Google Unveils Virgo Network to Power the Next Decade of Mega-Scale AI Training

Building the Campus as a Computer

Google announced Virgo Network at Google Cloud Next '26, representing a major shift in how AI infrastructure is built. This specialized data center fabric moves beyond general purpose networking to create what Google calls a campus-as-a-computer architecture. It serves as the primary backbone for the updated AI Hypercomputer, specifically designed to handle the explosive growth of foundation model parameters.

Unprecedented Scale and Speed

The technical specifications of Virgo are staggering, supporting up to 134,000 chips within a single data center. It delivers 47 petabits per second of non-blocking bandwidth, which is a fourfold increase over previous generations. When combined with Google’s software orchestration, the network can scale a single training cluster to over one million chips across multiple global sites.

Architected for the Agentic Era

This new fabric is essential for the transition toward agentic AI workflows. Unlike simple chatbots, these advanced agents require continuous multi-step reasoning and massive real-time data throughput. By using a flat, two-layer topology and high-radix switches, Virgo minimizes latency and ensures that communication bottlenecks do not stifle the performance of the latest TPU 8t and 8i accelerators.

Resilience and Goodput

Reliability remains a core focus, as a single failing component in a 100,000-chip cluster can halt an entire training run. Virgo utilizes a multi-planar design that provides robust fault isolation and deep telemetry. This allows the system to automatically detect and reroute traffic around degraded nodes, aiming for a productive computing goodput exceeding 97 percent for the most demanding workloads.

Broad Ecosystem Support

While optimized for Google’s custom silicon, Virgo also offers significant support for industry-standard hardware. It is fully compatible with Nvidia’s Vera Rubin architecture, allowing for clusters of up to 80,000 GPUs in one location. This flexibility ensures that enterprise customers can leverage Google’s cutting-edge networking solutions regardless of their specific choice in AI accelerators.