Google publicly targets the 'agent-type AI' super-large data center network 'Vergo'…… enhancing latency and fault response capabilities

Google has unveiled ultra-large-scale data center network systems and cross-cloud infrastructure for the era of “Agentic AI.” As environments where artificial intelligence calls external tools in milliseconds become increasingly common, competition over infrastructure to reduce latency and reliably process large-scale computing is unfolding across the board.

On the 24th, Google released its new AI infrastructure network system, “Virgo Network.” The system is designed to boost the overall data center communication speed not only within accelerator clusters, but also in the back-and-forth communications among memory, compute, and storage resources.

The key is a “flattened” network architecture. This approach lowers bottlenecks by reducing the number of network layers the data passes through during transmission. According to Google, the Virgo Network can connect up to 134,000 chips, including the 8th-generation TPU 8t processors used for training. Its bidirectional bandwidth can reach up to 47 petabits per second (Pbps). The company states that, compared with the previous generation, bandwidth per accelerator has increased by more than four times.

A particularly notable part of this release is that, besides straightforward speed competition, it places comprehensive emphasis on “resilience.” In ultra-large-scale AI clusters, failures, latency issues, and performance slowdowns of some devices are nearly unavoidable. Google says that to address this, they combine a “visibility” function that carefully monitors the overall network status with automatic rerouting/recovery software. Its features also include applying an independent switching plane to ensure that overall throughput does not sharply fluctuate even when network latency or failures occur.

Google describes the Virgo Network not as a simple expansion of existing data centers, but as an independent infrastructure product, with a design philosophy of “designing the entire campus like a supercomputer.” The company argues that by monitoring system status in increments of less than 1 millisecond, it can optimize issues related to instantaneous congestion and buffer management across the full lifecycle of both hardware and software. This is especially important in environments where agent AI needs to handle tool calls, reasoning, and retrieval-augmented generation (RAG) tasks at the same time.

Also released: a connection and security layer that breaks down cloud-edge boundaries

In addition to the data center network, Google also released a cloud connection and security layer tailored to agent AI workloads. The company summarizes this update into four pillars: “Elastic Computing,” “Secure Cross-Cloud Connectivity,” “Unified Data Layer,” and “Digital Sovereignty.”

First, “Elastic Computing” is an architecture intended to more efficiently handle fluctuations in AI agent (Agent) demand. AI services often experience sudden surges in requests. Google explains that by increasing the accessibility of the CPU, they can respond to this demand faster and more economically. In particular, its design provides CPU resources optimized for tasks such as inference, agent orchestration, and retrieval-augmented generation, to complement shortages of graphics processing units (GPUs).

To that end, Google has applied C4N and M4N CPUs to virtual machines (VMs) based on Google Compute Engine and Google Kubernetes Service. Google states that this system can handle up to 95 million packets per second, outperforming major hyperscale cloud providers by up to 40%.

In the area of secure cross-cloud connectivity, the “Agent Gateway” is positioned at the core. This controller is responsible for monitoring AI agents’ access permissions and fundamentally controlling and managing protocols such as the Model Context Protocol (MCP) and the agent-to-agent protocol (A2A). Its explanation notes that in a multi-cloud environment, the gateway can ensure visibility and protection for data flows moving across different networks.

“AI goes where the data is,” rather than “moving the data”

The unified data layer focuses on enabling AI to directly understand and utilize data that enterprises have distributed across multiple storage repositories. Google says that “intelligent storage” injects metadata into data objects, thereby transforming traditional “static data” into AI-readable knowledge assets.

Once this architecture is established, it can apply semantic search to information in various formats—such as spreadsheets, documents, PDFs, and images—and enable automatic annotation and insight extraction. This approach aims to reduce the “island” phenomenon where data is trapped in specific repositories, helping AI agents find the information they need more quickly.

A “Knowledge Directory” released alongside it visually connects enterprise knowledge, helping AI agents better understand business processes and context. Google emphasizes that with this approach, optimization for AI learning and response can be achieved without separately moving data. In other words, instead of moving data to a central location, it makes the AI model run within the “private environment” where the data resides.

This release shows that the focus of AI competition is rapidly shifting from model performance to the network, security, and data infrastructure that supports running models. In particular, with the formal introduction of “agentic AI” into enterprise business operations, low latency, high resilience, and multi-cloud control are very likely to become core competitive advantages. Google’s move has been interpreted as a signal that competition among major technology companies for dominance in AI infrastructure is intensifying.

TP AI notes: This article uses a language model based on TokenPost.ai for summarization. The main content may be omitted or may not match the facts.