Google Cloud Accelerates into the Agentic AI Era with ‘Ironwood’ TPU and Advanced Agent Software
Introduction to the Agentic AI Era and Google Cloud's Preparations
Google Cloud is making a decisive move to lead the burgeoning agentic artificial intelligence (AI) era, marked by a significant unveiling of new hardware, sophisticated AI models, and robust software development tools at its recent NEXT conference. This strategic initiative signals a profound shift in how AI will be developed and deployed, moving beyond responsive systems to proactive, insightful agents capable of complex reasoning and autonomous operation. The company’s multi-pronged approach, encompassing custom silicon, advanced models, and developer-centric software, positions it as a key enabler for businesses looking to leverage the next generation of AI capabilities.
The ‘Ironwood’ TPU: Powering the Next Generation of AI
At the heart of Google Cloud’s new AI infrastructure is ‘Ironwood,’ the seventh-generation Tensor Processing Unit (TPU). This custom-designed chip represents a monumental leap in performance and efficiency, engineered specifically to meet the exponentially growing demands of advanced AI models, particularly those characterized as “thinking models.” Google states that Ironwood is twice as power-efficient as its predecessor, a critical factor in managing the immense energy requirements of large-scale AI computations. The scalability of Ironwood is a standout feature, with pods designed to accommodate over 9,000 chips. When fully scaled, a single Ironwood pod can deliver an astounding 42.5 exaflops of compute power. For context, this far surpasses the capabilities of the world’s leading supercomputers, such as El Capitan, which offers approximately 1.7 exaflops per pod. This massive increase in computational power is directly attributed to innovations in TPU architecture, including advanced liquid cooling and optical switching, which have reportedly resulted in 100-fold improvements in sustained performance over conventional designs. This enhanced performance is crucial for serving the observed tenfold year-over-year increase in demand for training and serving AI models.
Enhanced Infrastructure: Networking and Software Runtimes
Complementing the raw power of the Ironwood TPU, Google Cloud is also enhancing its underlying infrastructure to better support AI workloads. The company is making its proprietary advanced networking technology, Google Cloud WAN, available to customers for the first time. This provides access to the same planet-scale network that underpins Google’s global services like Gmail, YouTube, and Search, offering unparalleled connectivity and performance. Furthermore, Google is democratizing its internal machine learning runtime, ‘Pathways,’ developed by Google DeepMind. Pathways on Google Cloud enables customers to efficiently scale model serving across hundreds of TPUs, ensuring exceptional performance and simplifying the management of large-scale AI deployments. This integration of cutting-edge hardware and optimized software infrastructure creates a powerful platform for developing and deploying sophisticated AI agents.
Google’s Gemini Models: Driving Reasoning and Efficiency
Central to Google Cloud’s AI strategy are its advanced Gemini models. Gemini 2.5 Pro, a highly capable reasoning model accessible through Vertex AI, is designed to tackle complex problems by employing multi-step thought processes, making it ideal for demanding applications such as drug discovery, financial modeling, and risk management. Recognizing the need for more accessible and efficient models for everyday use cases, Google is introducing Gemini 2.5 Flash. This model is optimized for speed and high-volume interactions, capable of generating real-time summaries, assisting with basic coding tasks, and performing function calls where responsiveness is paramount. Gemini 2.5 Flash is expected to be widely adopted for powering AI agents, given its balance of performance and cost-effectiveness.
Empowering AI Agent Development with New Software Tools
The proliferation of AI agents necessitates robust tools for their development and management. Google Cloud is addressing this need with a suite of new software offerings. The Agent Development Kit (ADK) is a unified development environment designed to streamline the process of building, testing, and operating AI agents. With ADK, developers can reportedly create multi-agent systems with fewer than 100 lines of code, incorporating creative reasoning and strict guardrails to steer agent behavior. The platform aims to enable a rapid transition from concept to production, often within a week. To further facilitate agent creation and integration, Google Cloud has launched ‘Agent Garden.’ This resource provides a collection of ready-to-use samples and tools, making it easy for users to connect agents to over 100 pre-built connectors, custom APIs, and other integration workflows. Agent Garden also supports the Model Context Protocol (MCP), an emerging industry standard for connecting data with AI models.
Fostering Interoperability and an Agent Ecosystem
Google Cloud is not only developing its own agent technologies but also actively fostering an ecosystem of interoperability. The company is supporting MCP, which is gaining traction as a standard for data-model interaction. Additionally, Google Cloud has announced its own Agent to Agent (A2A) protocol. Unlike MCP, which focuses on connecting agents to AI models and tools, A2A is specifically designed to enable agents to call and connect with other agents, promoting collaboration and distributed intelligence. To further catalyze this ecosystem, Google Cloud is launching an AI Agent Marketplace, where customers can discover and select partner-developed AI agents. Complementing this is Google Agent Space, a platform designed to serve as a central hub for organizations to share information and manage AI agents among their employees.
Specialized Agents for Data and Security
Google Cloud is also extending its agent capabilities to specialized domains, particularly in data engineering, data science, and data analytics. New specialized data agents are being integrated directly into BigQuery pipelines to streamline data pipeline creation. Other agents are being introduced for data preparation tasks, such as transformation and enrichment, and for anomaly detection. Brad Calder, vice president and GM of Google Cloud, highlighted that these agents cover the entire data engineering lifecycle, from metadata generation and catalog automation to maintaining data quality. Data scientists will benefit from a new agent within Google’s Colab notebook, designed to assist with feature engineering, model selection, and iterative development. Data security is also a significant focus, with the introduction of two new data engineering agents: one for analyzing security threats and another for detecting malware. These specialized agents aim to automate complex tasks, improve efficiency, and enhance the security posture of data operations.
Visibility and Interaction with Gemini Code Assist
For developers working with AI agents, particularly in coding tasks, Google Cloud is rolling out its new Gemini Code Assist Kanban board. This tool provides a real-time display of the tasks that Google AI agents are currently working on, and crucially, it allows users to interact with these agents. This feature enhances transparency and control, enabling developers to monitor progress, provide feedback, and collaborate more effectively with their AI coding assistants. The integration of such tools signifies a move towards more collaborative and human-in-the-loop AI development processes.
Conclusion: A Comprehensive Strategy for the Agentic AI Future
Google Cloud’s announcements at NEXT underscore a comprehensive and ambitious strategy to lead the agentic AI era. By combining the immense power of the new ‘Ironwood’ TPU with advanced reasoning models like Gemini, and a robust suite of developer tools including ADK and A2A, Google is building an end-to-end platform for AI innovation. The focus on specialized agents, ecosystem development, and enhanced infrastructure demonstrates a clear vision for how AI will integrate into business operations, driving efficiency, enabling new capabilities, and unlocking unprecedented insights. The company
AI Summary
Google Cloud is making a significant push into the agentic AI era with a comprehensive suite of announcements at its NEXT conference. At the core of this strategy is the unveiling of ‘Ironwood,’ the seventh-generation Tensor Processing Unit (TPU). This new TPU is designed to be twice as power-efficient as its predecessor and is engineered to handle the exponentially growing demands of advanced AI models, particularly those referred to as "thinking models" like Gemini 2.5. Ironwood boasts an impressive scale, with pods capable of supporting over 9,000 chips, delivering a staggering 42.5 exaflops of compute power per pod. This represents a substantial leap in computational capability, significantly outperforming current supercomputers like El Capitan, which offers 1.7 exaflops per pod. Google highlights that this immense power is necessary to meet the observed 10x year-over-year increase in demand for training and serving AI models. Innovations in the TPU architecture, including liquid cooling and optical switching, have reportedly led to 100-fold improvements in sustained performance compared to conventional designs. Beyond hardware, Google Cloud is enhancing its networking capabilities by making its internal Google Cloud WAN technology available to customers, providing access to the same planet-scale network that powers Google’s core services. Furthermore, the company is democratizing its internal machine learning runtime, ‘Pathways,’ developed by Google DeepMind, allowing customers to scale model serving efficiently across hundreds of TPUs. The Gemini family of models is central to this push, with Gemini 2.5 Pro, a powerful reasoning model available via Vertex AI, capable of complex problem-solving. A more accessible version, Gemini 2.5 Flash, is being introduced for everyday use cases, offering fast responses and high-volume interaction capabilities, making it suitable for tasks like real-time summarization and basic coding. These reasoning models are poised to be instrumental in the development of AI agents. To support this, Google Cloud is launching a new Agent Development Kit (ADK), described as a unified development environment that simplifies the building, testing, and operation of AI agents, enabling the creation of multi-agent systems with minimal code and robust guardrails. Complementing the ADK is ‘Agent Garden,’ a collection of ready-to-use samples and tools that facilitate agent integration with over 100 pre-built connectors and custom APIs. Google is also actively supporting the Model Context Protocol (MCP), an emerging industry standard for connecting data with models, and has introduced its own Agent to Agent (A2A) protocol, designed for direct agent-to-agent communication. The company is further fostering an agent ecosystem through an AI Agent Marketplace and Google Agent Space, a platform for organizations to share information about AI agents internally. Specialized data agents for data engineering and data science are also being integrated into services like BigQuery, enhancing tasks such as data pipeline creation, data preparation, and anomaly detection. For data scientists, a new agent in Google’s Colab notebook will assist with feature engineering and model development. Security is also a focus, with new agents for threat and malware analysis. Finally, the Gemini Code Assist Kanban board will provide real-time visibility and interaction capabilities for AI agents working on coding tasks. This multi-faceted approach underscores Google Cloud’s commitment to empowering developers and businesses to harness the full potential of the agentic AI era.