Google launches Workspace Intelligence semantic layer and TPU 8t/8i chips with 2.8x training performance

Google announced Workspace Intelligence at Cloud Next 2026, a semantic understanding layer that connects data across Gmail, Docs, and other Workspace applications to provide context for Gemini AI features. The company also unveiled its eighth-generation Tensor Processing Units with separate architectures for training and inference.

Workspace Intelligence details

Workspace Intelligence creates what Google describes as a background intelligence layer that understands semantic relationships between emails, documents, meeting notes, and other workspace data. According to Google, the system handles information gathering, situational awareness, and personalization by learning individual work patterns and communication styles.

The layer powers several new features:

• Ask Gemini in Chat: A conversational interface for complex tasks including document generation, file search, meeting scheduling, and daily briefings. Integrates with third-party tools including Asana, Jira, and Salesforce
• Gemini in Docs: Creates infographics grounded in business data, edits multiple images simultaneously, and triages document comments
• Gemini in Slides: Generates presentations that adhere to company templates and visual styles
• Gemini in Sheets: Conversational spreadsheet building and editing

Google explicitly branded this capability as "Workspace Intelligence" rather than positioning it as a Gemini extension, though the company notes end users won't need to be aware of the underlying layer.

TPU 8th generation specifications

Google introduced two distinct TPU architectures this generation:

TPU 8t (training):

• 2.8x better price/performance than previous generation
• Single superpod scales to 9,600 chips
• 2 petabytes of shared high bandwidth memory
• 121 ExaFlops of compute
• 10x faster storage access via TPUDirect
• Near-linear scaling up to 1 million chips in a single logical cluster using Virgo Network

TPU 8i (inference):

• 80% better performance-per-dollar than previous generation
• 288 GB high-bandwidth memory paired with 384 MB on-chip SRAM (3x more than previous generation)
• 19.2 Tb/s Interconnect bandwidth (doubled)
• Powered by custom Axion Arm-based CPUs with doubled physical CPU hosts per server
• Collectives Acceleration Engine reduces on-chip latency by up to 5x
• Boardfly architecture reduces maximum network diameter by over 50%

Pricing for TPU 8t and 8i was not disclosed.

What this means

Google's decision to brand the semantic layer separately from Gemini suggests a strategic shift toward infrastructure-level AI capabilities rather than model-centric positioning. The dual TPU architecture mirrors industry trends of optimizing hardware specifically for training versus serving, with the 2.8x training improvement potentially accelerating Google's ability to iterate on Gemini models.

The Workspace Intelligence integration creates vendor lock-in through contextual understanding that deepens with usage—data moats that competing productivity suites cannot easily replicate. Third-party integrations (Asana, Jira, Salesforce) indicate Google is positioning this as an enterprise workflow hub rather than isolated productivity tools.

The TPU 8i's emphasis on Mixture of Expert (MoE) model optimization aligns with current frontier model architectures, suggesting Google expects MoE to remain dominant in production deployments.