At its annual Cloud Next conference in Las Vegas (April 22–24, 2026), Google Cloud went all‑in on autonomous agents. CEO Thomas Kurian titled the opening keynote “The Agentic Cloud” and unveiled a sweeping reorganization of its AI stack. The most significant move: retiring Vertex AI as a standalone brand and folding it into the Gemini Enterprise Agent Platform — a new platform for building, scaling, governing and optimizing AI agents. Alongside the platform shift, Google announced two eighth‑generation TPUs (8t for training, 8i for inference), made its seventh‑gen Ironwood TPU generally available, introduced Virgo interconnect fabric, and unveiled new security agents, Workspace Studio (no‑code agent builder), and a production‑grade A2A (Agent‑to‑Agent) protocol.
With over 32,000 attendees, 700+ breakout sessions, 490+ sponsors and 260 announcements, Next ’26 marked Google’s most aggressive push yet to own the full stack — from silicon to employee inbox. This review covers the biggest announcements, the strategic bets behind them, and what they mean for enterprises, developers and AI builders.
Big Picture: Why “Agentic Enterprise” Changes Everything
Google’s central thesis is that isolated AI tools are no longer enough. Instead, companies need an environment where autonomous agents work alongside employees — handling multi‑step tasks, interacting across systems, and maintaining context over days.
Key driver: 75% of Google Cloud customers already use AI in their businesses. But managing agents at scale requires a new foundation: secure, connected to internal systems, and optimized for performance, cost, and governance.
Google’s differentiating bet: Vertical integration. Unlike competitors that hand customers separate components (models, compute, storage, security tools) and expect them to integrate everything, Google offers a unified stack where everything is co‑developed. Gemini Enterprise Agent Platform, TPUs, Gemini Enterprise app, Workspace Studio, and the new security layer all work together out of the box.
Competitive context:
- Google Cloud exited Q4 2025 with the fastest growth rate among the big three cloud providers: 50% year‑on‑year (compared to AWS and Azure)
- OpenAI’s enterprise revenue now accounts for 40% of its total; OpenAI’s Operator browser agent scores 87% on complex benchmarks
- Anthropic has launched an enterprise marketplace, and its Model Context Protocol has reached 10,000 servers with 97 million monthly SDK downloads
Google is fighting from third place in cloud market share, but its vertical integration (model, runtime, silicon, Workspace distribution) is a competitive advantage neither OpenAI nor Anthropic can replicate.
The Core Announcements At A Glance
| Category | Key Announcement | Status / Availability |
|---|---|---|
| Agent Platform | Gemini Enterprise Agent Platform (evolves Vertex AI) | Generally Available |
| Agent Building (Low‑Code) | Workspace Studio build agents in Gmail, Docs, Sheets | Rolling out to business/enterprise |
| Agent Building (Code‑First) | Agent Development Kit (ADK) v1.0, graph‑based orchestration | Generally Available |
| Agent Runtime | Sub‑second cold starts, long‑running agents (multi‑day) | Generally Available |
| Agent Memory | Memory Bank for persistent context across sessions | Generally Available |
| Agent Governance | Agent Identity (cryptographic IDs), Agent Registry, Agent Gateway | Preview / Available |
| Pre‑Built Agents | Agent Garden — templates for customer service, data analysis, coding | Available |
| TPU (7th Gen) | Ironwood — 4.6 petaFLOPS per chip, 9,216‑chip superpod | Generally Available (April 2026) |
| TPU (8th Gen) | TPU 8t (training, 9,600‑chip pod) + TPU 8i (inference, 288GB HBM) | Preview (available late 2026 / 2027) |
| Networking | Virgo interconnect (47 petabits/sec, links 134k TPUs) | Announced |
| Agentic Data Cloud | Cross‑cloud Lakehouse, Knowledge Catalog | Available |
| Agentic Defense | 3 new Security Operations agents + Wiz deep integration | Preview / General Availability |
| Agentic Taskforce | Gemini Enterprise for CX + Workspace updates | Rolling out |
| A2A Protocol | Agent‑to‑Agent communication (v1.0) | In production at 150+ organizations |
Gemini Enterprise Agent Platform: The Death Of Vertex AI
Vertex AI, launched in 2021, was Google Cloud’s managed platform for training, tuning, and deploying AI models. At Next ’26, Google rebranded and expanded it into the Gemini Enterprise Agent Platform — a comprehensive platform for building, scaling, governing, and optimizing agents.
What stays the same: Existing Vertex AI APIs remain backward compatible. Organizations that already built on Vertex AI can access the new brand directly inside their Google Cloud console without migration.
What’s new: To use the new agent capabilities (Agent Runtime, Memory Bank, Agent Registry, etc.), enterprises need to enable the Agent Platform plan. The platform is organized around four pillars: Build, Scale, Govern, Optimize.
Build: Two Paths To Agent Development
Agent Studio (low‑code visual canvas) is generally available. It allows users to design agent reasoning loops, set triggers, and build schedule‑based automation without writing code.
Agent Development Kit (ADK) is a modular, model‑agnostic framework that now includes a graph‑based orchestration layer for defining structured logic across multiple cooperating agents. Google reports that more than 6 trillion tokens are processed monthly through ADK. The framework supports Python, Java, and Go, and is compatible with third‑party orchestration tools.
Agent Garden supplements both paths with prebuilt templates for use cases including financial analysis, code modernization, and invoice processing.
Scale: Agents That Run For Days
Earlier AI agents were built around short, discrete interactions. Agent Platform’s scale layer supports a different model: agents that persist, accumulate context, and execute tasks over extended periods.
- Agent Runtime: delivers sub‑second cold starts and supports long‑running agents that maintain state for multiple days
- Memory Bank: provides persistent context storage across sessions, allowing agents to carry forward information from previous interactions without requiring users to restate it each time
- Already in production at several large enterprises — GE Appliances is running more than 800 agents on the platform
Govern: Identity, Registry & Access Controls
Governance is a core layer of the new platform. As agent deployments scale, security and compliance become critical.
- Agent Identity: each agent receives a verifiable cryptographic ID, creating an auditable record of actions tied to defined access policies
- Agent Registry: centralized catalog of approved agents, tools, and capabilities
- Agent Gateway: enforces policy on agent‑to‑agent and agent‑to‑tool connections, with awareness of MCP and the A2A protocol
- Model Armor: runtime protection for model and agent interactions, integrating with Agent Gateway and Agent Runtime
Optimize: Evaluation & Observability
A new security dashboard maps relationships between agents and underlying infrastructure, helping teams identify vulnerabilities and maintain oversight as deployments scale. Monitoring features include models that flag unusual agent behavior and tools that surface potential threats such as unauthorized access or suspicious network activity.
Model Garden: Access To 200+ Models
The platform gives access to over 200 leading AI models through Model Garden, including:
- Google’s own: Gemini 3.1 Pro, Gemini 3.1 Flash Image (Nano Banana 2), Lyria 3, Gemma 4
- Third‑party models: Anthropic’s Claude Opus, Sonnet, and Haiku
- Open models: Llama and others
Google’s inclusion of competing models (including Claude) reflects a deliberate positioning: bet that enterprises will choose the platform for its infrastructure and governance capabilities, not solely for its own models.
Workspace Studio: No‑Code Agents Inside Gmail, Docs & Sheets
The most consumer‑facing announcement at Next ’26 is Workspace Studio — a no‑code platform that lets business users build and deploy AI agents directly inside Gmail, Docs, Sheets, Drive, Meet, and Chat by describing automations in plain language.
Example: A user can type “every Friday, ping me to update my tracker” and Gemini creates the automation.
Integrations: Workspace Studio connects to third‑party applications including Asana, Jira, Mailchimp, and Salesforce, and can call external APIs via webhooks or run custom logic through Apps Script.
Availability: Rolling out to Google Workspace business, enterprise, and education customers.
Gemini Enterprise App: AI Front Door For Every Employee
Complementing the developer‑focused Agent Platform, the Gemini Enterprise app brings the best of Google AI to every employee. It includes:
- Agent Designer: low‑code tool for building agents within daily workflow
- Inbox: manage agent activity and notifications across all deployed agents
- Skills, Projects: organize and reuse agent capabilities
- Long‑running agents: agents that maintain state over multiple days
The Gemini Enterprise app serves as the “front door for AI for every employee,” integrating agents directly into how teams work.
TPU v8: Splitting Training And Inference
Google made its seventh‑generation TPU, Ironwood, generally available to cloud customers at Next ’26. Ironwood delivers 4.6 petaFLOPS of peak FP8 compute per chip — roughly four times the performance of its predecessor Trillium — with 192 gigabytes of HBM3e memory and 7.37 terabytes per second of memory bandwidth. A single Ironwood superpod links 9,216 chips into a unified system delivering 42.5 exaFLOPS of compute, more than 24 times the capacity of El Capitan, currently the world’s most powerful supercomputer.
Ironwood positions Google directly against Nvidia’s Blackwell B200. Both deliver roughly 4.5–4.6 petaFLOPS and 192GB of HBM. Google’s advantage is at cluster scale, energy efficiency (roughly twice the performance per watt of Trillium and 2.8 times that of Nvidia’s H100), and the economics of inference on custom silicon designed specifically for the task.
TPU 8t (Training) And TPU 8i (Inference)
For the first time, Google split its flagship TPU into two specialized chips, moving away from a one‑size‑fits-all model.
TPU 8t offers nearly three times the compute performance compared to Ironwood, with up to twice the performance per watt. It introduces native FP4 support, which helps overcome memory bandwidth bottlenecks and doubles MXU throughput while maintaining accuracy for large models. Additionally, SparseCore, a new accelerator, complements the existing Matrix Multiply Unit to handle irregular memory access patterns of embedding lookups, keeping the chip busy rather than stalling. Both chips use fourth‑generation liquid cooling to support higher compute density while controlling power usage.
TPU 8i prioritizes low latency for AI agents, which require systems that can reason through tasks, run workflows, and repeatedly interact with tools. Its 384 MB of on‑chip SRAM allows an AI model’s active working set to reside entirely on the chip, significantly reducing response times for complex multi‑agent tasks.
Google claims an 80% improvement in price‑performance for the 8i and up to a 2.8x gain for the 8t, allowing enterprises to serve nearly twice the customers at the same cost. Both TPU 8t and 8i will be available through Google Cloud later this year (preview), supporting frameworks including JAX, PyTorch, SGLang, and vLLM.
Virgo Interconnect: 47 Petabits Per Second
Google introduced Virgo, a new “megascale” scale‑out fabric that can link up to 134,000 TPUs in different data centers at speeds up to 47 petabits per second of bi‑directional bandwidth. This massive interconnect capability enables AI clusters to scale beyond what any single data center can house, making distributed training and inference across geographically separated facilities practical.
A2A Protocol v1.0: Agent‑To‑Agent Communication
The Agent‑to‑Agent (A2A) protocol v1.0 is now in production at over 150 organizations. A2A enables agents built on different platforms (Google, AWS, Salesforce, etc.) to communicate and collaborate directly, breaking down silos and enabling multi‑agent systems that span the enterprise technology stack. Managed MCP (Model Context Protocol) servers are also available across Google Cloud services, with Apigee acting as an API‑to‑agent bridge.
Project Mariner: Web‑Browsing Agent
Project Mariner is a new web‑browsing agent capable of navigating websites, filling forms, and performing multi‑step online tasks autonomously. It represents Google’s entry into the browser‑agent space — competing directly with tools like OpenAI’s Operator. Project Mariner can handle tasks such as booking appointments, filling out applications, and retrieving information from complex web interfaces without human supervision.
Agentic Data Cloud & Agentic Defense
Agentic Data Cloud introduces a new, AI‑native architecture including a cross‑cloud Lakehouse and Knowledge Catalog, allowing organizations to use their data at the speed and scale required by agentic AI. This architecture is designed to work seamlessly with the Gemini Enterprise Agent Platform, ensuring that agents have access to the data they need without performance bottlenecks.
Agentic Defense combines Google’s Threat Intelligence and Security Operations with Wiz’s Cloud and AI Security Platform to prevent, detect, and respond to threats. Key components include:
Customer Adoption & Real‑World Use Cases
Google highlighted several customers already running agents in production:
Pricing & Availability Summary
| Product | Status | Availability / Pricing Notes |
|---|---|---|
| Gemini Enterprise Agent Platform | Generally Available | Access via existing Vertex AI console; enables new agent capabilities |
| Gemini Enterprise app (for employees) | Generally Available | Included with Google Workspace business/enterprise plans |
| Workspace Studio | Rolling out | Included with Google Workspace business/enterprise/education |
| Ironwood TPU (7th gen) | Generally Available | Access via Google Cloud (April 2026) |
| TPU 8t / TPU 8i (8th gen) | Preview (code names Sunfish / Zebrafish) | Available late 2026 / 2027 |
| Virgo interconnect | Announced | Availability not specified |
| Project Mariner | Preview | — |
| A2A protocol v1.0 | In production | Available at 150+ organizations |
| Agentic Defense (Security Operations agents) | Preview | — |
| Google Cloud Fraud Defense | Generally Available | — |
Note: TPU 8t and 8i are built on TSMC’s 2nm process (advanced node) and are expected to enter production in late 2026 / early 2027.
Pros And Cons Of Google Cloud Next ’26 Announcements
| Pros ✅ | Cons ❌ |
|---|---|
| Vertically integrated stack (silicon → platform → app) eliminates integration headaches for enterprises | TPU 8t and 8i are still in preview — not available for production workloads until late 2026/2027 |
| 200+ models in Model Garden (including Anthropic Claude) gives customers choice, not lock‑in | A2A protocol adoption is early — cross‑platform agent interoperability will take time to mature |
| Sub‑second cold starts and multi‑day persistent agents enable entirely new classes of applications (sales prospecting, approval workflows) | Project Mariner web agent is still preview — reliability and safety benchmarks not yet public |
| Ironwood TPU generally available now, competing directly with NVIDIA Blackwell on raw specs | Migrating from existing Vertex AI workloads to Agent Platform requires enabling new plan (though APIs are backward compatible) |
| 80% better inference price‑performance on TPU 8i — major cost advantage for high‑scale AI deployments | Limited regional availability for new TPU instances initially |
| Agentic Defense with 60‑second security alert triage (vs 30 minutes manual) addresses critical enterprise pain point | Wiz integration still rolling out — not yet available in all Google Cloud regions |
| Workspace Studio opens agent building to non‑developers (business users) — true democratization of AI | Learning curve for Agent Studio (low‑code) and ADK (code‑first) — requires training |
| Over 32,000 attendees and 260 announcements confirm Google Cloud is a major force in enterprise AI | With over 260 announcements, the sheer volume can overwhelm customers trying to prioritize |
Google Cloud Next ’26 Vs Competitors
| Aspect | Google Cloud | Microsoft Azure | AWS | OpenAI / Anthropic |
|---|---|---|---|---|
| AI chip strategy | TPU (in‑house), split training/inference (8t/8i) | Maia (in‑house), Nvidia partnership | Trainium, Inferentia (in‑house) | Reliant on Nvidia / cloud providers |
| Agent platform | Gemini Enterprise Agent Platform | Azure AI Agent Service (Copilot Studio) | Bedrock Agents, SageMaker | Operator (OpenAI), Claude Code (Anthropic) |
| 100% vertical integration | Silicon → platform → app → Workspace | Azure + OpenAI (partnership) | AWS + third‑party models | Model‑only — rely on cloud providers |
| Inference economics | 80% better price‑performance (TPU 8i claim) | Azure Maia — limited public data | Inferentia — cost‑optimized | Dependent on Nvidia — higher cost |
| Enterprise distribution | Workspace (Gmail, Docs, Sheets, Meet) | Microsoft 365 (Teams, Office) | No native productivity suite | No native distribution |
| Agent‑to‑agent protocol | A2A v1.0 (production at 150 orgs) | In development | Not announced | Competitive (not cooperative) |
| Security agents | 3 new SecOps agents + Wiz deep integration | Microsoft Security Copilot | Amazon Security Lake | Limited |
Frequently Asked Questions (FAQs)
Is Vertex AI going away?
Vertex AI as a standalone brand is being retired and replaced by the Gemini Enterprise Agent Platform. However, existing Vertex AI APIs remain backward compatible. Organizations that built on Vertex AI can access the new platform directly without migration. New agent capabilities (Agent Runtime, Memory Bank, Agent Registry) require enabling the Agent Platform plan.
When will TPU 8t and 8i be available?
Both TPU 8t and 8i are currently in preview. They are built on TSMC’s 2nm process and are expected to enter production in late 2026 or early 2027, with general availability through Google Cloud following.
What is the A2A protocol, and why does it matter?
The Agent‑to‑Agent (A2A) protocol v1.0 enables AI agents built on different platforms (Google, AWS, Salesforce, etc.) to communicate and collaborate directly. This avoids vendor lock‑in and allows enterprises to build multi‑agent systems that span their entire technology stack. A2A is already in production at over 150 organizations.
How does Workspace Studio differ from the Agent Development Kit?
Workspace Studio is a no‑code platform for business users, accessible inside Gmail, Docs, Sheets, and other Workspace apps. Users describe automations in plain language (“every Friday, ping me to update my tracker”) and Gemini creates the automation. The Agent Development Kit (ADK) is a full‑code framework for developers building complex, multi‑agent systems.
What is the “Agentic Enterprise”?
“Agentic Enterprise” is Google Cloud’s term for an organization where AI agents work alongside human employees, automating complex workflows, interacting across internal systems, and adapting in real time. The term reflects the shift from generative AI (chatbots, content creation) to autonomous agents that can plan, execute, and learn.
Is Google winning against AWS and Microsoft?
In absolute market share, AWS remains #1 and Azure #2. However, Google Cloud exited Q4 2025 with the fastest growth rate among the three at 50% year‑on‑year. Next ’26 demonstrates Google’s aggressive bet on vertical integration: owning the model, the runtime, the silicon, and the distribution channel through Workspace — an advantage neither competitor can easily replicate.
