By Yohai Schwiger

Israeli networking company DriveNets has announced the first commercial deployment of its Scale-Across architecture, a technology that enables geographically separated data centers to operate as a single AI cluster. According to the company, this is the world’s first commercial implementation of the concept.

The deployment is part of Project Redwood, an initiative by U.S.-based AI infrastructure provider WhiteFiber to build distributed AI infrastructure. The project connects two data centers located approximately 84 kilometers (52 miles) apart via a high-speed Ethernet network, allowing AI workloads to access GPUs across both facilities as though they were installed in a single location.

The system delivers 111.2 Tbps of bandwidth while maintaining an end-to-end latency of just 0.9 milliseconds between the two sites. According to WhiteFiber, this latency is close to the theoretical physical limit imposed by the speed of light in optical fiber.

The deployment also integrates WEKA, which provides the storage and data layer, enabling the entire infrastructure to function as a single logical system rather than two independent facilities connected by a network.

The announcement follows DriveNets’ recent expansion of its AI Fabric portfolio with new systems based on Broadcom’s Tomahawk 6 switch silicon, which also support Scale-Across architectures. The WhiteFiber deployment represents the first commercial implementation of the technology disclosed by the company.

Solving AI’s Next Infrastructure Bottleneck

The architecture addresses one of the industry’s emerging challenges.

Building larger AI clusters is no longer limited primarily by GPU availability. Increasingly, hyperscale operators face constraints related to electrical power, land availability and cooling capacity at individual data center sites.

Rather than continuing to expand a single facility, Scale-Across allows computing resources to be distributed across multiple locations while still functioning as one unified AI infrastructure.

Making Two Facilities Behave Like One

Achieving this is far from trivial.

Training large AI models requires thousands of accelerators to exchange enormous volumes of data continuously. Even small increases in latency or packet loss can significantly reduce training efficiency.

According to WhiteFiber, Project Redwood is designed not simply to connect two data centers with a fast network, but to make them behave as a single GPU cluster from the perspective of AI software frameworks.

To accomplish this, the project relies on DriveNets’ Fabric Scheduled Ethernet (FSE) technology, which orchestrates network traffic to eliminate packet loss while maintaining deterministic performance under the heavy communication loads generated by large AI clusters.

WhiteFiber also claims it achieved 111.2 Tbps using only part of the available optical spectrum—performance the company says is roughly twice the capacity reported in comparable field deployments. The company plans to scale the architecture further and has filed patent applications covering key elements of the design.

A New Model for AI Factories

WhiteFiber provides GPU-as-a-Service infrastructure and develops AI computing platforms for enterprise customers. The Nasdaq-listed company (NASDAQ: WYFI) operates facilities in the United States, Iceland and Europe, and recently announced a multi-year contract worth more than $160 million to build AI infrastructure near Paris.

The company promotes a vision of distributed AI Factories, where computing resources located across multiple facilities can be aggregated into a single logical platform.

For DriveNets, the deployment represents more than another customer win. Only last month, the company raised $410 million to accelerate the expansion of its AI Fabric business, arguing that the next generation of AI infrastructure will increasingly rely on high-performance Ethernet as an alternative to InfiniBand.

The WhiteFiber project provides the first commercial proof point for one of the most ambitious capabilities in DriveNets’ portfolio. If Scale-Across gains broader adoption among AI infrastructure providers, it could fundamentally reshape how future AI superclusters are built—allowing them to expand beyond the physical limits of a single data center.

DriveNets today announced a $410 million Series D financing round led by Bessemer Venture Partners and Atreides Management, with participation from AMD, Red Dot Capital, and existing investors including Pitango and D1 Capital. According to industry estimates, the round values the company at approximately $8.5 billion, a sharp increase from its 2022 financing, which reportedly valued DriveNets at more than $2.5 billion.

Beyond the size of the round, one detail stands out. Unlike most software companies, DriveNets says a significant portion of the capital will be used to build readily available inventory of networking platforms for large-scale AI deployments. Such language is unusual for a software vendor and suggests that DriveNets is increasingly moving beyond supplying networking software alone. Instead, the company appears to be financing infrastructure components in advance to support major customer projects. DriveNets also reports a pipeline of orders and projects exceeding $1 billion.

Founded in 2015 by Ido Susan and Hillel Kobrinsky, DriveNets set out to disrupt the traditional networking equipment market. Rather than relying on expensive proprietary routers from vendors such as Cisco Systems, Juniper Networks, and Nokia, the company developed software that allows clusters of standard switching hardware to operate as a single, large-scale carrier-grade router. The model has been adopted by major customers including KDDI, Comcast, and AT&T. According to DriveNets, AT&T now runs most of its core network on the company’s technology.

In recent years, DriveNets has extended the same architectural philosophy into AI data centers. Once known primarily as a telecom networking company, it now positions itself as a provider of a full-stack AI networking fabric, offering software, integration, workload orchestration, and networking management for large AI compute clusters.

That shift places DriveNets in more direct competition with NVIDIA. While NVIDIA has built much of its AI infrastructure strategy around InfiniBand technology acquired through its purchase of Mellanox Technologies, DriveNets is promoting an alternative based on open, multi-vendor Ethernet networking. The company argues that this approach enables customers to combine AI accelerators from different manufacturers within the same cluster, reducing dependence on a single vendor’s proprietary stack.

AMD’s participation in the funding round reflects that vision. Both companies are advocating for heterogeneous AI infrastructure, where different types of AI accelerators can operate together within the same data center environment. If DriveNets originally set out to transform how telecom operators build networks, this latest financing suggests a much broader ambition: becoming the networking layer that connects the next generation of AI infrastructure.

[Photo Credit: Shauli Lendner]