at Cloud Next 2026

Generated on: Last updated: 1 sources
View source

Google Announces Workspace Intelligence and New TPU Chips at Cloud Next 2026

At its Cloud Next 2026 event, Google unveiled a new system for personalized AI across its productivity suite and introduced the eighth generation of its custom AI chips.

Workspace Intelligence: A Unified AI Layer for Work

Workspace Intelligence is described as a system that understands semantic relationships between data in Gmail, Docs, and other Workspace apps, as well as active projects, collaborators, and company-specific information. It creates an intelligence layer across digital workflows including meeting notes, emails, and files.

The system leverages Google's search capabilities and Gemini reasoning for information gathering, situational awareness, and personalization. It powers several new and enhanced features:

  • In Gmail: AI Inbox and AI Overviews.
  • In Google Chat: An "Ask Gemini" feature that can perform tasks like document generation, file search, meeting scheduling, daily briefings, and integration with third-party tools like Asana, Jira, and Salesforce.
  • In Google Docs: Gemini can create infographics, edit multiple images simultaneously, and respond to comments.
  • In Google Slides: It generates slide decks adhering to company templates and visual styles.
  • In Google Sheets: It conversationally builds and edits spreadsheets.

Workspace Intelligence retrieves relevant emails, chats, files, and web information to create drafts that match user voice, brand style, and company templates. Notably, this intelligence system is branded separately from the underlying Gemini functionality.

Next-Generation Tensor Processing Units

Google announced two distinct TPU architectures, one optimized for training AI models and another for running them (inference).

TPU 8t (for Training)

This chip aims to reduce frontier model development cycles from months to weeks. Key specifications and improvements include:

  • Performance: Offers 2.8x better price/performance than the previous generation.
  • Scale: A single TPU 8t superpod scales to 9,600 chips with two petabytes of shared high bandwidth memory, delivering 121 ExaFlops of compute.
  • Speed: Features 10x faster storage access with TPUDirect.
  • Software: Provides near-linear scaling for up to a million chips in a single logical cluster using Virgo Network, JAX, and Pathways software.

TPU 8i (for Inference)

This chip provides 80% better performance-per-dollar than the previous generation, allowing companies to serve nearly twice the customer volume at the same cost, according to Google. Its features include:

  • Memory: Pairs 288 GB of high-bandwidth memory with 384 MB of on-chip SRAM (3x more than the previous generation).
  • Architecture: Uses custom Axion Arm-based CPUs with non-uniform memory architecture (NUMA) for isolation.
  • Interconnect: Doubles Interconnect (ICI) bandwidth to 19.2 Tb/s for efficiently running Mixture of Expert (MoE) models.
  • Network: Features Boardfly architecture, reducing maximum network diameter by more than 50%.
  • Latency: Includes a Collectives Acceleration Engine (CAE) reducing on-chip latency by up to 5x.

Additional Note: The article states that Google Cloud sponsored lodging costs for the event but had no input on editorial coverage.