Google Cloud Next '26: TPU 8th Gen, Gemini 3.1, and the Agentic Enterprise Platform — What Was Actually Announced
Google Cloud Next '26 ran from April 22 to 24, 2026, in Las Vegas. CEO Sundar Pichai delivered the keynote. The announcements spanned custom silicon, frontier AI models, an enterprise agent platform, a cross-cloud data architecture, and a major security acquisition integration. This article covers what was confirmed at the event and in official Google blog posts.
8th Generation TPUs: Two Purpose-Built Chips
Google's most significant infrastructure announcement was the 8th generation of its Tensor Processing Unit, split into two distinct chips designed for different AI workloads: TPU 8t (training) and TPU 8i (inference).
TPU 8t — Built for Training Scale
The TPU 8t is designed for large-scale model training. Key confirmed specifications:
- 9,600 chips networked in a single superpod
- 2 petabytes of shared high-bandwidth memory across the superpod
- 121 ExaFlops per pod
- Inter-chip interconnect (ICI) bandwidth: 2× the prior generation (Ironwood)
- Storage access: 10× faster than prior generation
- Compute performance: approximately 3× per pod versus Ironwood
- Performance-per-watt: up to 2× better than Ironwood
- Target goodput: over 97%
- Near-linear scaling to 1 million chips in a single logical cluster
The TPU 8t uses Google's TPUDirect technology to pull training data directly into the chip, Optical Circuit Switching (OCS) for automatic failure rerouting, and real-time telemetry across tens of thousands of chips simultaneously. It is paired with Google's Virgo Network fabric.
TPU 8i — Built for Inference
The TPU 8i is purpose-built for the low-latency requirements of agentic AI workloads — systems that reason across multiple steps and make sequential decisions. Confirmed specifications:
- 288 GB HBM per chip
- 384 MB on-chip SRAM — 3× more than Ironwood — allowing larger key-value caches to remain entirely on-silicon
- ICI bandwidth: 19.2 Tb/s (doubled versus prior generation)
- 80% better performance-per-dollar for inference versus Ironwood
- 1,152 TPUs per pod using the new Boardfly hierarchical topology
- Network diameter reduced by over 50% through the Boardfly architecture
- On-chip latency reduced by up to 5× through the Collectives Acceleration Engine (CAE)
- Performance-per-watt: up to 2× better than Ironwood
Both chips run Google Axion ARM-based CPU hosts and use fourth-generation liquid cooling. Both support JAX, MaxText, PyTorch, SGLang, and vLLM. Both are co-designed with Google DeepMind.
Gemini 3.1 and the Model Family
Google announced Gemini 3.1 Pro and Gemini 3.1 Flash as generally available. The Gemini 3.1 family includes enhanced reasoning for multi-step tasks and is available on Google Cloud for enterprise workloads.
The broader model announcements at the event included Veo 3.1 Lite for video generation and Lyria 3 Pro for music generation, both part of the expanded Gemini family.
Google reported that its first-party models now process more than 16 billion tokens per minute via direct API use by customers — up from 10 billion tokens per minute the prior quarter.
Gemini Enterprise Agent Platform
Google launched the Gemini Enterprise Agent Platform — a unified platform for building, deploying, governing, and optimising AI agents at enterprise scale. It combines Vertex AI capabilities with new orchestration and governance tooling.
New features confirmed at the event:
- Agent Designer: builds schedule-triggered or event-triggered agents with no-code configuration
- Long-running agents: executes complex business processes that span hours or days without human intervention
- Inbox: centralised view of agent activity across the organisation
- Skills: shortcuts for repetitive task sequences that agents can invoke
- Canvas: document creation and editing within the agent interface, without switching applications
The platform also includes agent identity management — each agent has a distinct identity for audit trails — and observability tooling for monitoring agent behaviour across deployments.
Agentic Data Cloud: Zero-Copy Access Across AWS and Azure
Google introduced the Agentic Data Cloud, a new AI-native data architecture built around a cross-cloud lakehouse. The confirmed components:
- Cross-cloud lakehouse with zero-copy access to data stored in AWS and Microsoft Azure — organisations do not need to migrate data into Google Cloud to query or analyse it
- Knowledge Catalog for grounding AI agents in enterprise-wide semantic context — agents can access structured metadata about what data exists, where it lives, and what it means
- Deep Research Agent for autonomous intelligence gathering and synthesis across enterprise data sources
The cross-cloud lakehouse capability is a direct response to the multi-cloud reality most enterprises operate in. 94% of enterprises now use multiple cloud providers; requiring data migration as a prerequisite to AI analysis has been a significant adoption barrier.
Security: Wiz Integration and Agentic SecOps
Google completed its $32 billion acquisition of Wiz, the cloud-native application protection platform, and announced the first integrations at Cloud Next.
New agentic security tools confirmed:
- Dark Web Intelligence Agent: uses Gemini to build a real-time security profile from dark web activity — monitoring for credential exposure, threat actor mentions, and data leakage
- Threat Hunting Agent: proactively searches for novel attack patterns that signature-based detection systems would miss
These tools are part of Google's broader Agentic SecOps framework, which applies AI agents to security operations workflows — alert triage, threat investigation, and response — that currently require significant human analyst time.
Hardware: NVIDIA Vera Rubin NVL72 on Google Cloud
Google confirmed it will be among the first hyperscalers to offer NVIDIA Vera Rubin NVL72 systems, integrated into the Google AI Hypercomputer architecture. The Vera Rubin NVL72 — 72 Rubin GPUs in a liquid-cooled rack-scale system — is expected in H2 2026.
Google also announced a chip partnership with Marvell Technology to co-develop a custom media processing unit and inference-optimised TPU variant. Marvell becomes Google's third chip partner after Broadcom and MediaTek.
Capital Commitment
Sundar Pichai reaffirmed Google's $175–185 billion capital expenditure plan for 2026, with approximately half directed toward Cloud infrastructure. This follows the AI infrastructure investment announcements from Microsoft ($190 billion capex) and Amazon (approximately $200 billion capex) for the same period — a combined $565–575 billion in hyperscaler infrastructure investment in a single year.
What This Means for Philippine Google Cloud and Workspace Users
For Philippine businesses already on Google Workspace or Google Cloud, the Cloud Next announcements translate to a near-term set of capabilities:
Gemini 3.1 is available now in Workspace Business and Enterprise plans. The enhanced reasoning capability is most visible in tasks like document analysis, meeting summary, and multi-step research — where Gemini 3.1's improvements over prior versions are measurable.
The Gemini Enterprise Agent Platform requires Workspace Enterprise or Google Cloud setup. For organisations wanting to build custom agents — automating procurement workflows, customer query handling, or internal knowledge retrieval — the Agent Designer is the starting point.
The Agentic Data Cloud's cross-cloud lakehouse is most relevant to organisations with data in multiple environments. Philippine businesses running a hybrid of Google Cloud and Azure, or using Azure for M365 and Google Cloud for other workloads, can now analyse cross-environment data without migration.
Wiz-backed security tooling will roll out through Google Cloud Security Command Center. For Philippine businesses in regulated industries — banking, healthcare, BPO — the Agentic SecOps capability provides an AI-assisted layer over security monitoring that previously required dedicated analyst staffing.
For Philippine organisations evaluating or expanding their Google Cloud or Workspace deployment, get in touch.
Talk to our Cloud & I.T. team →
