At Cloud Next, Google bets on AI everywhere

There's been a lot of discussion recently on how AI applications are evolving but based on many of the announcements that Google made at the Cloud Next event in Las Vegas, it seems increasingly clear that hybrid is where many of these developments are headed.

To be clear, Google made an enormously and impressively broad range of announcements at Cloud Next and not a single press release specifically mentioned Hybrid AI.

However, when you take a step back and analyze how several of the pieces fit together and look ahead to where the trends the company is driving appear to be going, the concept of GenAI-powered applications (and agents) that leverage a combination of the public cloud, enterprise private clouds and even smart devices-that is, Hybrid AI-appear inevitable.

A few highlights first. On the cloud infrastructure front, Google made several big debuts at Cloud Next, most of which focus on the increasing range of computer architecture options coming to customers of GCP.

Most notably, the company took the wraps off their 7th generation TPU processor, codenamed Ironwood, its in-house designed alternative to GPUs and the first to be specifically designed for inferencing workloads. In addition to 10x improvements in raw performance versus previous generations, what's impressive about the latest versions is the extent of high-speed chip-to-chip connectivity options that Google will be offering between them.

Taking a page from the Nvidia NVLink book, Google's latest AI Hypercomputer architecture lets up to 9,216 of these Gen 7 TPUs be interconnected into a single compute pod, providing plenty of bandwidth for even the largest of the new chain-of-thought based reasoning models starting to become available. In fact, Google claimed that maxing out a system could deliver up to 42.5 exaflops, more than 24x the computer power of today's fastest supercomputer.

ADK framework showing how you can build multi-agent systems

Another big theme from the Cloud Next keynote was around agents, including the tools to build them, to connect them to one another, and to integrate them more easily with a variety of LLMs.

Building on the company's previous Agentspace announcement - which allows enterprise employees to use Google's multi-modal search capabilities across enterprise data and build their own agents in a low code/no code manner-Google also debuted a new Agent Development Kit for developers as part of its Vertex AI platform.

Even more importantly, the company announced its Agent2Agent (A2A) protocol, which is an effort to standardize the means by which different agents can "talk" to each other and share information. A2A builds upon and is compatible with Anthropic's Model Context Protocol (MCP) which was introduced last year and is quickly gaining traction in the AI world.

In fact, it's Google's strong MCP support across a range of products that it introduced here at Cloud Next that really led to the hybrid AI conclusions I made earlier. MCP offers a standardized way for models to connect to a variety of different data sources – instead of having to deal with proprietary APIs – and provides a standardized means by which models can expose the various functions they're able to perform on these data sets.

In the process, this means that MCP both solves some big challenges in creating AI-powered applications that can tap into local data resources and opens up a world of intriguing possibilities for creating distributed AI applications that can tap into data sources, other models and other computing infrastructure across different physical locations. It's this capability that makes MCP so intriguing-and it's likely a big reason support for the nascent standard is growing so rapidly.

Google made the potential impact of MCP much more real by announcing it is now also allowing organizations to bring Gemini models, Agentspace and other AI tools into their private cloud/on-prem datacenter environments via the Google Distributed Cloud in the third quarter of this year. This is a hugely important development because it means that companies building apps with Google Cloud-based tools can use them across many different environments.

So, for example, it would be possible for an organization to tap into the essentially unlimited resources of Google's public cloud infrastructure to run certain functions with certain models and data sets stored there, while running other functions on different models that access data behind the firewall within their private cloud or datacenter environments.

This solves the data gravity problem that many organizations have been struggling with as they start to think about tapping into the powerful capabilities of today's most advanced LLMs because it essentially allows them to have the best of both worlds. It gives them massive cloud-based compute with data stored in the public cloud and local compute with the large and often most valuable proprietary data sets that many organizations still keep (or may want to repatriate) within their own environments.

Plus, it's even possible to extend the distributed nature of the computing environment to PCs and smartphones, particularly as the availability of devices with more powerful AI acceleration capabilities increases. While this last step likely won't happen overnight, it will become a critical capability as companies look to reduce the electricity demands and costs of their AI applications down the road.

Speaking of on-device capabilities, Google also announced several enhancements to their Workspace productivity offering at this year's Cloud Next. New AI-powered features include automation-focused Workflows, audio features in Docs and more. These build on many previous AI-powered functions that Google brought into Workspace earlier this year, including no-cost access to the most advanced version of the Gemini model, new data analysis functions in Sheets, document analysis and summarization across all the Workspace applications and more.

As with previous Cloud Next events, there were many more announcements that Google discussed across areas such as databases, code creation tools, the Firebase agent creation studio, Cloud WAN private network access, security improvements and much more.

It's a bit overwhelming to make sense of it all, to be honest, but it just shows how tremendously fast cloud-based offerings continue to expand, particularly with the integration of the even faster moving AI foundation model developments.

Ultimately, though, it's clear that Google is using its long history of AI developments as well as the recent advancements it's made with Gemini models and other AI tools as a clear differentiator for Google Cloud. In the process, they're continuing to position themselves in a unique way not only for current applications but also for hybrid AI applications down the road.

Bob O'Donnell is the founder and chief analyst of TECHnalysis Research, LLC a technology consulting firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on X @bobodtech