Last updated: 

July 2026

WhiteFiber announces cross-data-center networking solution that will transform the world of AI compute

WhiteFiber's cross-data-center networking solution links two data centers 83km apart into one GPU supercluster: 111 Tbps and 0.9ms latency. See how it works.

News

Lorem ipsum dolor sit 1

Distance has always been a tax on distributed compute. Split a workload across two facilities and you inherit latency, jitter, and a system that behaves like two systems pretending to be one.

‍

We just ran the numbers that say otherwise.

‍

In R&D testing, WhiteFiber's cross-data-center networking solution connected two geographically separated data centers into a single logical GPU supercluster, operating as one unified system rather than two linked environments.

‍

The result: 111 Tbps across 83km of dark fiber, already roughly double the capacity of comparable published full-spectrum field trials, using only a portion of available fiber spectrum. Guaranteed round-trip latency of 0.9ms sits within 8% of the physical limit for light in fiber over that distance. Full-fiber lighting testing is planned prior to commercial launch later this quarter.

‍

‍

‍

Why this matters

‍

Most infrastructure treats a data center as the boundary of a cluster. This new solution treats it as an implementation detail. That distinction matters for the workloads that don't fit inside one facility's power envelope, one site's compliance boundary, or one region's blast radius. Scale, resilience, and compliance stop being tradeoffs against each other.

‍

"These results validate what we set out to prove: that geographic distance does not have to be a constraint on AI infrastructure. It is the foundation for a new class of AI compute, one that delivers the performance of a single supercluster with the resilience and flexibility of a distributed system. We are excited to bring this capability to market."

Sam Tabar, CEO of WhiteFiber

‍

Getting there took precision, not a shortcut.

‍

"Achieving 111.2 Tbps with sub-millisecond latency across 83 kilometers is the result of months of precise, disciplined engineering. Every element of this system was designed and validated to perform at this level."

Tobias Ford, Principal Design Engineer at WhiteFiber

‍

‍

‍

Built with the right partners for the job

‍

WhiteFiber's cross-data center networking solution was built in collaboration with DriveNets, which provides the high-performance network fabric connecting both sites, and WEKA NeuralMesh, which provides high-performance data and memory infrastructure across the cluster. Networking, storage, and compute engineered together, not stitched together after the fact, which is the same principle behind everything else we build.

‍

‍

‍

Where this goes next

‍

The architecture was proven for GPU superclustering, but the applications extend further: telecommunications, edge computing, sovereign AI, and anywhere geography and data residency currently force a choice between scale and control.

‍

We'll share more on availability, architectural design, and pricing alongside the commercial launch later this quarter.

‍

‍

Want to talk through what a distributed supercluster looks like for your environment?

Let's connect

‍

‍

FAQs

‍

What is WhiteFiber's cross-data center networking solution?

‍

It is WhiteFiber's distributed GPU supercluster architecture that links two geographically separated data centers into a single logical system rather than two connected environments. In R&D testing it reached 111 Tbps across 83km of dark fiber at 0.9ms guaranteed round-trip latency, with commercial launch planned later this quarter.

The idea is to treat the data center boundary as an implementation detail rather than the edge of a cluster, so workloads that exceed one site's power envelope, compliance boundary, or regional blast radius can still run as one system.

‍

‍

Can you run a single GPU cluster across two separate data centers without a latency penalty?

‍

Yes, and that is what WhiteFiber's cross-data center networking solution demonstrated. It connected two data centers 83km apart into one unified GPU supercluster with guaranteed 0.9ms round-trip latency, within 8% of the physical limit for light traveling through fiber over that distance. The two sites operate as one logical system, not two linked environments.

Distance normally adds the latency and jitter that make split workloads behave like two systems pretending to be one. These results indicate geographic separation does not have to be a hard constraint on AI infrastructure performance. Full-fiber lighting testing is planned before commercial launch.

‍

‍

What bandwidth and latency did WhiteFiber's cross-data center networking solution achieve?

‍

It reached 111.2 Tbps across 83km of dark fiber while using only a portion of the available fiber spectrum, roughly double the capacity of comparable published full-spectrum field trials. Round-trip latency was a guaranteed 0.9ms, within 8% of the theoretical limit for light in fiber over that distance.

Because those numbers came from only part of the spectrum, there is meaningful headroom before full-fiber lighting testing, which is scheduled ahead of commercial launch later this quarter. The system was engineered with DriveNets for the network fabric and WEKA NeuralMesh for data and memory infrastructure.

‍

‍

What can WhiteFiber's cross-data center networking solution be used for?

‍

It was proven for GPU superclustering, running AI training and inference workloads that do not fit inside a single facility's power envelope, compliance boundary, or regional blast radius. The same architecture extends to telecommunications, edge computing, and sovereign AI, anywhere geography and data residency otherwise force a tradeoff between scale and control.

For regulated and sovereignty-sensitive buyers, that means scale, resilience, and compliance stop competing with one another. WhiteFiber plans to share availability, architectural detail, and pricing alongside the commercial launch later this quarter.