Google Cloud has used its annual Next conference in Las Vegas to pitch businesses on moving from chatting with AI models to running whole teams of AI agents, unveiling a new agent platform, two eighth-generation TPUs and a data cloud built for agentic work. The event runs through Friday.

The centrepiece is the Gemini Enterprise Agent Platform, which Google describes as a full-stack system for building, scaling, governing and optimising agents. It brings together a low-code agent studio, agent-to-agent orchestration so agents can delegate work to each other, a registry for indexing agents across an organisation, agent identity for traceable authorisation, a gateway for real-time policy enforcement and observability dashboards for auditing what agents do.

Agents built on the platform can draw on Google's most advanced model, Gemini 3.1 Pro, as well as models from Anthropic, including Claude Opus 4.7. Google also revealed support for the latest Gemini 3 models in air-gapped and connected environments, aimed at customers in financial services, healthcare and other regulated industries.

Two new TPUs, split by job

In infrastructure, Google launched two versions of its eighth-generation tensor processing unit, each designed for a different job. The TPU 8t is built for training and scales up to 9,600 TPUs with 2 petabytes of shared, high-bandwidth memory in a single superpod. Google says it should deliver three times the processing power of Ironwood, its previous generation, and up to twice the performance per watt.

The TPU 8i is aimed at cost-effective inference, the work of running a trained model. It can directly connect 1,152 TPUs in a single pod and should deliver 80 percent better performance per dollar for inference than the previous generation, which Google says can translate into millions of concurrent agents running cost effectively.

Google Cloud chief executive Thomas Kurian said the split was deliberate. "We felt that people would want systems that were more optimized for training, and separately, systems that were more optimized for inference," he said on an analyst briefing, adding that power efficiency mattered because it would become a constraint as customers scaled.

A data cloud aimed at Databricks

Google also launched its Agentic Data Cloud, a cross-cloud lakehouse standardised on Apache Iceberg that lets customers leave data in Amazon Web Services and query it instantly, with Microsoft Azure interoperability coming later this year. It offers zero-copy access to Databricks, Salesforce, ServiceNow and Snowflake, alongside a knowledge catalog meant to ground agents in trusted business context.

The package includes Gemini-powered notebooks and a deep research agent that works across structured and unstructured data. Google said its Lightning Engine for Apache Spark offers twice the price-performance of Databricks Photon on AWS S3 for large datasets. Enterprises already using the Agentic Data Cloud include Flipkart, Lowe's, American Express and Vodafone.

$750 million to grow the partner network

Alongside the products, Google announced a $750 million fund of resources and incentives for its 120,000-member partner ecosystem, covering AI value assessments, Gemini proofs of concept, agent building and deployment, upskilling and teams of embedded forward-deployed engineers. It also introduced Rapid Enterprise Migration, a service it says can move whole organisations from Microsoft 365 to Google Workspace up to five times faster.

On security, Google said it will combine its Threat Intelligence and Security Operations with the cloud and AI security platform of Wiz, which it bought in March for $32 billion, to create an Agentic Security Operations Center for threat detection, prevention and response.

Google also shared usage figures: about three-quarters of Google Cloud customers now use its AI products, 330 customers each processed more than 1 trillion tokens over the past 12 months, and its first-party models now process more than 16 billion tokens per minute through direct API use, up 60 percent quarter on quarter. Even so, Google said only about 25 percent of organisations have successfully moved AI into production at scale, which it sees as the opportunity ahead.