Google Cloud transformed its Next 26 conference into a statement of intent: enterprise artificial intelligence is no longer just a tool for assistance but an operational layer capable of executing processes, coordinating agents, querying data, enhancing security, and acting on corporate applications. The company summarizes this with a term that will shape much of its commercial messaging this year: the “agentic enterprise.”
Thomas Kurian, CEO of Google Cloud, made it clear: the experimental phase of generative AI is beginning to give way to a true deployment stage within large organizations. According to data shared by the company, nearly 75% of its cloud customers are already using Google’s AI products, and its proprietary models process over 16 billion tokens per minute through direct API use by clients.
Gemini Enterprise, the new command center for agents
Google Cloud Next 26’s major bet is the Gemini Enterprise Agent Platform, a platform designed to create, deploy, govern, and optimize AI agents within companies. This is not just a rebrand; Google presents it as the evolution of Vertex AI into a broader environment where agents can integrate with internal tools, third-party applications, corporate data, and security systems.
The platform includes features such as Agent Studio, Agent-to-Agent Orchestration, Agent Registry, Agent Identity, Agent Gateway, and Agent Observability. Collectively, these components aim to solve one of the main challenges companies face: how to prevent each department from creating its own uncontrolled, untraceable agents without common security standards.
Google intends for Gemini Enterprise to be the gateway to AI for employees and customers. The app incorporates Agent Designer, long-duration agents, an inbox to manage activity, projects with contextual memory, Skills for repetitive tasks, and Canvas to create and edit documents and presentations without leaving the environment.
The approach is ambitious. An agent is no longer seen as a simple chatbot but as a system capable of following instructions, consulting internal sources, executing tasks, requesting human approval when needed, and maintaining context over long processes. This concept can be useful in areas such as customer service, operations, finance, marketing, internal support, data analysis, or document management.
However, the risk lies in complexity. The more agents operate within a company, the greater the need for controls, auditing, and clear boundaries. That’s why Google places strong emphasis on identity, observability, control gateways, and anomaly detection. The question is no longer just if an agent responds correctly but whether it acts within the organization’s authorized parameters.
| Announced Area | What It Offers | Why It Matters |
|---|---|---|
| Gemini Enterprise Agent Platform | Development, orchestration, governance, and observability of agents | Moving from isolated pilots to controlled enterprise deployments |
| Gemini Enterprise app | Agent Designer, Inbox, Skills, Projects, and Canvas | Brings agent creation and usage closer to non-technical employees |
| TPU 8t and TPU 8i | New chips for training and inference | Strengthen Google’s own infrastructure amidst large-scale AI pressure |
| Agentic Data Cloud | Multicloud Lakehouse, Knowledge Catalog, and Data Agent Kit | Connecting agents with reliable data and business context |
| Agentic Defense | Threat intelligence, SecOps, and Wiz integration | Responding to new risks in agents, cloud, data, and AI applications |
New TPUs for training and inference
The infrastructure was another major focus of the event. Google introduced its eighth generation of TPUs, with two distinct lines: TPU 8t, focused on training, and TPU 8i, designed for low-latency inference and large-scale agents.
According to Google, TPU 8t can scale up to 9,600 TPUs and 2 PB of high-bandwidth shared memory within a single superpod. The company claims it delivers three times the processing power of Ironwood and up to double the performance per watt. TPU 8i, on the other hand, is optimized for inference, utilizing a new topology called Boardfly, with direct connection of 1,152 TPUs in a pod. Google states this variant offers 80% more performance per dollar in inference compared to the previous generation.
These figures should be read as manufacturer claims but indicate a clear trend: AI infrastructure is becoming more specialized. Powerful accelerators are no longer sufficient; inference, reasoning, and agent execution require low latency, efficient memory, fast networks, and controlled costs.
Google also reinforced its message of openness by recalling that its AI Hypercomputer combines TPUs, Axion CPUs, and NVIDIA GPUs. The company assured that it will be among the first providers to offer NVIDIA Vera Rubin NVL72, alongside instances based on Blackwell and Hopper, already available or announced in its portfolio.
In storage, Google highlighted Managed Lustre with up to 10 TB/s performance to A5X or TPU 8t via RDMA, and improvements in Rapid Storage, which increased from 6 TB/s to 15 TB/s. In networking, it introduced Virgo Networking, an AI-optimized infrastructure aimed at connecting Vera Rubin NVL72 systems or superpods of TPU 8t in supercomputers with hundreds of thousands of accelerators.
Data, security, and productivity: AI wants to be in every part of the business
The third pillar of the announcement is Agentic Data Cloud, an architecture designed to enable agents to work with dispersed corporate data across different systems. Google announced a multicloud Lakehouse based on Apache Iceberg, Knowledge Catalog to contextualize corporate data, Data Agent Kit for Gemini-assisted data science, and Deep Research Agent for deeper business analysis.
The goal is clear: if agents are to make decisions or perform tasks, they need reliable, up-to-date data with business context. Connecting a model to a document repository is not enough. AI must understand relationships, permissions, context, metadata, and internal policies. This is where Google aims to differentiate itself with a semantic layer over both structured and unstructured data.
In security, Google introduced Agentic Defense, a concept that integrates Google Threat Intelligence, Security Operations, and Wiz platform. The company also announced specialized agents for threat detection, detection engineering, investigation, and remediation. The message aligns with the broader deployment of agents: as companies automate critical processes, they must also automate aspects of security.
Security concerning agents will be a major debate in the coming years. Risks such as instruction injection, unauthorized data access, tool misuse, or reasoning errors could have more severe consequences when agents are not just responsive but also act. That’s why Google emphasizes Agent Identity, Agent Gateway, Model Armor, anomaly detection, and security dashboards.
Productivity also gained importance with Workspace Intelligence, a semantic layer for Gmail, Docs, Drive, Meet, Chat, Sheets, and Slides. Google wants Workspace to evolve from a suite of applications into an environment with shared context, capable of summarizing information, preparing documents, organizing projects, and executing tasks via Gemini Enterprise.
A direct battle for the new enterprise cloud
Google Cloud Next 26 confirms that competition in the cloud has entered a new phase. AWS, Microsoft, and Google are no longer only competing over VMs, storage, or databases. They are vying to become the platform where companies build agents, connect data, run models, and automate entire processes.
For Google, the opportunity is twofold: leverage its historical strengths in models, data, search, productivity, and proprietary infrastructure; and continue expanding its market share in a cloud landscape still dominated by strong positions from Microsoft Azure and AWS.
While Google’s proposal appears technically robust on paper, it also raises important questions for clients. Companies will need to evaluate actual costs, platform dependence, integration with existing systems, agent portability, regulatory compliance, data sovereignty, and maturity of governance tools.
The shift toward the agentic enterprise will not be immediate or uniform. Some sectors like banking, retail, telecom, security, healthcare, and customer service are likely to adopt these systems quickly. Others will need more time due to regulatory, cultural, or technical reasons. Still, the market trend is clear: enterprise AI is moving from individual assistance to coordinated process automation.
Google Cloud has laid out a comprehensive architecture for this transition. The next step is to prove it can turn this compelling narrative into real, secure, measurable, and profitable deployments for companies that demand operational results rather than isolated experiments.
Frequently Asked Questions
What is Gemini Enterprise Agent Platform?
Gemini Enterprise Agent Platform is Google Cloud’s new platform to create, deploy, govern, and optimize AI agents within enterprises. It combines capabilities from Vertex AI with new features for orchestration, security, observability, identity, and agent management.
What is the difference between TPU 8t and TPU 8i?
TPU 8t is geared towards large-scale AI model training, while TPU 8i is designed for inference, low latency, and large-scale agent deployment. Google presents them as specialized architectures for different stages of the AI lifecycle.
What does Agentic Data Cloud offer?
Agentic Data Cloud aims to connect AI agents with reliable, contextualized enterprise data across various environments. It includes a multicloud Lakehouse based on Apache Iceberg, Knowledge Catalog, Data Agent Kit, and Gemini-assisted analysis capabilities.
Why does Google speak of a “agentic enterprise”?
Google uses the term “agentic enterprise” to describe organizations where AI agents not only answer questions but also execute processes, coordinate tasks, work with internal data, and act according to corporate policies on security and governance.
via: cloud.google