WhiteFiber claims to have connected two data centers separated by 83 kilometers to operate as a single logical GPU cluster. Initial R&D testing achieved an aggregate bandwidth of 111.2 Tbps over dark fiber and round-trip latency of 0.9 milliseconds, according to the company’s published data.
The result points to a solution for one of the main limitations of large-scale artificial intelligence infrastructure: finding enough power, cooling, and space in a single location. However, the company has not yet published the technical details and load tests needed to verify how much of that optical performance actually reaches distributed training applications. Commercial launch is scheduled for Q3 2026.
The key points of WhiteFiber’s distributed cluster in 20 seconds
- The test connected two data centers 83 kilometers apart.
- WhiteFiber reports an aggregated bandwidth of 111.2 Tbps.
- The guaranteed round-trip latency was 0.9 milliseconds.
- The connection uses dark fiber and only a portion of its available spectrum.
- The company plans to light up the full spectrum before launch.
- DriveNets provides the Ethernet network fabric between sites.
- WEKA NeuralMesh supplies the data layer and memory.
- WhiteFiber aims to operate both centers as a single logical GPU cluster.
- Commercial availability is planned for Q3 2026.
- Details such as the number of GPUs, their models, or training efficiency have not been disclosed.
- No independent results currently validate these claims.
- Patent applications are still pending and do not guarantee approval.
The 111.2 Tbps equals a theoretical maximum of about 13.9 terabytes per second. This is a significant capacity for metropolitan interconnection, but should be understood as the aggregate bandwidth of the infrastructure—not the per-GPU available speed or the effective AI training performance.
Distance no longer a physical limit, but it persists in software
WhiteFiber presents its architecture as a way to overcome the traditional data center boundary. Instead of building the entire cluster within a single facility, the company distributes computation across two locations and uses a high-capacity network to make them function as one system.
This approach responds to a growing challenge. Large AI clusters are not solely limited by how many GPUs a company can buy. They also require electric power, transformers, cooling, space, connectivity, and permits. Gathering all these resources into one data center can take years or be impossible at a specific site.
Splitting infrastructure allows access to two power reserves and expands capacity without waiting for a single site to host more racks. It also makes it easier to segregate failure domains and keep certain datasets at their origin location.
The challenge arises when a task must use GPUs in both centers. Distributed training exchanges gradients, parameters, and states frequently. While fiber speed prevents bandwidth from being an obvious bottleneck, latency still impacts synchronization time.
The physical fiber path explains much of the 0.9 milliseconds latency. Light travels through fiber at about 200,000 km/sec. Covering 83 km, returning, and traversing network equipment leaves little margin to reduce this figure. WhiteFiber claims its result is roughly 8% above the physical limit of the link’s path.
Being close to that minimum doesn’t mean the distance has no consequences. A delay of 0.9 ms equals about 900 microseconds—far above typical intra-data center latency times. In repeated collective operations, each wait can accumulate, keeping expensive GPUs idle without productive work.
Research shows that wide-area networks for training cause idle cycles even when using specialized techniques to share bandwidth and organize communication. Experimental systems like ATLAS significantly improve times over conventional setups but require tailored training planning aligned with network behavior.
Therefore, the key metric isn’t only optical capacity. WhiteFiber must demonstrate what models can be trained across sites, what parallelism strategies are used, and how much efficiency is retained compared to running the same job on a single data center.
111.2 Tbps alone doesn’t prove a GPU supercluster
WhiteFiber states that the two centers operate as a single supercluster, not as two connected environments. While this distinction is important, it has not been documented in enough technical detail.
The company has not disclosed how many GPUs participated in testing, what generations they were, or how they were distributed. Nor have they published results of collective operations such as AllReduce, AllGather, or All-to-All, which are essential to evaluate distributed training behavior.
Additional data on application performance are also missing. A transport test can saturate links with synthetic traffic near nominal capacity. A real model introduces different communication patterns, synchronization barriers, varying transfer sizes, and periods when the link isn’t fully utilized.
Indicators to assess the solution include scalability efficiency, tokens processed per second, GPU utilization, training step times, and how these compare to an equivalent cluster in a single site.
Stability metrics are also crucial: latency variation, packet losses, retransmissions, behavior during fiber failures, and re-routing times. Low average latency can coexist with spikes that impact synchronous tasks.
WhiteFiber claims both sites carry live traffic simultaneously, and if one fails, the other continues without switch-over delays. While the network can keep the second site operational, maintaining an ongoing training process also depends on software, checkpoints, and resilience to reduced GPU counts.
If a run uses resources from both centers and one fails, merely keeping the remaining network active isn’t enough. The system must remove lost processes, rebuild communication groups, and recover the model state—highlighting the difference between infrastructure continuity and application continuity.
DriveNets for networking and WEKA for data management
WhiteFiber developed the solution with DriveNets and WEKA. DriveNets supplies the high-performance Ethernet fabric connecting the data centers, while WEKA NeuralMesh provides the data infrastructure and memory for the cluster.
Choosing Ethernet reflects a broader trend in AI hardware. Leading operators seek open fabrics that can grow with equipment from different vendors and leverage optical telecom technologies.
Nonetheless, Ethernet must be much more predictable than conventional enterprise networks. Distributed training generates massive, concurrent data flows and can suffer congestion when many servers communicate with the same destinations. Queue management, load balancing, and reacting to degraded links are as important as nominal speed.
The storage layer also plays a vital role. The two centers need access to datasets, checkpoints, and training states without creating additional bottlenecks. WhiteFiber discusses distributed data and memory infrastructure but has yet to specify what information is replicated, how consistency is maintained, or what performance is achieved reading and writing across sites.
The company states that sensitive data can stay at its source site, with only gradients traversing the link. This organization can help meet data residency requirements but does not guarantee privacy or compliance by itself.
Various studies have shown that gradients can, under certain conditions, leak training data, or recover images or text snippets. Using this architecture for healthcare or finance data would require additional controls, such as encryption, isolation, secure aggregation, data minimization, and legal assessments.
Merely storing raw data locally doesn’t automatically ensure compliance with regulations like DORA or industry standards. Compliance depends on operators, where derived data is processed, who accesses it, infrastructure auditability, and recovery procedures.
An energy solution for large clusters
WhiteFiber’s approach is particularly relevant given the context. The demand for AI infrastructure is rising faster than available power capacity in some markets. Using fiber to connect multiple sites creates a larger logical capacity without concentrating all power in one location.
This model enables utilization of facilities with available energy but insufficient scale to compete independently. It also offers a way to expand clusters without rebuilding or waiting years for new electrical connections.
Telecom operators could benefit from this evolution. They have metro fiber, distributed sites, and experience with optical networks. If compute resources can be integrated with acceptable performance loss, some network hardware assets may be repurposed as AI infrastructure nodes.
The approach aligns with technological sovereignty projects. An organization could keep datasets in different regions and collaborate on training without transferring original files. While architecture alone doesn’t ensure legal compliance, it offers more options than centralizing all data at a single site.
Resilience is another advantage. Two centers reduce dependence on a single substation, cooling system, or building. However, interconnection introduces potential critical points—unless optical routes and equipment are fully redundant and independent.
WhiteFiber has not specified how many physical paths are used, whether the 83 km represents a single route or multiple, or how the cluster would behave after a simultaneous cut. Such details will be necessary to assess claims of availability and capacity continuity.
Commercial launch needs more comprehensive proof
WhiteFiber plans to light the full spectrum of fiber before market entry in Q3 2026. They also intend to publish details on architecture, availability, and pricing at that time.
The initial test proves that transmitting over 100 Tbps between two centers dozens of kilometers apart with near-physical-limit latency is possible. However, claiming that both form a single supercluster requires further evaluation.
Confirmed validation will require real GPU tests, comparisons to a single-site setup, and protocol details. Cost analysis for fiber, optics, networking gear, and redundancy capacity to maintain resilience will also be critical.
Economics can be as crucial as performance. Lighting dark fiber with dozens or hundreds of channels involves transceivers, amplification, maintenance, and specialized personnel. The solution will make sense when it’s more cost-effective or faster than expanding a single site, not solely when it’s technically feasible.
The 111.2 Tbps figure indicates optical transport can cease being the main obstacle in metropolitan distances. The key question remains: can training software, data layers, and daily operations exploit this bandwidth without significantly sacrificing GPU performance?
WhiteFiber has put forward a credible approach to extending AI factories beyond a single building. The Q3 launch must demonstrate that this architecture works under real workloads and not just as laboratory-scale interconnection.
Frequently Asked Questions
What has WhiteFiber achieved in its tests?
The company reports 111.2 Tbps of aggregate bandwidth and 0.9 ms round-trip latency between two data centers 83 km apart.
Are the two centers already functioning as a single commercial supercluster?
WhiteFiber claims to have validated the architecture during R&D tests, but commercial service will not be available until Q3 2026.
Is 111.2 Tbps the speed of each GPU?
No. This is the aggregate optical interconnection bandwidth. Each GPU’s performance depends on its number, internal network, and workload type.
Does storing data locally at each site ensure privacy?
Not necessarily. Location helps, but gradients and other derived data can still leak sensitive information. Additional technical, organizational, and legal controls are needed.
