Google Cloud Ironwood TPU on a rack. (Source: Google Cloud)
Google Cloud is gearing up for the agentic AI era in a big way, and its showing off its new wares this week at its NEXT conference. The company unveiled a slew of new AI models and new software for developing and managing AI agents, as well as the seventh generation of the processor at the heart of its AI Hypercomputer, a TPU dubbed Ironwood, which Google says is twice as power efficient as the previous generation.
Google Cloud is seeing AI workloads shifting from model training to inference workloads, which is a trend that Nvidia also observed during its recent GTC conference. The seventh-gen Ironwood TPU was built from the ground up for inferencing at scale, according to Amin Vahdat, the company’s vice president of ML, systems and cloud AI. And oh my, what scale.
“Ironwood will scale to over 9,000 chips per pod to meet the exponentially growing demands of thinking models like Gemini 2.5,” Vahdat said during a press conference on Monday. “This scale will deliver a staggering 42.5 exaflops of compute per pod.”
For perspective, the world’s number one supercomputer, El Capitan, supports 1.7 Exaflops per pod, Vahdat said. By comparison, Ironwood running on Google Cloud’s TPU-based AI Hypercomputer will deliver more than 24 times the compute power of El Capitan, he pointed out.
Much of that compute power will go toward serving the burgeoning demand for AI workloads, he said. “We’ve seen a 10x year over year increase in demand for training and serving models,” Vahdat continued. “Innovations throughout the TPU architecture, such as liquid cooling and optical switching, have led to 100 times improvements in sustained performance relative to conventional architecture design.”
Google Cloud has made a few other improvements to its service to help customers put all that power to use. For instance, its making its internal advanced networking technology to, dubbed Google Cloud WAN, available to customers for the first time.
“Our customers can now tap into the same planet scale network that powers Google’s globally available services, including Gmail, YouTube, and search,” Vahdat said. “No other technology company can offer this to its customers.”
Google Cloud’s seventh-generation TPU, Ironwood. (Source: Google Cloud)
It also is making its own internal machine learning runtime, dubbed Next Pathways, available to customers. “Developed by Google DeepMind, Pathways on Google Cloud allows customers to scale out model serving to hundreds of TPUs with exceptional performance,” Vahdat said.
Google develops one of the world’s most capable foundation models, Gemini 2.5 Pro. The reasoning model, which is available through its Vertex AI service, is capable of breaking up complex problems and using multi-stepped thought processes to deliver accurate answers in demanding environments, such as drug discovery, financial modeling, and risk management, Vahdat said.
Soon Google Cloud customers will have a more affordable version of that model, dubbed Gemini 2.5 Flash. “Gemini 2.5 Flash is more affordable for everyday use cases,” Vahdat said. “The model gives cloud customers the ability for fast responses and high volume customer interactions. It can quickly generate real time summaries of documents or news, and can assist with basic coding tasks and function calling where responsiveness is important.”
Reasoning models such as Gemini 2.5 Flash will be widely used for AI agents, which are rapidly progressing in capability and usability. Google Cloud is using its NEXT conference to roll out a slew of additional software to help customers develop and manage their new robot workers.
For starters, Google Cloud is rolling out a new Agent Development Kit (ADK), which it bills as a “unified development environment” that “makes it easy to build, test and operate these agents,” Vahdat said.
“With ADK, customers can easily build a multi-agent system in under 100 lines of code and precisely steer agent behavior with creative reasoning and strict guardrails,” the Google VP said. “Customers can go from concept to testing, with real data and assets, to running with security and compliance in production in less than a week.”
Ironwood’s FLOPS per watt. (Source: Google Cloud)
Since growing new crops of AI agents will be so important, why not have a garden devoted to it? That’s essentially what Google Cloud is enabling with its aptly named Agent Garden, which Vahdat called a collection of ready to use samples and tools directly accessible in SDK. The Agent Garden will make it easy for users to connect agents to 100 plus pre-built connectors, as well as to custom APIs, other integration workflows, or data stored in customers cloud systems. It will also support Model Context Protocol (MCP), the new protocol developed by Anthropic to connect data with models
Google Cloud is supporting MCP, which appears to have the early lead in the search for industry standard protocols. But there’s also room for an Agent to Agent protocol, which is something that Google Cloud just announced. A2A, as it’s called, will be geared at enabling agents to call and connect to other agents, as opposed to AI models and tools, which is the focus with MCP, Vehdat said.
But wait, there’s more agentic AI from Google Cloud! The company is rolling out an AI Agent Marketplace where customers can search for and select from a slew of partner-developed AI agents to use in their Google Cloud environment. And Google Cloud is also launching Google Agent Space, which is designed to provide organizations a clearinghouse of sorts to share information about AI agents to employees.
Google Cloud also provides a slew of AI agents to handle a range of data engineering, data science, and data analytics tasks. It is using Google Cloud Next to unveil enhancements to these agents, too.
The company is launching a handful for new specialized data agents for data engineering and data science at NEXT, according to Brad Calder, vice president and GM of Google Cloud. Its adding agents directly into BigQuery pipelines to build data pipelines. It’s also adding agents to perform data prep tasks, such as transformation and enrichment, and another specifically for anomaly detection.
Google Cloud’s Agent Engine functioning in AgentSpace. (Source: Google Cloud)
“We deliver agents for all aspects of the data engineering lifecycle, from catalog automation metadata generation to maintaining data quality to data pipeline generation,” Calder said during the press conference.
Data scientists will appreciate the new agent in Google’s Colab notebook, which will help with a range of tasks, including feature engineering, model selection, and training and iteration. Data security is also a focus for Google Cloud’s agentic development, and to that end, it is launching new two data engineering agents, one that analyzes security threats and another that analyzes malware.
Finally, Google Cloud is rolling out its new Gemini Code Assist Kanban board, which provides a real time display of the tasks that Google AI agents are working on, and also gives them the ability to interact with the agents.
Google Cloud has a ton more news at the show (the book of blogs it shared with reporters was nearly 200 pages). Keep BigDATAwire bookmarked for the most relevant bits.
This article first appeared on BigDATAwire.
Related
