Google Cloud Next 2026: Agentic AI Highlights and TPU Generation 8

1 min read
google-cloudgemini-enterpriseagentic-aitpuworkspace-intelligencevoice-agentswiz-securityai-infrastructure
View as Markdown
Originally from vm.tiktok.com
View source

My notes

Watch on TikTok Tap to open video

Summary

Recap of Google Cloud Next 2026 keynote highlights showcasing Gemini Enterprise as an end-to-end agentic platform, 8th-gen TPU hardware (TPU-AT for training, separate inference platform), Workspace Intelligence eliminating context fragmentation, and live production agent deployments like YouTube TV’s bilingual voice support. Heavy push on “open AI stack” positioning against walled-garden competitors, with named partner expansion (Accenture, BCG, Deloitte, McKinsey).

Key Insight

  • Gemini Enterprise is positioned as the application layer where business runs, while the agent platform is for technical teams to build/govern. Two distinct surfaces, not one.
  • TPU-AT specs (training-optimized): 121 exaflops FP4 per pod, 9,600 TPUs in 3D torus topology, ~3x compute perf per pod vs Ironwood, 2x interconnect bandwidth, 2 PB shared bandwidth memory per super pod.
  • Native in-MXU quantization removes VPU overhead, meaningful for anyone running cost models on training compute.
  • Axion N48 ARM instances: 2x price/performance, 80% better perf/watt vs comparable x86. Always-on, no cold starts, pitched specifically for agent workloads.
  • Knowledge catalog + Gemini extracts schemas across PDFs to surface cross-document connections (the “base 204 contains soy” allergen demo). The differentiator vs basic RAG: building entity graphs, not just chunking.
  • YouTube TV voice agent is at 100% production traffic for NFL Sunday Ticket / plan support. Demoed mid-call language switch (English to Spanish). Real customer-facing reference, not a sandbox demo.
  • Wiz integration narrative: agentless inventory of code+cloud, security graph flags internet-exposed agents with sensitive data access, auto-suggests fixes routed to dev tools. “Vibe-coded agents from finance” called out explicitly as a threat surface.
  • Workspace Intelligence “context tax” framing: search across Drive/Chat/Docs by meeting context, not keywords. Direct shot at Microsoft Copilot.
  • Open-stack pitch is the strategic wedge, multi-model, multi-cloud, run-where-data-lives. Differentiation against OpenAI/Microsoft lock-in.
  • Google Cloud is now offering NVIDIA Vera Rubin NVL72 among the first hyperscalers, keeping the “any chip” optionality alive even while pushing TPUs.