Google unveils updates to its Agentic Data Cloud: borderless Lakehouse, Knowledge Catalog and databases for agents

Google Cloud has published an announcement detailing what's new in its Agentic Data Cloud, which it presents as a "System of Action" designed to activate all of a company's data, deliver accuracy and trust, and operate proactively and at low cost. According to the company, organizations such as PayPal, Siemens, Telus, Vodafone and Walmart are already achieving business results with the platform. The text is a product announcement from the vendor itself, so the performance figures cited come from Google and not from independent evaluations.

The announcement is organized into four areas: a "borderless" Lakehouse based on Apache Iceberg, a strengthened Knowledge Catalog as a context layer for agents, new experiences for data and business professionals with Gemini, and infrastructure designed to operate at agent scale.

Borderless Lakehouse. Google argues that enterprise data lives across multicloud environments, on-premises systems and SaaS platforms such as Confluent, SAP, Salesforce and Workday, and that the Lakehouse, built on the open Apache Iceberg format, lets companies activate that data without moving it or paying variable egress fees. The company says Iceberg data under management has more than tripled over the past year. Among the new features: Gemini runs directly on the Lakehouse, so business users can query records spread across clouds in natural language; cross-cloud interconnect with caching for AWS and Azure becomes generally available, with a private connection and no variable egress costs, and the Lakehouse stores retrieved remote data blocks locally in Google Cloud; and open Iceberg catalog federation is offered with Databricks Unity, Snowflake Horizon and AWS Glue, along with zero-copy integrations with SAP, Workday and Salesforce. The Lakehouse integrates with Knowledge Catalog to index distributed and unstructured data in cloud storage, in order to give agents context and keep token costs under control.

As a customer case, Delivery Hero talabat says that, through single-copy Iceberg data federation from Amazon S3 tables into BigQuery, it cut data movement costs by roughly 40% and made streaming transactional events queryable across both clouds within minutes of occurring.

AlloyDB and Spanner Omni. To give agents access to both historical analytical data and real-time transactional data, Google is announcing new AlloyDB integrations with the Lakehouse. Direct Iceberg access in AlloyDB allows Iceberg tables in cloud storage to be queried without ETL pipelines or copies, and live operational data to be combined with historical Lakehouse data in a single PostgreSQL query. Tables are registered in the Lakehouse runtime catalog and appear automatically in Knowledge Catalog. Google claims this access is up to 7 times faster on aggregation queries than Amazon's comparable PostgreSQL-compatible service. In addition, Spanner Omni, now generally available, brings "Google-grade" consistency, scale and multi-model capabilities to hybrid agentic applications in on-premises, edge and multicloud environments.

Knowledge Catalog as a context engine. Google describes it as a universal context engine that turns raw data into shared business context for agents. According to its internal benchmarks, it makes agents nearly 4 times more accurate, with up to 33% lower latency and up to 75% fewer tokens. New features include: self-learning enrichment (in preview), which learns from user intent and from human and agent interactions, and extracts information from unstructured sources such as Google Workspace, GitHub, Jira, Confluence and SharePoint; reusable context sets in the Open Knowledge Format (OKF), which can be governed by context engineers and data stewards; and a new studio-style experience (in preview) with a conversational interface, visualization of the data estate across clouds and SaaS applications, and review, auditing and approval of proposed updates with version control and rollback.

In addition, Knowledge Catalog is enabled by default in Gemini (in preview): when connecting BigQuery, AlloyDB, Cloud SQL or Spanner, it resolves business terms, locates trusted tables and determines join paths before generating SQL, while respecting each user's permissions. Google is also announcing general availability of grounding in trusted semantic models from BigQuery, Looker (LookML), Databricks, Snowflake, dbt and SAP, so that agents apply the same governed formulas (for example, net revenue or customer churn) that business teams use. Metadata federation extends to Oracle AI Database@Google Cloud, Snowflake, Databricks Unity Catalog and third-party catalogs such as Atlan and Collibra.

Experiences for professionals. Google wants to move from assistants that support individual contributors to professionals who orchestrate autonomous agents. For data and ML engineers, Gemini adds machine learning skills and tools: the engineer describes the desired outcome in natural language and Gemini generates PySpark code, provides notebooks to edit and test it, trains models, and diagnoses and fixes pipeline issues. For business users, Gemini builds and saves queries with operational reporting skills integrated with BigQuery and Knowledge Catalog; once saved, reports can be run on demand at no token cost. For developers who prefer their own tools, the Data Agent Kit is now generally available and brings Google Data Cloud tools into IDEs and CLIs such as VS Code, Claude Code and Antigravity via native MCP tools.

Agent-scale infrastructure. Google is introducing AlloyDB Express (in preview), an edition aimed at builders and coding agents, with instant provisioning in under 2 seconds, pause and resume with scale-to-zero, secure private IP connectivity, built-in observability and a small footprint of 1–2 vCPUs and 20 GB of storage. For production, "PostgreSQL for agents" in AlloyDB (in preview) lets thousands of agents collaborate by querying live transactional data with full workload isolation; Google cites sub-millisecond I/O, aggregate scan throughput of terabits per second and more than 3 million queries per second, with compute that scales with active reasoning sessions and drops back to zero when idle.

In processing, Managed Service for Apache Spark extends, according to Google, its price-performance advantage over market alternatives to more than 3 times thanks to Lightning Engine improvements, with faster startup on Spark 4.0 at no extra cost and a premium tier combining enhanced hardware and software acceleration, enabled by default. In BigQuery, the company says that over the past year queries have become more than 40% faster and processing costs have fallen by more than 50% thanks to optimizations such as fluid scaling and built-in workload acceleration. BigQuery Graph becomes generally available, offering petabyte-scale graph analytics over relational tables without ETL. Finally, two Google Research models run natively in BigQuery, both in preview: TimesFM 3.0, for generating forecasts directly in SQL without training custom models, and TabFM, which brings predictive machine learning to BigQuery.

Overall, the announcement reflects Google's bet on positioning its data and database offerings as the layer on which enterprise AI agents operate, with an emphasis on open interoperability (Iceberg, MCP, catalog federation), context governance and token cost control. It is worth noting that several features remain in preview and that the performance and accuracy metrics come from Google itself, not from independent evaluations.