Google’s AI innovations at Cloud Next 2025: What CIOs need to know

At Cloud Next 2025, Google announced several updates that could help CIOs adopt and scale agents while reducing integration complexity and costs.

The event focused on providing enterprises with an AI-optimized platform and open frameworks that make agents interoperable.

During his one hour forty minute-keynote, Thomas Kurian, CEO of Google Cloud showcased updates around most of the company’s offerings, including new large language models (LLMs), a new AI accelerator chip, new open source frameworks around agents, and updates to its data analytics, databases, and productivity tools and services among others.

Cost-performance optimizations via new chip

One of the major updates announced last week was Google’s seventh generation Tensor Processing Unit (TPU) chip — Ironwood — targeted at accelerating AI workloads, especially inferencing.

With the new TPU, Google said it wants to offer better performance per watt per dollar than any TPUs it released earlier and this is welcome news for CIOs as they often have to do more with less or constrained resources.

CIOs are under pressure to accommodate the exponential rise in inferencing workloads within their budgets, fueled by the adoption of LLMs for running generative AI-driven applications.

“Ironwood brings performance gains for large AI workloads, but just as importantly, it reflects Google’s move to reduce its dependency on Nvidia, a shift that matters as CIOs grapple with hardware supply issues and rising GPU costs. Taken together, these tools aim to make enterprise AI more practical to deploy, scale, and manage,” said Kaustubh K, practice director at Everest Group.

Inferencing workloads are also going to see another growth spurt with the adoption of agentic AI — employing digital agents to complete tasks without manual intervention — as LLMs will start to proactively reason before they take steps to complete a task or user request.

Solving the agentic DevOps problem with open frameworks

Last week also saw Google announcing new open frameworks — the Agent Development Kit (ADK) and the Agent2Agent (A2A) protocol — to help enterprises build, manage, and connect multiple agents, even across different ecosystems.

While the focus of the ADK is to help developer teams build an AI agent faster with less complexity, albeit with enough controls to manage it, the A2A protocol is targeted at helping enterprises connect agents that are built on different ecosystems or vendor platforms.

Analysts believe that the frameworks will be a boon for CIOs and help solve challenges around building agents or agentic applications.

“Google is quietly redefining agent lifecycle management as it is destined to become the next DevOps frontier. It is going to be massive for IT as agentic DevOps is going to be one of the biggest new operational headaches for enterprises going forward,” said Dion Hinchcliffe, lead of the CIO practice at The Futurum Group.

“The big strategic play is standardizing how a developer team in an enterprise version, monitors, secures, and retires them across hybrid clouds, and Google’s the first hyperscaler to frame that coherently,” Hinchcliffe added.

Everest Group’s Kaustubh said that CIOs may find Google’s emphasis on open standards and hybrid deployment options practical, especially for integrating AI into existing environments.

“These features offer flexibility without requiring major platform shifts, which could be appealing for phased or use case-specific adoption,” Kaustubh added.

Explaining further how Google’s strategy differs from rivals, such as AWS and Microsoft, Hinchcliffe said, where Microsoft is optimizing for “AI as UX layer” and AWS is anchoring on “primitives,” Google is carving out the middle ground — a developer-ready but enterprise-scalable agentic architecture.

“This strategy could make it a real differentiator for CIOs who want agents that are interoperable, observable, and enterprise-governed, not just chatbots bolted onto SaaS,” Hinchcliffe added.

While Microsoft offers agent-building capabilities via Copilot Studio and Azure Studio inside Azure AI Foundry, AWS offers agent-building capabilities via Amazon Bedrock.

Smaller LLMs and other updates

At Cloud Next 2025, Google also introduced specialized LLMs for video, audio, and images in the form of Veo 2, Chirp 3, and Imagen 3.

According to analysts, these specialized LLMs might help enterprises achieve more accuracy on video, audio and image generation-related tasks while reducing costs to a certain extent.

Specialized LLMs and smaller, faster Gemini variants directly address cost-performance optimization — an unsolved issue in enterprise AI scaling, according to Hinchcliffe.

“For CIOs, these updates make it more realistic to embed LLMs in edge devices, private data stores, or verticalized apps without overspending,” Hinchcliffe said.

To enterprises with productivity, Google, last week, updated its productivity suite with new agents via Google Workspace and introduced new CES agents.

However, Hinchcliffe said the real value of these updates for any CIO lies in the fact that these agents or updates help bridge the gap between enterprise task automation and structured, governable AI workflows.

“This is something that most vendors gloss over in their demonstration of new capabilities,” Hinchcliffe said. Other updates that Google announced included a slew of data analytics, databases, networking, and security updates along with a new Application Design Center.