WhiteFiber announced R&D results for a distributed AI infrastructure architecture that connects geographically separated GPU clusters into a single logical supercluster. The company said its testing achieved 111.2 Tbps of bandwidth across 83 kilometers (51.6 miles) of dark fiber while maintaining a guaranteed 0.9 millisecond round-trip latency. WhiteFiber plans a commercial launch in Q3 2026 following additional testing across the full fiber spectrum. The project, internally known as Project Redwood, was developed with networking partner DriveNets and AI data infrastructure provider WEKA.
Rather than treating two facilities as independent GPU clusters connected through conventional data center interconnect (DCI), WhiteFiber’s architecture enables them to operate as a unified AI cluster. The approach addresses one of the industry’s emerging infrastructure challenges: limited power availability at individual data center campuses. By allowing GPU capacity to expand across multiple facilities, operators can build larger AI clusters while improving resiliency and meeting regional data sovereignty requirements. WhiteFiber said it has filed patent applications covering the implementation.
DriveNets supplied the Ethernet-based AI fabric interconnecting the two NVIDIA H200 GPU clusters, while WEKA provided the distributed data and memory infrastructure through its NeuralMesh platform. According to DriveNets, conventional DCI technologies struggle with synchronized AI traffic bursts generated during large-scale model training. Its AI Fabric employs Fabric Scheduled Ethernet (FSE), Virtual Output Queuing (VOQ), deep buffering, and cell-based load balancing to maintain lossless transport between sites while minimizing congestion. WhiteFiber said the architecture could also support future telecommunications, edge computing, and sovereign AI deployments beyond distributed GPU clusters.
• Demonstrated bandwidth: 111.2 Tbps across 83 km (51.6 miles) of dark fiber
• Guaranteed round-trip latency: 0.9 milliseconds
• Commercial launch targeted for Q3 2026
• Built around two geographically separated NVIDIA H200 GPU clusters
• DriveNets supplied the Ethernet AI fabric connecting both sites
• WEKA NeuralMesh provides distributed storage and memory infrastructure
• WhiteFiber has submitted patent applications covering the architecture
• Intended applications include distributed AI training, inference, sovereign AI, telecommunications and edge computing
“These results validate what we set out to prove: that geographic distance does not have to be a constraint on AI infrastructure. This 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.” — Sam Tabar, CEO of WhiteFiber
🌐 Analysis
The announcement represents one of the industry’s first public demonstrations of a commercial “scale-across” AI cluster rather than the more familiar “scale-up” architecture confined to a single data center. As AI clusters grow beyond 100,000 GPUs, power availability—not rack density or networking technology—is increasingly becoming the limiting factor. Vendors across the ecosystem, including NVIDIA, Broadcom, Arista Networks, Cisco, Nokia, DriveNets, and several optical networking suppliers, are developing architectures that extend AI fabrics across metro distances while preserving the low latency required for distributed training.
DriveNets’ role is particularly notable because the deployment relies on Ethernet rather than InfiniBand. The company has positioned Fabric Scheduled Ethernet as an alternative for large-scale AI networking, using advanced congestion management, deep buffering, and end-to-end scheduling to support lossless transport over metropolitan distances. The deployment also highlights the growing importance of coordinated advances across networking, storage, optical transport, and GPU orchestration as AI infrastructure evolves beyond the boundaries of individual data centers.
🌐 We’re tracking the latest developments in AI infrastructure. Follow our ongoing coverage at: https://convergedigest.com/category/ai-infrastructure/
| WhiteFiber Profile | |
| Date | July 9, 2026 |
| Headquarters | New York, New York, USA |
| CEO | Sam Tabar |
| Business | AI infrastructure, GPU cloud, HPC data centers, AI networking |
| Core Technology | Distributed GPU superclusters, high-performance AI cloud infrastructure, cross-data-center AI networking |
| Latest Milestone | Validated 111.2 Tbps across 83 km with 0.9 ms round-trip latency using a distributed GPU supercluster architecture. |
| Technology Partners | DriveNets (AI Fabric), WEKA (NeuralMesh), NVIDIA H200 GPUs |
| Target Markets | AI training, AI inference, hyperscalers, NeoClouds, sovereign AI, enterprise AI infrastructure |
| Commercial Availability | Q3 2026 (planned) |
| Key Industry Trend | Scale-across AI clusters overcoming single-site power constraints by treating multiple data centers as one logical GPU system. |