# How Google’s Virgo Fabric Signals Shift in AI Network Design — retained record

> Retained news record from news:datacenterknowledge.

Canonical: https://highsignal.app/data/records/983fa615f5bbe78c

[← sources](https://highsignal.app/data)


source: [news](https://highsignal.app/data/news) (news:datacenterknowledge)  published 2026-04-27 16:00 ingested 2026-04-28 08:15

[original source ↗](https://www.datacenterknowledge.com/infrastructure/google-s-virgo-fabric-signals-shift-in-ai-network-design)

## extract

How Google’s Virgo Fabric Signals Shift in AI Network Design
A flatter topology and higher bandwidth reflect how hyperscalers are reshaping networks for large-scale AI clusters.
Google’s new Virgo data center fabric reflects how AI workloads are reshaping network design in hyperscale environments.
In a company blog post, Google describes Virgo as a fabric built for large AI clusters – spanning tens of thousands of accelerators – within its AI Hypercomputer architecture. The design uses a flatter, two-layer topology to cut latency and increase bandwidth across the fabric.
The shift underscores a broader change: AI training and inference demand consistent performance at scale, not just peak throughput. These synchronized workloads exchange data continuously across high-volume east-west paths; each step depends on tight coordination, and a single lagging node can slow the entire job.
Ron Westfall, vice president and analyst at HyperFrame Research, said Google is treating that variability as a system-level risk rather than a networking side effect.
“Google’s transition to the Virgo fabric reimagines the data center as a Campus-as-a-Computer,” he said, adding that the design “treats tail latency as a critical hardware reliability issue” and isolates AI training traffic to keep large clusters synchronized.
Google said Virgo is designed to scale to clusters of more than 100,000 accelerators, with a focus on maintaining high bisection bandwidth and minimizing latency across the system. The company also emphasized resilience. The fabric uses multiple independent switching planes and deep telemetry to detect congestion or failures and reroute traffic without interrupting workloads. At this scale, localized failures are expected. The design goal is to prevent those disruptions from propagating across the cluster.
Fewer Layers, Less Variance
Traditional data center networks rely on multi-tier Clos architectures and oversubscription to balance cost and utilization. AI workloads break that model. They generate sustained east-west traffic, keeping links busy and exposing contention points. Google said Virgo replaces traditional three-tier designs with a two-layer fabric, reducing hop count between nodes and limiting the opportunity for queuing delays.
Sameh Boujelbene, vice president at Dell’Oro Group, said topology plays a direct role in controlling latency.
“Flattening reduces hop count and creates more direct, predictable paths between accelerators,” she said, “which is critical for synchronized workloads.”
Westfall noted that the design targets variability more than raw congestion. “Flattening reduces the cumulative probability of queuing delays at intermediate hops,” he said, helping prevent synchronized workloads from stalling on a single delayed packet. Still, at larger scales, Westfall added, flattening alone isn't sufficient. Systems depend on traffic distribution and optical interconnects to prevent congestion from concentrating as networks simplify.
A Segmented Data Center Fabric
Virgo also operates within a broader, segmented architecture that separates traffic types inside the data center. Google distinguishes between tightly coupled accelerator communication, large-scale east-west traffic across clusters, and north-south traffic connecting to storage and external services.
That segmentation reflects a move away from general-purpose networking toward fabrics tuned for specific workload patterns. It reinforces a larger shift: the data center network is no longer a single, uniform system but a set of coordinated layers aligned to how AI workloads execute.
The same pressures are shaping vendor offerings aimed at AI clusters. Nvidia is positioning its Spectrum-X platform as an Ethernet-based AI fabric that combines switches and DPUs to manage congestion and maintain consistent performance across GPU clusters.
Meanwhile, Broadcom continues to supply high-radix switching silicon, including its Tomahawk and Jericho lines, which underpin many large-scale Ethernet fabrics. Arista Networks is focusing on AI backend networks with software tied to traffic management, telemetry, and load balancing across distributed clusters.
Westfall said designs like Virgo raise expectations for how those platforms perform. “The shift reinforces the need to treat tail-latency consistency as the primary success metric,” he said, pointing to high-radix switching and tighter hardware-software integration to keep workloads in sync.
He added that hyperscalers retain an advantage by co-designing infrastructure across compute, networking, and software layers.
Boujelbene said vendors are advancing those capabilities but face structural limits.
“Vendors can deliver better congestion control, telemetry, and tighter switch-to-NIC integration,” she said. “But hyperscalers still have an advantage because they control the full stack.”
A New Direction of Travel
Virgo is specific to Google, but the pattern is broader. Hyperscalers are flattening topologies, increasing path diversity, and tying network behavior more closely to workload demands. Google’s design treats the data center as a unified system, where compute, storage, and networking operate as a coordinated fabric rather than independent components.
Westfall said that the level of coordination remains difficult for vendors to replicate.
“Hyperscalers co-design the network with the AI stack, treating the data center as a single, software-defined computer,” he said. “That’s a level of integration general-purpose vendors can’t fully package.”
Boujelbene said that the gap is likely to persist as AI systems scale.
“Hyperscalers design across the entire system – from accelerators to software and operations,” she said. “That level of integration is difficult to replicate in a general-purpose environment.”
As AI systems continue to scale, these integrated design choices are poised to redefine data center networks – starting with hyperscale environments and eventually influencing enterprise deployments.

## retained text

No document body was retained for this record — only its metadata was stored.