AI won’t wait for your infrastructure
Google Cloud has just published a new report, prepared together with GBK Collective based on responses from 1,402 technology leaders from 12 countries, and the conclusion is compelling: 83% of organizations acknowledge that their current infrastructure is not ready to support agentic AI in real production.
The reason? AI agents are no longer limited to answering questions: they investigate, decide, and execute tasks autonomously, chaining hundreds of actions in seconds. And that requires a very different infrastructure from what most companies have today.
These are the four keys of the study, highlighting what industry leaders are doing and what Google Cloud proposes in each case:
1. A new infrastructure standard
An agent needs persistent memory and continuous access to several data sources at once. In fact, inference already accounts for 47% of AI spending, ahead of model training.
Leaders are migrating to cloud platforms that integrate storage, computing, and AI tools in a single place. Sabre, for example, left 19 of its own data centers for the cloud and today saves $150 million a year with more than 70 AI solutions in production.
Google Cloud proposes an “all-in-one” infrastructure that combines hardware, development tools, and advanced models, to avoid wasting time connecting systems manually.
2. Without governance there is no control or security
More autonomy means more access to systems and data, and therefore more risk. Four out of five leaders point to security and governance as the biggest barrier to scaling AI.
For this reason, the majority consider it a critical requirement to centralize control in a single platform instead of scattered tools. Dun & Bradstreet, for example, uses centralized security tools to detect threats without slowing down its teams.
Secure AI Framework presents a different approach, building security within the agent itself from day one, not as an afterthought patch. Added to this are Model Armor and Security Command Center, two tools that work together so that any AI-related threat is detected and managed from the same dashboard, without relying on separate systems.
3. Hybrid and multicloud are now the standard
Neither everything in the cloud nor everything on-premises: model training needs the power of the cloud, but many real-time tasks cannot afford the round trip to a remote data center.
That is why hybrid adoption has risen from 41% to 52% in just one year. Recursion, for example, combines its own supercomputer with the Google cloud, cutting its costs in half.
Added to this is another increasingly decisive factor: digital sovereignty. Organizations are already prioritizing infrastructures that keep their data under the laws of their own country.
Google Cloud responds to this dual need with Google Distributed Cloud and Cross-Cloud Network, which allow operating without losing centralized control regardless of where the data is located.
4. Energy efficiency is already a condition to operate
91% of leaders already take energy consumption into account when choosing their hardware, pressured both by the limited availability of energy and by increasingly strict regulations.
Faced with this, many organizations choose to combine different types of hardware depending on the task they must solve. AXIA Energia, Brazil’s largest utility, is a good example: thanks to climate models in the cloud, it anticipates storms up to 10 days in advance and thus avoids supply cuts for 60 million people.
Google Cloud addresses this challenge with AI Hypercomputer, an infrastructure designed to maximize the performance obtained per unit of energy consumed.
In the era of agentic AI, it is no longer the one who trains the largest model that wins, but the one who builds the most resilient, secure, and well-integrated system. That is why organizations plan to invest heavily in infrastructure this year.
Because agents do not work like a traditional application: they need persistent memory and continuous access to several systems at once. Many current infrastructures are designed for occasional responses, not for this constant and chained activity.
Download the full eBook “State of Infrastructure in the Agentic AI Era” by Google Cloud and Luce for free to discover how to build the technological foundation that will drive your company tomorrow.
Frequently asked questions
Why might my current infrastructure, which works well today, not be enough for agentic AI?
Because agents do not work like a traditional application: they need persistent memory and continuous access to several systems at once. Many current infrastructures are designed for occasional responses, not for this constant and chained activity.
Is it necessary to choose between public cloud, private cloud, or on-premises?
No, and in fact the majority trend goes exactly in the opposite direction: combining several environments depending on the task. The public cloud provides power to train models, and on-premises or “edge” provides the speed needed for real-time tasks.
Does security slow down the speed of AI adoption?
According to the study, it is exactly the opposite: organizations with strong governance move faster and with more confidence. Without a clear security framework from the start, scaling AI becomes slower, more expensive, and riskier.
