Google’s Virgo Network: Inside the Megascale Fabric Wiring 1 Million TPUs Together

What Google Actually Announced at Next ’26

At Google Cloud Next 2026 in Las Vegas on April 22, Google’s networking team published the technical details of Virgo Network — a new megascale data center fabric purpose-built for AI training and inference at hyperscale. Three numbers anchor the announcement: 134,000 TPUs linked in a single fabric, more than 1 million TPUs across multiple sites stitched into one cluster, and up to 47 petabits per second of non-blocking bisectional bandwidth.

The architecture is a flat, two-layer non-blocking topology built on high-radix switches with multi-planar design and independent control domains. Compared to the prior generation, Virgo delivers 4x the bandwidth per accelerator and a 40% reduction in unloaded fabric latency for TPUs. It supports both Google’s own Ironwood (TPU 8t) silicon — capable of 121 exaflops in a single 9,600-chip superpod with 2 petabytes of shared memory — and NVIDIA’s upcoming Vera Rubin platform, with up to 80,000 Rubin GPUs per data center and 960,000 GPUs across multiple sites.

For context: a single Virgo fabric now wires together more chips than most public clouds run for AI workloads in their entire footprint two years ago. The “campus-as-a-computer” metaphor that Google used in 2023 has been quietly replaced by “globe-as-a-computer.”

Why a Flat Two-Layer Topology Matters

Most legacy data center networks use Clos or fat-tree topologies with three or more layers. Each extra layer adds latency, cabling complexity, and failure domains. Virgo’s two-layer non-blocking design is a deliberate engineering bet that high-radix switches — switches with hundreds of ports of equal bandwidth — let Google flatten the hierarchy without sacrificing scale.

The practical payoff for AI training is brutal in its simplicity. In synchronous data-parallel training, every gradient step is bottlenecked by the slowest tail of the all-reduce collective. Cut fabric latency 40% and you cut the gradient barrier 40%. Multiply across millions of gradient steps in a frontier-model training run and you save weeks of wall-clock time and tens of millions of dollars in idle accelerator hours.

Multi-planar design with independent control domains is the second engineering bet. By splitting the fabric into parallel planes that fail independently, Google reduces the blast radius of a single switch or controller fault — a direct response to the lesson the rest of the industry learned from Cloudflare’s November 2025 outage and AWS’s December 2025 us-east-1 incident: at hyperscale, blast radius is the metric that matters more than peak throughput.

How “1 Million TPUs as One Cluster” Actually Works

Stitching a million accelerators across multiple data centers into one logical training cluster is a problem that nobody had publicly solved before 2026. The bandwidth between sites is typically 100-1000x lower than intra-site bandwidth, and the latency is 10-100x higher. Naive multi-site training collapses into communication overhead.

Google’s three-layer architecture answers this by separating concerns. The scale-up domain handles intra-pod chip-to-chip communication using the Inter-Chip Interconnect (ICI) at 19.2 Tb/s for TPU 8i. The scale-out accelerator fabric is the east-west RDMA-based layer that Virgo Network actually targets — this is where most of the AI-specific bandwidth investment lives. The Jupiter front-end network handles north-south traffic for storage, ingress, and inter-zone connectivity.

The 4x bandwidth improvement is concentrated in the middle layer because that’s where modern training collectives spend most of their time. Google’s accompanying announcement of Cloud Managed Lustre at 10 TB/s storage bandwidth (a 10x year-over-year jump) closes the storage-side bottleneck so accelerators don’t sit idle waiting for shards.

The Competitive Bar This Resets

Microsoft’s Maia 100 cluster at Azure was reported to scale to roughly 100,000 chips per region as of late 2025. AWS Trainium2 UltraServers scale to 64 chips per node and clusters of “tens of thousands” per region. Meta’s Grand Teton clusters target 24,000 GPUs. Against this baseline, Virgo’s 134,000 TPUs in a single fabric and 1M+ across sites is roughly an order of magnitude ahead on the cluster-size axis.

The Vera Rubin numbers — 80,000 GPUs per site, 960,000 across sites — also signal something subtler. Google is positioning Virgo Network as cloud infrastructure that runs both Google silicon and NVIDIA silicon at the same scale. This matters because the mid-2026 GPU constraint is no longer raw die supply (TSMC has caught up) but the networking and power infrastructure that lets you actually use the chips. Customers locked into NVIDIA roadmaps but unhappy with their current cloud’s networking now have a credible Google alternative.

The third-order effect is on power and water. A 960,000-GPU site at Vera Rubin’s expected 1.5-2 kW per accelerator implies 1.4-1.9 GW of IT load — larger than the entire grid draw of several mid-size African countries. Site selection, water rights, and grid interconnect timelines now constrain Virgo’s actual rollout more than the silicon does.

What This Means for Cloud Buyers in Africa and the Middle East

1. Stop comparing AI clouds on per-GPU price — start comparing on cluster topology

Most procurement RFPs in Algeria, Morocco, the GCC, and East Africa still ask for “$/A100-hour” or “$/H100-hour” as the headline metric. Virgo makes that comparison meaningless for any workload above a few thousand chips. A 64-GPU job will run roughly the same anywhere; a 16,000-GPU job will run 30-50% faster on Virgo because of the latency reduction, and a 100,000-GPU job won’t run elsewhere at all. Rewrite your scorecard to weight three things over price: maximum single-job cluster size, intra-site fabric bandwidth, and cross-site clustering capability. If your sovereign cloud strategy precludes Google, document the size ceiling you accept.

2. Negotiate “fabric SLA” terms, not just availability SLAs

Standard cloud SLAs cover compute and storage availability — they say nothing about fabric latency or bandwidth-degradation events. With Virgo-class infrastructure, a 10% bandwidth degradation on the east-west fabric can crater training throughput while every dashboard still reports “available.” Push your account team for fabric-level metrics in the SLA: p99 east-west latency, percent of bisectional bandwidth available, time-to-detect for fabric incidents. Google has these numbers internally; ask them to expose them as customer-visible SLOs.

3. Plan for the egress-tax shift before you sign multi-site training agreements

Multi-site training across a 1M-TPU cluster generates colossal cross-region traffic. Today’s cloud egress pricing assumes you train in one region and serve in another — multi-site training rewrites that assumption. Verify with your vendor whether cross-site fabric traffic counts as “egress” (priced per GB) or “internal” (free). The same workload billed as egress vs. internal can vary by a factor of 100x in monthly cost. Get the answer in writing before committing to a multi-site training architecture, because the line between “fabric” and “egress” is currently being drawn in real time across all three hyperscalers.

4. Use Virgo as leverage in Vera Rubin allocation talks with NVIDIA partners

If your shortlist includes an NVIDIA-only path (CoreWeave, Lambda, sovereign-cloud H200/Rubin builds), Virgo’s 80,000-Rubin-per-site capacity is leverage. NVIDIA’s Vera Rubin allocation in 2026-2027 will be tight; Google’s announced numbers force every NVIDIA-only provider to either match or explain the gap. Use that publicly disclosed ceiling as the floor in your allocation negotiations.

The Sovereignty Question Virgo Forces

Virgo’s “globe-as-a-computer” framing collides head-on with the sovereign-cloud regulations now spreading across Africa, the Gulf, and Europe. Algeria’s Law 18-07 mandates local hosting of personal data. The UAE’s Federal Decree-Law 45 of 2021 imposes similar restrictions. The EU’s pending Data Act fragments cross-border data flows further.

A 1M-TPU multi-site cluster only delivers its advertised performance if data and gradients can flow freely across sites. The moment a customer’s data is pinned to a single jurisdiction, the cluster effectively shrinks to whatever lives in that jurisdiction’s data centers. For Algeria, where Google has no domestic region, that means Virgo Network is functionally a 0-TPU offering for any workload subject to Law 18-07. The same is true for any jurisdiction without a Google AI region.

This is the structural lesson: hyperscale AI infrastructure and data-sovereignty law are now in direct collision, and Virgo Network is the clearest expression yet of what’s at stake. Customers in regulated jurisdictions will face a choice between training on the world’s largest cluster (and accepting a sovereignty hit) or training on a domestically compliant cluster that is one to two orders of magnitude smaller. There is no middle path that the network engineering can paper over.

For Algerian CTOs evaluating frontier-model training in 2026-2027, the practical move is to separate workloads by sensitivity tier: regulated personal data stays on a domestic compliant region (Oracle, Microsoft sovereign, or future ATM Mobilis cloud builds), while non-sensitive research and pre-training workloads can use a Virgo-class cluster abroad. The sooner that bifurcation is engineered into the data architecture, the cheaper it is to maintain.

Frequently Asked Questions

Sources & Further Reading

• Introducing Virgo Network: Megascale Data Center Fabric — Google Cloud Blog
• AI Infrastructure at Next ’26 — Google Cloud Blog
• Google Cloud Next ’26 Recap — Google Blog
• Google Cloud Unveils Virgo Network to Power Next-Generation AI Data Centers — TechAfricaNews
• Google Cloud Next ’26: Gemini Enterprise Agent Platform Leads AI-Centric News — Virtualization Review