Google's Single-Datacenter Problem
A 15-hour power outage at Google Cloud last week didn't bring down the whole zone. It brought down three services.
The culprit? An upstream electrical fault that cascaded into cooling failure inside a single datacenter serving Google's Europe-West4-A region. The victims: VMware Engine (GCVE), Bare Metal Solutions (BMS), and NetApp Volumes. All three. All offline. All relying on infrastructure that Google uses a discrete datacenter for.
The rest of the zone kept humming. The rest of the region stayed up. But if your workload lives on one of those three services, you were on your own for half a day.
As organizations accelerate their adoption of cutting-edge computing and autonomous systems, the assumption that cloud abstraction automatically handles facility-level redundancy is becoming a dangerous oversimplification.
Why the Hidden Dependencies Matter
Google's incident report was refreshingly blunt. An electrical fault occurred on the utility grid upstream of the datacenter, disrupting electrical distribution gear and cooling equipment. Google hasn't explained how an upstream failure caused that disruption, but they did say they "proactively turned down workloads in order to protect customer data from any risks posed by running infrastructure in a high temperature environment."
Translation: the cooling failed, the temperature rose, and Google killed the workloads to prevent hardware damage. It's a reasonable call from an engineering standpoint. It's a terrible one from a customer standpoint.
We've asked Google if they had generators or other energy sources at this site, and if so, why workloads still had to be turned down. No response at the time of writing. Google told us their incident analysis "is currently ongoing" and promised a follow-up once available.
The real issue, according to analysts, is transparency. Customers are generally told to use multiple zones and regions for resilience. They're rarely given visibility into whether a particular managed service has a single-datacenter dependency within a zone.
"The underlying architecture is not necessarily unusual," said Biswajeet Mahapatra, principal analyst at Forrester. "AWS, Azure, and Google all operate services that rely on dedicated hardware, storage platforms, or tightly coupled infrastructure that may not be distributed across multiple facilities in the same way as core compute and storage services."
Evaluating AI Cloud Infrastructure Companies in India for Critical Workloads
This lack of transparency makes due diligence increasingly difficult, especially for organizations trying to build and scale autonomous systems. When assessing AI cloud infrastructure companies in India or globally, simply looking at region/zone availability is no longer enough.
Organizations must demand concrete documentation on the physical distribution of specialized services. Does the service rely on tightly coupled hardware or distinct storage platforms that cannot, by design, be distributed across multiple facilities? If the answer is "yes," you must architect your application layer—not the infrastructure layer—to account for that single point of failure.
This isn't theoretical. Gartner Director Analyst Adrian Wong reminded The Register of the 2023 outage at Google Cloud's europe-west9-a region, caused by a water leak originating in a non-Google portion of the facility. Google uses a tool called "Spanner" to replicate data across zones, but in the flooded zone, Spanner's configuration didn't work once one building became unavailable.
"It is very hard to figure out how an individual region is architected," Wong said. "Our customers are often surprised by that."
Understanding Agentic AI and Embodied Agents
The complexity of these resilience regimes matters because the workloads themselves are becoming more autonomous. As we move deeper into this era of AI-driven operations, distinguishing between simple automation and true intelligence is critical.
What is agentic AI?
According to industry benchmarks from IBM and Google Cloud's own definitions, agentic AI marks a shift from passive, predefined automation to active task completion. Unlike standard automation, which follows a rigid, deterministic path, agentic AI systems are designed to reason, plan, and utilize tools to achieve a given objective. They operate with a degree of autonomy, making decisions based on intermediate feedback loops.
Google Cloud differentiates this by highlighting that agentic AI can interact with tools and maintain context—a level of adaptability that makes resilience even more important. A system that is constantly planning and executing in real-time cannot afford the "black box" of infrastructure downtime.
What is an embodied agent?
An embodied agent takes this autonomy further, possessing a physical (or simulated) presence that interacts directly with the world. Through sensors and actuators, an embodied agent perceives its environment, reasons, and executes physical or digital actions. Whether the body is a robotic arm, an autonomous vehicle, or a complex simulation hub, the infrastructure supporting these agents requires consistent, low-latency, and highly reliable state maintenance.
Any interruption, like the single-datacenter failure described earlier, doesn't just stop a server. It interrupts the agent's ability to interact and "see" its environment.
The Road Ahead
The reality is that hyperscalers will continue to operate services that rely on dedicated, tightly coupled infrastructure. That's not going to change. The responsibility, therefore, shifts back to the customer: to dig deeper into the architectural promises, interrogate vendor incident reports, and explicitly account for single-facility dependencies in the overall design.
Knowing how Google—or any other vendor—plans to prevent future incidents is, by itself, insufficient. Customers need to know which architecture choices were made to optimize for service capability and performance, and how those choices trade off against traditional facility-level redundancy.
As we build AI-centric systems, our reliance on the underlying layer has never been greater, and our due diligence must evolve to match that dependence.
If you're looking for more context on how cloud outages are reshaping infrastructure strategy, check out Cloud Outages Just Got Personal: What AI and Cloud Infrastructure Companies in India Need to Do Now, which covers broader patterns we're seeing across the industry.
The bottom line? When your infrastructure is supposedly in a "zone," you expect a level of redundancy that spans physical facilities. But sometimes, a zone is just one building with a power problem.
