Cloud Roadmap Series: Discover secure and unified infrastructure for agents - Engish transcript

Hi everyone. I'm excited to welcome you to the Google Cloud Roadmap series and we're going to be talking with the advancements we have in infrastructure that can help keep you secure and really embrace all the challenges and excitement around agents. I'm Drew Bradstock and I'm the senior director of product for orchestration and Kubernetes. >> Hi, I'm Muninder Singh Sambi, vice president for Google distributed cloud. >> And Muninder and I are quite excited to join you and walk through some of the conversations we had with our customers and what happened on stage at Next.

So, our agenda today is pretty jam-packed, so definitely stay near the end for the wrap-up. We're going to walk through some of the new innovations we have within Google Cloud. Really the exciting changes we brought within dynamic compute. How to secure your connectivity, which is definitely a challenge for all of us in market. Where we're going with a unified data layer. As well as a lot of the excitement and challenges of hybrid and sovereign cloud.

And then we'll bring it all back together and next steps that all of you can take. So, with that, let's get going. So, first, let's look at your traditional infrastructure. And it is what you expect. You've got your core apps, you've got databases, you've got ERP systems. And they really are your foundation and need absolute and unwavering reliability. They can't afford to go down. But, we just surveyed over 1,400 tech leaders and honestly, the results were a bit of a wake-up call.

Only 17%, that's not even one in five, feel confident their current setup can actually handle these new agentic workloads we're talking about. And agentic AI isn't just a new feature or buzzword. It's a massive stress test on everything you and Google has built over the past decade. When you have agents acting at machine speed, that safe infrastructure of yesterday is going to rapidly show its age. Also, the speed of the game has fundamentally changed.

If you're trying to build for this next wave using a lot of your same old foundations, you're going to hit a wall and a pretty hard wall at that. So, let's talk about the four pillars that Google's cloud infrastructure relies on to help you innovate in today's agentic era. First of all, there's dynamic compute. Secondly, secure connectivity. Third, having a unified data layer. And finally, ensuring you have digital sovereignty.

Let's start with what we call dynamic compute. It consists of Google Compute Engine and Google Kubernetes Engine to build a core of dynamic computing. It consists of a workload-optimized portfolio of machines and all our new investments in custom silicon. The agentic enterprise changes the architecture and introduces the need for a broad range of CPUs to complement AI accelerators. And this is a really key point to realize.

Often when people think of agentic, they think of GPUs and TPUs and AI infra, but it's very hard. The agentic enterprise runs on standard compute. Now, it also runs on Google Cloud TPUs or Tensor Processing Units and GPUs, which are great for core AI training and your inference workloads. We just pushed the frontier again and announced our eighth-generation Tensor Processing Units. These Google Cloud TPUs can can the core of the infrastructure strategy.

Designed to train the largest AI models and to serve millions and millions of agents across your business. For agentic inference and execution, CPUs are running agentic planning, orchestration, and response verification. Let's point out one specific family, which is our C4 series Intel base instances with matrix mass support, which are great for these use cases, as well as for very small model inference. All while providing the high velocity environment needed to lead the agentic world of tomorrow.

Now, modernizing core apps is a crucial part of innovating the agentic era. And you can accelerate their transformation with GKE, providing containerized API-driven platform to incrementally add AI services, agentic workloads, and scalable cloud-native components without rewriting your entire core. And we've just made at Cloud Next some exciting announcements about dynamic infrastructure that powers these core apps. Now, one thing I want to touch on is often people think, well, I don't have time to build a modernize.

I can't work towards this. The reality is that you have the opportunity to really optimize and innovate. You can leverage a lot of our hardware to be able to find cost savings such you now have the budget, the time, and that core infrastructure to tackle what Mininder and I are going to be walking through. Now, on the Google side, first, we've got the brand new network optimized VM family, starting with the C4N series.

C4N is designed for IO-intensive workloads such as security appliances, analytics, and media streaming. And with the World Cup right around the corner, or depending when you're watching it occurring, there's going to be a lot of media streaming going on, whether you like it or not in your office. The C4N VMs can process 95 million packets per second, up to 40% faster than other hyperscalers. They also deliver up to 400 gigabits of VM to VM network bandwidth, a 4x increase per vCPU over the standard C4.

C4N's performance can handle both the ever-increasing scale of your applications and the growing demand from agents. We also introduced the M4N series. Running Oracle workloads on M4N can reduce the TCO by over 20% compared to hyperscalers, which is a huge delta for database workloads. This helps enable you to run Oracle more efficiently and reinvest your savings right back into modernization with AI. We also announced on the storage side Hyperdisk Extreme on M4N and C4N, which can now support nearly 1 million IOPS and 25 gigabits per second throughput.

Whether it's M4N enabling millions of Oracle transactions or C4N tackling high-volume IO on storage task, Hyperdisk Extreme enables maximum per core performance with significantly fewer cores, giving you that benefit you need for these next-gen workloads we're walking through. We also recently introduced Google Axion N4A, which delivers up to 2x better price performance to meet the single-thread performance needs of AI and your foundational workloads alike.

Customers can offload state management to these powerful CPUs, ensuring that TPUs and GPUs are never starved for context. And we just announced, back at Next, GK Agent Sandbox, which is built on open-source gVisor technology. This new capability senses agent bursts and responds instantly, starting agent sandboxes in less than a second in scaling up to 300 sandboxes per cluster, which will help you enable your agents to perceive, reason, and act without a pause.

Because agents act as proxies for users, we've made sandboxing a native no-code feature of GKE. Our agent sandboxes isolate sensitive tasks like code generation at the kernel, storage, and network level, giving you total control over agent activities. In fact, GKE agent sandbox, combined with Google accident M4A, offers up to 30% better price performance over other hyperscalers. With this combination, you can build agents without compromising on cost or choice.

Now, we've talked a lot about dynamic computing, and I'm happy to hand it off to Maninder, who's going to walk through the other advancements that we have coming. >> Thank you, Drew. For compute to be truly dynamic, it cannot be constrained by the boundaries of a single data center, or even a specific cloud. Over the last 25 years, we have watched the network evolve from getting businesses online during the internet era to supporting massive scale of the streaming era, and eventually migrating mission-critical workloads in the cloud era.

But now, in the AI era, a new set of challenges have emerged. Your infrastructure must extend to wherever your data lives and wherever your agents need to act. To address exactly these challenges, we developed a global cross-cloud network that is not bound by static limits or topology. It supports and helps secure your dynamic compute, core applications, agents, and AI services. Google's cross-cloud network connects your agents, data, AI core services, and users across clouds and even on premises.

We launched cross-cloud network 2 years ago, and now it is being used by 65% of the Fortune 100. And it now handles more than 27 exabytes of data per month. We continue to expand the cross-cloud network capabilities, making sure you as our customers are equipped to respond to the demands of the AI era. In December 2025, we announced a new cross-cloud interconnect collaboration with AWS, allowing you to privately connect your multi-cloud environments in minutes versus weeks.

Today, it is generally available. And now we are also announcing private connectivity to both CoreWeave and Newman, which are in preview. Now, let's talk about beyond connectivity. Let's talk about observability and management. To allow for observability of cross-cloud agents, applications, and networks, we announced cloud network insights that allows you as a customer to monitor the performance and experience of your agents, your web applications, your networks across Google Cloud, other public clouds, and even your own data centers.

In addition, we also announced the network domain agent that can be used to automate and simplify network operations. Next, let's turn to network security. The threat landscape, as you all know, is evolving faster than ever with AI-driven attacks. And staying ahead requires the latest defenses. We realize how important security is, and our approach has always been to have a built-in security versus bolted-on. This requires us to actually offer you as our customers real-time protection against zero-day attacks.

With this, we are very pleased of announcing new advanced protection capabilities across our network security products. Starting with the advanced malware sandbox for our Cloud NGFW, which is powered by Palo Alto Networks. This allows to support and protect against thousands of protection rules. We're also extending this same level of protection for thousands of rules for vulnerabilities and zero days for Cloud Armor that is powered by Talos.

Now, both of these capabilities are fully integrated into our cross-cloud network, ensuring real-time intelligent security threat protection against AI attacks. Now, let's talk about agents. As most of you know, you're building innovative AI-powered agents. But, taking that innovation of production has some real challenges. Many organizations that we talk to are facing a fragmented landscape of what we like to call agent sprawl, and I'm sure many of you are witnessing it.

Here, the agents are scattered across diverse platforms without consistent governance or even security. This creates gaps, blind spots. It leaves your organization exposed to risk that traditional architectures were just not built to handle. To bridge exactly these gaps, we're very pleased to announce Agent Gateway, which is part of our Gemini Enterprise Agent platform. Together with Agent Identity, Agent Registry, and Model Armor, Agent Gateway helps you as a customer to enforce enterprise-grade governance and strong data protection at scale.

Now, we understand your agents need to talk to other agents. They need to talk to models, MCP tools, and APIs in Google Cloud, cross-cloud, or even on-premise. Agent Gateway allows you to do that in a simple and easy way. It has the ability to inspect and govern the connections in your AI solutions. Authenticating and securing every call to and from from your agents, tools, and supporting services. It integrates with identity, IAM services, and AI safety services from Google Cloud's Model Armor, Agent Identity, and Agent Registry.

Being built on open foundations, Agent Gateway also enables a broad ecosystem of third-party services and identity across access services that gives you the best of choice from our marketplace services. Now, we can't forget what is crucial to all things AI, and that is data. So, let's talk about our third pillar, which is the unified data layer. In the legacy world, you have had a fleet of data stores that is holding vast amounts of structured and unstructured data that was often misunderstood or dark.

This information was scattered across silos, whether it be Google Cloud Storage, databases, and data lakes on Google Cloud, or data stores on any other platform. In the agentic era, we are now allowing you to move from data to knowledge. And we are launching the knowledge catalog that provides a comprehensive semantic map of your entire enterprise data landscape, enabling you to fully understand your data and provide context to your AI agents.

Context is especially important for unstructured data. The challenge is that this cloud storage that knows about an object name and its size, but it actually doesn't know what it contains. And to solve exactly this problem, we are also announcing Smart Storage in Google Cloud Storage that makes objects self-describing. Let me get into some details. with smart storage. You can now automatically annotate images, extract document entities, and find data using semantic search.

This enables intelligent search across the entire storage estate, allowing you to easily select the right subset of data for AI agents and AI training use cases. No custom pipelines, no copies. Teams can easily find the right training data in seconds versus weeks. We also launched the cross-cloud lakehouse. We understand that your data lives everywhere, and your infrastructure should not dictate where it stays. With cross-cloud lakehouse, which is built on an open Apache iceberg standard, your agents can natively read catalogs and object stores from any vendor.

So, whether your data is stored in Google Cloud, storage, databases like AlloyDB and Spanner, or a data warehouse such as BigQuery or any on-prem or cloud environment, it is easily accessible to all your agents. This allows for real-time reasoning across your entire data landscape with absolutely no preparation required. You move from a world where you were managing silos to a world where you're actually creating knowledge across your entire application estate.

Now, let's talk about our fourth pillar, which is hybrid and sovereign cloud. Up till now, we've discussed how to build and optimize infrastructure in the public cloud. But, the reality of the enterprise world today is that not every workload belongs there or can be moved into the public cloud. To unlock the true value of the Agentic Enterprise, we have to look at how we extend the same capabilities that I talked about in two environments that require absolute control.

Now that balance between innovation and control is at the heart of what organizations are wrestling with. When they specially navigate on sovereign requirements today. We find that in many of our conversations with global enterprise, with public sector leaders, defense and intelligence, the practical operational need usually falls into three distinct categories. The first being regulatory compliance. This is about ensuring organizations can quickly adapt to shifting local data regulations and regional privacy laws.

Second, it's about local operations. For many, simple data residency is not enough and we understand that. The data and the infrastructure must be actively operated, managed and supported locally by authorized personnel. Third, it's about total isolation. For the most highly regulated environments, the sovereign governments, the defense and intelligence, there's a hard line. The infrastructure must run fully disconnected from the public internet with zero external dependency.

Our goal at Google is to ensure that you don't have to choose having the power of Google's AI while meeting these operational realities. To solve this, Google Cloud doesn't force a one-size-fits-all model. Instead, we provide a full spectrum of sovereignty options tailored to where your workload needs to sit on the spectrum. To meet customers exactly where their compliance and architecture needs are, we have built the most comprehensive sovereign cloud cloud platform.

Spanning from our public cloud all the way to our physical premises. We have the Google Cloud data boundary, which is the assured workloads that can be run in our region. We have the Google Cloud dedicated that is operated by local providers. And then we have our Google distributed cloud that runs on premise in your environment. Let me walk you through the solutions. On the Google Cloud data boundary, we provide strict data residency and access controls on our public cloud.

Support for many compliance requirements across the globe. If you have access to Google Cloud products and capabilities to manage and control your encryption keys. We are also taking this commitment further by introducing confidential external key management. It brings the benefit of confidential computing to provide access control to your keys with a verifiable hardware isolated environment. In countries that mandate cloud operations to be managed by local independent entities, we have now introduced Google Cloud dedicated that is designed to meet local sovereignty rules and operated by independent regional providers.

It also comes with GPUs to support AI workloads in your region. For example, in France, we have partnered with Thales that can actually operate a cloud for Europe. It has already achieved the highest cloud security certification, which is the SecNumCloud 3.2. We will be expanding this capability to Germany later this year. And lastly, we provide Google distributed cloud. What makes Google distributed cloud uniquely is the flexibility.

It brings the power of Google Cloud directly to your physical locations and can be deployed in a connected state giving you hybrid operations with cloud management ease. It can also be deployed in a fully air-gapped environment that requires complete offline isolation. While we bring new innovations, we're also bringing some of our premier foundation models. Uh state-of-the-art Gemini on Google distributed cloud in both the connected as well as the air gapped environment.

With this capability, our customers no longer have to choose between cutting-edge innovation and strict compliance. We're delivering Gemini directly into your perimeter. And to ensure you have the performance that is required to run these frontier models, we are very thrilled to announce that Gemini on GDC is going to be supported on NVIDIA Blackwell family as well as the NVIDIA Blackwell Ultra family. Some of you might know it as B200 and the B300 hardware accelerators.

This all comes together to deliver the Google distributed cloud AI engine. Customers now have the option of using our A4 compute instances along with the latest hardware accelerators from NVIDIA. We continue to add more and more innovations on data and storage with supporting up to 30 IOPS per gigabyte block storage. It meets your real-time AI inferencing needs. Now, this infrastructure is going to be powered with foundational models of Gemini as well as OSS models like Gemma.

It also comes with a AI inferencing service that we call as AI gateway that provides body-based routing to make sure that your infrastructure is running effectively. This foundation now allows you to build and run your sovereign agents and applications in both a connected as well as an air gapped environment. As you can see, we have covered a lot of ground today highlighting our most exciting infrastructure announcements.

I can't wait to see what you will all do with these innovations. And on that note, welcome to the era of agentic enterprise. Everything we talked about is life. Please feel free to check out the QR code. Thank you all for attending today. >> Hi, I'm Mandy and welcome to AI Boost Bites where every week we teach you the skills to supercharge your work with AI. We've all been there, presenting a slide with a bar chart, trying to explain what the data really means.

But what if the chart could tell the story for you? What if you could show the data changing and evolving over time? Today, I'll show you how to use Gemini Canvas to turn a static set of numbers into a fun animated bar chart race. Bringing your data to life and captivating your audience. Let's make some data dance. My goal is to visualize store sales data over several years. Instead of just showing a final chart, I want to show year-by-year changes in a more exciting way.

So, I'll ask Gemini to build it for me. Create a web app that shows the top countries with the biggest population increase over the last 50 years in a fun animated way. Use a bar chart where the bar race against each other to show how the rankings change over time. And here it is. Gemini Canvas built an interactive app that shows the bar racing against each other as time progresses. You can instantly see which country was leading in which year.

This is a much more powerful way to show trends >> [music] >> than a simple static image. Now, let's refine our story. The animation is a bit fast for a presentation. So, I'll ask, "Slow down the animation speed so I can talk through the changes each year." Perfect. Now, I can easily pause and discuss key moments. Creating data visualization like this often involves importing and referencing the data, and Canva can help make that an intermediate step in a larger goal.

Now, instead of a boring slide, I have a dynamic engaging visual that will keep my audience hooked. I can share a direct link to this app or embed it in my presentation. So, how can this change how you work with data? First, it turns your data into a compelling narrative. You can easily show growth, competition, and trends over time. Second, it makes your presentation more memorable. An animated chart is far more engaging than a static one.

And third, it helps you understand your own data better. Seeing the data evolve can reveal insights you might have otherwise missed. Ready to get started? Now is your turn. Earn your badge by completing the challenge. Thanks for watching and have fun. >> [music]