Google Cloud Next ’25: Unpacking Data & AI’s Innovative Edge

In April 2025, Google Cloud Next hosted over 32,000 attendees, marking its largest customer event to date. This year’s conference reflected a significant inflection point for Google Cloud, demonstrating impressive growth across its full-stack product portfolio, customer base, and partner ecosystem. The 229 announcements underscored the breadth and depth of Google Cloud’s advancements.

From the financial side, Google Cloud recorded its first profitable quarter in Q1 2023, achieving an operating income of $191 million. Last week, the company reported first-quarter Google Cloud revenue of $12.3 billion, demonstrating significant growth in a short amount of time.

Having had the privilege of attending Google Cloud Next ’25 and immersing myself in the latest innovations, I’ve analyzed Google Cloud’s overarching strategy and documented the key announcements in this blog post. To provide a structured understanding, I’ll be using my own classification of these announcements, visually represented in Figure 1.

Figure 1: Key Google Cloud Next ’25 Data and AI Announcements

Let’s start with the lowest layer of the stack, the infrastructure layer.

Infrastructure

Google Cloud offers a comprehensive full-stack platform that spans from custom hardware to end-user applications. In this section, we focus on the latest advancements in its infrastructure layer, including innovations in custom silicon, AI model deployment, network architecture, and security frameworks. These updates reflect Google Cloud’s continued investment in optimizing performance, scalability, and trust at every level of the stack.

TPUs

The backbone of training and serving foundation models is the GPU. At Google, that role is filled by its custom Tensor Processing Unit (TPU). At Next ’25, Google Cloud introduced its 7th-generation TPU, called Ironwood — a purpose-built, energy-efficient chip optimized for faster inference. Compared to its predecessor, TPU v5 (Trillium), Ironwood delivers 5x the peak compute performance, 6x the high-bandwidth memory (HBM) capacity, and is twice as power efficient.

Ironwood is a key component of the AI Hypercomputer, Google Cloud’s workload optimized infrastructure helps accelerate AI workloads at scale. This integrated supercomputing system is designed for efficiently training, fine-tuning, and serving AI workloads. It comprises performance-optimized hardware such as GPUs/TPUs, networking and storage, software (JAX, PyTorch, vLLM, XLA, Keras), and flexible consumption like Dynamic Workload Scheduler.

Figure 2 highlights the latest enhancements across the AI Hypercomputer stack, including improvements in compute, memory, interconnect, and AI model orchestration.

Figure 2: Google Cloud’s AI Hypercomputer

Hyperdisk Exapools is a new block storage optimized for large training jobs. Rapid Storage is a new zonal Cloud Storage bucket offering sub-1ms latency and high throughput for data-intensive AI applications. New inference capabilities in Google Kubernetes Engine (GKE) designed to reduce serving costs, decrease latency, and increase throughput for AI workloads.

Jensen Huang, the charismatic CEO of NVIDIA, made his obligatory appearance to herald the arrival of their groundbreaking Blackwell architecture within the Google Cloud ecosystem. Blackwell, the successor to the formidable Hopper architecture, promises a significant leap in processing power, especially for trillion-parameter scale AI models. In addition, Google Cloud announced the availability of A4 and A4X Virtual Machines that are powered by NVIDIA’s Blackwell B200 and GB200 GPUs. The A4 VM specifically features eight Blackwell GPUs interconnected via fifth-generation NVIDIA NVLink.

Google Cloud introduced its first Arm-based CPU called Axion in 2024. This year, two operational databases, Cloud SQL and AlloyDB were made available on the corresponding C4A instances with 50% higher price-performance.

While security isn’t my primary area of expertise, the significant $32 billion acquisition of Wiz has undeniably injected fresh energy and attention into the cloud security landscape. Wiz maintained its independent presence with its own booth at the event, due to the acquisition still awaiting regulatory clearance. Once finalized, this strategic move is expected to significantly bolster Google Cloud’s hybrid and multi-cloud security offerings. Adding another layer to this dynamic is the fact that Palo Alto Networks, a major player in the cybersecurity space, is a prominent Google Cloud customer.

Models

The narrative surrounding Google’s position in the fiercely competitive race for AI supremacy has taken a dramatic turn. While there was a sentiment in the industry that the very originators of the groundbreaking Transformer architecture had, perhaps ironically, ceded ground in the pursuit of the most performant AI models, the arrival of the Gemini 2.5 family has emphatically rewritten that script. They have outperformed almost the entire field across a comprehensive suite of rigorous benchmarks.

Some of the standout features for Gemini 2.5 models include:

• Enhanced Reasoning Capabilities
• Large context window of 1M tokens (2M coming soon)
• Multimodal (text, images, audio, video, code repositories)
• Multi-lingual
• Live APIs for dynamic, real-time interactions with AI applications.
• Excellent price performance — Gemini 2.5 Flash is one of the lowest cost models.

The next generation of Gemini reasoning models includes Gemini 2.5 Pro and its smaller version Gemini 2.5 Flash. Imagen 3 text-to-image model with improved generation and editing capabilities. Chirp 3 is an audio generation and understanding model with Instant Custom Voice creation. Lyria is a text-to-music model and Veo 2, an advanced text or image to video generation model with new editing and camera control features.

All these models, along with over 200 third-party models, like Meta’s open-source Llama 4 are now available in Vertex AI Model Garden.

Finally, Model Armor was introduced for automated prompt and response protection. It protects sensitive data through integration with Google Cloud’s Sensitive Data Protection service (fka Cloud DLP) and blocks inappropriate content through safety and responsible AI filters.

Finally, Google Cloud offers indemnification for AI model outputs and the data used to train them in specific contexts. This provides customers with contractual protection against certain third-party intellectual property (IP) infringement claims related to the use of Google Cloud’s AI models and training data, reducing legal risks when deploying AI applications.

AI Tools and Platform

A compelling testament to Google Cloud’s AI-led progress is the remarkable growth of Vertex AI, Google Cloud’s unified machine learning platform, which now boasts a community of over 4 million developers. Adding to this intense AI focus was a significant emphasis on the development and deployment of AI agents, intelligent systems designed to autonomously perform tasks and workflows.

Agent Development Kit (ADK)

ADK is an open-source framework that simplifies the process of building sophisticated multi-agent systems. And, that too under 100 lines of vibe code! ADK supports the agent development lifecycle:

• Design agents with proper guardrails and orchestration controls.
• Develop with your choice of models and interact with your agents in human-like conversations with ADK’s bidirectional audio and video streaming capabilities.
• Iterate and test using different models and deployment options. ADK allows integration with remote agents and access to data sources using MCP. It also integrates with LangGraph.
• Deploy to production easily using the direct integration to Vertex AI Agent Engine.

AI Agent Marketplace within Google Cloud Marketplace enables buying, selling, and managing AI agents from trusted partners.

Google unveiled a Chrome integration, streamlining agent accessibility directly within the familiar browser environment. Furthermore, the introduction of a no-code Agent Assembler democratizes agent creation, enabling non-technical users to build custom AI assistants through an intuitive visual interface. Complementing these user-friendly tools were the announcement of innovative pre-built agents like the Idea Generation agent, designed to foster creativity and brainstorming, and the Deep Research agent, intended to automate and accelerate in-depth information gathering and analysis for employees. These advancements collectively aim to seamlessly integrate AI-powered assistance into daily workflows, enhancing productivity and unlocking new capabilities across the workforce.

Agent-2-Agent (A2A) Protocol

While the industry was still buzzing with wide support for Anthropic’s Model Context Protocol (MCP), Google surprised with the launch of another interoperability standard — Agent2Agent Protocol, aptly called A2A. This open protocol is as a result of collaboration with over 50 industry leaders.

Table 1 shows how A2A complements MCP.

Agents utilizing the A2A protocol advertise their capabilities through an “Agent Card,” which is a public metadata file typically formatted in JSON. This card serves as a directory of the agent’s skills, its network endpoint address, and the authentication requirements necessary to interact with it. Client agents use the Card to discover and identify the most appropriate agent for a specific task.

You can learn more about MCP here.

Agentspace

Agentspace is Google Cloud’s common portal that acts as a launch point for custom AI agents and for searching enterprise truth for an entire organization. It launched in December 2024. It uses a conversational interface to search databases like AlloyDB as well as 3rd party applications, such as Confluence, Google Drive, Jira, Microsoft SharePoint, and ServiceNow. Users can upload documents to synthesize, uncover insights, and use new ways of engaging with data using NotebookLM to accomplish complex tasks that require planning, research, content generation, and actions — all with a single prompt.​​

Data Analytics

BigQuery is the tour de force of Google Cloud’s data analytics story. This offering is no longer just a data warehouse but a whole ecosystem encompassing lakehouse, AI, governed data activation, and real-time streaming use cases. At the keynote, it was revealed that BigQuery has five times the number of customers compared to Snowflake and Databricks.

There are so many new capabilities to share that it could be a whole document by itself.

Multi-engine Data Platform

BigQuery was originally built, in 2011, on top of Dremel, Google’s internal, distributed query execution engine designed for ultra-fast SQL-style read-only queries over massive datasets in columnar storage format called Capacitor.

Over the years, it has transformed from a data warehouse to a multi-engine data platform supporting multi-structured data. Along with SQL and Spark, it supports Hadoop, Flink, Spark, Hive, and Trino. In 2022, Spark was introduced in a serverless form, allowing developers to run PySpark, Scala, R, or Spark SQL jobs inside BigQuery, alongside their regular SQL queries.

However Spark as a runtime engine is not always performant and BigQuery gives users multiple choices:

• Open-source Apache Spark in serverless (preview). This option has been tuned to be 2.7x faster processing than the prior year.
• SparkSQL with BigQuery pushdown
• Optimized Spark using C++-based OSS Velox and Gluten. The Google Cloud product teams are constantly experimenting with multiple engines and reserve the right to change the underlying engine to a more performant option in future. On a side note, BigQuery’s C++ engine is very similar to Databricks’ C++-based Photon execution engine rationale.

For developers who want more control over their Spark clusters, Google Cloud offers Dataproc. This product also has a serverless Spark option.

BigQuery Continuous Queries became GA at Next ’25. While BigQuery has long supported streaming ingestion, Continuous Queries now allows for continuous processing of that data using familiar SQL. This event-driven data platform does not require separate streaming engines, simplifying architectures and reducing latency.

BigQuery Pipelines and BigQuery Data Preparation also became GA at Next ’25. BigQuery Pipelines are powered by Dataform, a fully managed data transformation service that defines complex data transformations using SQL within a Git-based version control system. Google acquired Dataform in 2020. The pipelines are designed, executed and monitored in a low-code visual interface but it generates SQL code in pipe query syntax, which can be versioned in a Git repository.

Several capabilities mentioned in this section rely on Gemini models to analyze data and schema to provide intelligent suggestions for cleaning, joining, enriching, wrangling, and transforming data. Hence, let’s take a look at the role of LLM-generated metadata inside BigQuery next.

Built-in data governance

Every time BigQuery ingests raw data, the “Knowledge Engine” uses LLMs to profile the data, determine column names and generate descriptions. The Knowledge Engine is a knowledge graph and is built on Gemini. What you are witnessing here is a data catalog that is embedded inside BigQuery to enable ‘real-time data activation flywheel.’ With this new introduction, Dataplex, Google Cloud’s unified governance can now be accessed via three methods:

1. Standalone Dataplex UI
2. Build their own UI. Customers use Dataplex APIs, such as Ford, Carrefour, Walmart etc.
3. Headless and embedded inside BigQuery. This was introduced at Next ‘25

The embedded Dataplex capability creates the target enterprise relationship diagram, relevant tables and rates the relations in the ERD to create a social graph. It predicts common user questions, called ‘Data Insights.’. Finally, it creates a semantic model that is compatible with dbt and Looker.

The same process is used for structured or unstructured data that is stored in GCS and accessed via BigLake. One of Google Cloud’s guiding principles is to unify storage and metadata across BigQuery and GCS. This allows for single policies to be applied to all data types.

Since Dataplex is embedded, users can use SQL or PySpark to access data securely. In addition, users can also use their own IDE like VSCode or Jupyter to access BigQuery data.

Google Cloud has introduced Universal Catalog that combines all the metadata that has been created in the steps above. This new catalog contains technical metastore information along with the business metadata and runtime data such as usage statistics, query patterns, Apache Iceberg partitions, etc.

Data Agents

We have thus far looked at various compute use cases for Big Query and the role of LLM-generated metadata driving unified governance. These developments have led to the creation of four AI agents. At Next ’25, Google Cloud announced four agents:

• Data engineering agent (experimental): It is embedded in BigQuery pipelines and performs data preparation and maintains data quality rules with anomaly detection. As mentioned earlier, pipelines and data preparation are now GA but anomaly detection is in preview.
• Data science agent (GA): This agent is embedded in Colab and it automates every stage of the model lifecycle, e.g. feature engineering, intelligent model selection, scalable training, and faster iteration. Customers can also sign up to use this agent in BigQuery (experimental).
• Conversational agent (preview): Embedded in Looker, it not only allows users to do advanced analysis in natural language, but it also provides explainability into its ‘thinking’ to help reduce errors. This agent uses Looker’s semantic layer to resolve business terms in natural language prompts.
• Knowledge Engine (preview): Embedded in BigQuery, as mentioned above, this agent analyzes schema relationships, table descriptions, and query histories to generate real-time metadata and recommend business glossary terms.

These Gemini-powered agentic experiences are available to all customers at no additional costs.

AI Query Engine

This innovative engine transcends the traditional boundaries between structured and unstructured data, enabling users to seamlessly query and analyze both within a single, unified query.

Imagine the power of asking a question that spans your product catalog (structured data) and a collection of customer-uploaded images (unstructured data). For instance, a user could pose a query like: “Show me all the big buoyant blue bean bags from our product list that appear in customer photos tagged with ‘living room’ from the last quarter.” This capability goes far beyond simple keyword matching or basic joins.

The magic behind this lies in the intelligent co-processing of standard SQL with the real-time reasoning and understanding capabilities of Google’s Gemini models. When such a query is executed, the AI query engine leverages Gemini’s runtime access to vast real-world knowledge, its nuanced linguistic understanding of the query, and its sophisticated reasoning abilities to analyze the image content. It can identify objects, understand visual contexts, and relate them to the structured information in the product catalog based on attributes like color, product type, and even potentially subtle visual features.

This marks a significant step towards making unstructured data a “first-class citizen” within BigQuery, elevating it from a secondary data type requiring separate tools and processes to an integral part of the analytical fabric. This unified approach eliminates the complexities of disparate data silos and empowers users to derive richer, more holistic insights from their entire data universe.

Storage

Google Cloud’s goal is to unify the storage and querying of diverse data types within Google’s data warehouse. As a result, they announced a preview of BigQuery Multimodal Tables. For example, a product catalog information (structured) can be stored and analyzed alongside product images or Zoom call transcripts (unstructured) in the same table. This is possible via ‘object tables’ that read and index metadata from files stored on Google Cloud Storage (GCS) object store.

There has been so much emphasis on unstructured data and AI, that it has taken us so long to get to the other big movement in the data foundation — open table formats. BigQuery tables for Apache Iceberg (preview) helps connect Iceberg data to SQL, Spark, AI and third-party engines in an open and interoperable manner. It inherits all the benefits of the underlying BigQuery such as automated table management, serverless scale, advanced governance, high-performance streaming, auto-AI generated insights, and fine-grained access control.

Many of the BigQuery additions are designed to minimize operational overhead and deliver optimal price performance by making governance and orchestration autonomous and invisible. With so many new capabilities, BigQuery has unified its spend commit across data processing engines, streaming, and governance.

Databases

All Google Cloud operational databases — Spanner, AlloyDB, Cloud SQL, Bigtable, Firestore, and Memorystore embed generative AI technologies. Next ’25 witnessed unveiling of several significant new enhancements across each of these databases. New developments were also announced for partner databases, such as GA of Oracle Exadata X11M in 20 Google Cloud locations. Snap (formerly Snapchat) runs Oracle Exadata in Google Cloud.

In this section, we analyze some of the more important new developments.

MCP Support

Operational databases fill a special need that data analytics offerings don’t — low latency real-time transactional use cases. The biggest beneficiary in the future will be context-aware agents that need access to real-time data to improve reasoning, planning, and agentic decisions. As a result, Google Cloud launched ‘MCP Toolbox for Databases’ which supports Spanner, AlloyDB, Cloud SQL, Neo4j and Dgraph. The toolbox is an open-source MCP Server to connect AI agents to enterprise data in a standardized and secure manner. It includes boilerplate templates, end-to-end observability with OpenTelemetry integration, and security through OAuth2 and OIDC. The Toolbox can be used by any database vendor to build connectors to their offering.

AlloyDB AI

Several updates, in addition to the MCP Toolbox, have been made to AlloyDB as shown in Figure 3.

Figure 3: AlloyDB Next ’25 Updates

AlloyDB updates include:

• Agentspace: This allows the search engine to leverage Gemini’ advanced reasoning on unstructured and structured AlloyDB enterprise data in real-time.
• Natural Language: This uses supplied context and interactive intent clarification when querying the database. AlloyDB’s parameterized secure views also provide an extra layer of security when accessing data through agents and gen AI apps.
• ScaNN indexing: Continued enhancements in vector indexing now means Scalable Nearest Neighbor (ScaNN) indexing leads to vectors searches being up to 10x faster than the hierarchical navigable small world (HNSW) index in standard PostgreSQL.
• New models: Three new models include Gemini Embedding text model, a model to improve cross attention reranking, and finally, a multimodal embeddings model.
• AI query engine: Similar to the eponymous announcement in the data analytics space, this offering allows developers to augment SQL statements with natural language expressions.

Many of these offerings are in preview.

Firestore MongoDB compatibility

Firestores’ community of over 600,000 monthly active developers can now use their existing MongoDB application code, drivers, and integrations with the Firestore service. In other words, they can benefit from Firestore’s multi-region replication with strong consistency, scalability, five 9‘s uptime, and serverless pricing model.

Database Migration Service (DMS)

DMS now supports SQL Server to PostgreSQL migrations for Cloud SQL and AlloyDB. DMS migrations are inline, which means the source database doesn’t suffer downtime. DMS is another service that benefits greatly from the embedded Gemini model which can automate the most difficult migration steps, like converting Transact-SQL code and SQL Server-specific data types to their PostgreSQL equivalents.

Hybrid Multi-cloud / Sovereign AI

Google Cloud’s Sovereign AI offerings help organizations, particularly those in highly regulated industries and the public sector, meet their digital sovereignty requirements while leveraging the power of Google Cloud AI. This encompasses control over data residency, operational transparency, data access, and compliance with local regulations.

Google Distributed Cloud (GDC) is a suite of hardware and software offerings that extends Google Cloud infrastructure and services to various environments beyond Google’s own data centers and is a key component of Google Cloud Sovereign AI. GDC has different deployment options, including:

• GDC Edge: Run workloads at the network edge, closer to users and devices, for low-latency applications.
• GDC Hosted (air-gapped): A fully disconnected cloud environment operated by a partner, designed for highly sensitive workloads with strict data sovereignty and security requirements, often used by public sector and regulated industries.
• GDC Virtual: A software-only solution that extends a consistent development and operational experience to existing on-premises data center environments.

At Next ’25, Google announced that its Gemini models and Agentspace can now run on GDC locally and in air-gapped environments. Vertex AI can be deployed and run within GDC environments and key services, such as Vertex AI Search and Generative AI, have achieved FedRAMP High authorization. This allows Google Cloud to offer “Assured Workloads for FedRAMP.”

Conclusion

Google Cloud Next ’25 wasn’t just a showcase of incremental updates; it presented a cohesive vision of a future where AI is deeply integrated into every layer of the cloud stack, fundamentally transforming how businesses operate and innovate. The advancements across infrastructure, data analytics, databases, and AI tools paint a picture of a more intelligent, agile, and efficient enterprise. The emphasis on AI agents, for instance, signals a move towards a more autonomous and proactive approach to workflows, empowering employees and unlocking new levels of productivity. The integration of AI into databases and analytics platforms promises to democratize data access and insight generation, enabling a wider range of users to leverage the power of AI-driven decision-making.

The announcements at Next ’25 underscore how the hyperscalers in general, and Google Cloud in particular, are embracing a strategic, holistic approach to cloud adoption. As a result, Google Cloud’s customers who effectively leverage these advancements are positioning themselves to thrive in the rapidly evolving landscape of cloud data platforms. These customers are not just adopting new technologies, but also fostering a data-driven culture, investing in AI skills, and prioritizing data governance to ensure responsible and ethical AI implementation.

Google Cloud’s commitment to open standards and interoperability, as evidenced by its support for open table formats and the Agent2Agent protocol, is crucial for avoiding vendor lock-in and enabling seamless integration with existing systems. We look forward to organizations translating the potential of these innovations into tangible business outcomes, driving growth, efficiency, and a competitive edge in an increasingly AI-driven world.