DriveNets Targets AI's Core Economics with New 1.6 Tbps Fabric

RA'ANANA, Israel – July 01, 2026 – In the high-stakes race to build the digital infrastructure for artificial intelligence, the most critical battles are often fought over the silent, unseen plumbing that connects everything. Today, DriveNets, a company that has already reshaped the networks of telecom giants like AT&T, made a significant move into this arena, unveiling a new generation of networking platforms aimed squarely at the heart of the AI data center.

The announcement of its DriveNets 2600SL and 2601S platforms is more than a product launch; it's a strategic declaration. By leveraging Broadcom's latest Tomahawk 6 silicon to deliver massive 102.4Tbps capacity and next-generation 1.6Tbps port speeds, DriveNets is positioning open, standard-based Ethernet as the definitive challenger to the proprietary fabrics that have dominated the AI training space. The core of their argument is a simple, powerful economic principle articulated by co-founder and CEO Ido Susan: "The most expensive idle asset in the world today is an XPU waiting on the network."

The Billion-Dollar Bottleneck

As AI models scale to hundreds of thousands of processing units (XPUs), the network fabric has become the primary arbiter of performance and efficiency. An underpowered network leaves costly AI accelerators starving for data, tanking utilization rates and extending training times from days to weeks. DriveNets' new platforms are engineered to eliminate this bottleneck. Each switch offers 64 ports of 1.6Tbps Ethernet, a speed that industry analysts at 650 Group expect will constitute the majority of new AI port deployments by 2027. This raw capacity allows for the construction of massive clusters connecting tens or even hundreds of thousands of XPUs in highly efficient fat-tree architectures.

The collaboration with Broadcom is central to this capability. "As AI systems scale to hundreds of thousands of XPUs, the network fabric has become a critical factor in AI performance, infrastructure utilization, and overall token economics," said Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group. He noted that pairing the Tomahawk 6 silicon with DriveNets' software delivers "unprecedented performance and reliability at massive cluster scale," reinforcing the industry-wide view that "open Ethernet is becoming the foundation of the next-generation AI data center."

Operational Innovation: Redefining Token Economics

DriveNets' true operational innovation lies not just in the hardware's speed, but in its disaggregated, software-driven approach. The company claims its AI Fabric enables multi-vendor clusters to achieve performance on par with closed, single-vendor solutions—a clear shot at NVIDIA's dominant InfiniBand—but at a significantly lower cost. This strategy hinges on improving what the industry calls 'token economics.'

Token economics is the ultimate measure of an AI cluster's efficiency: the total cost to process a single unit of data, or 'token.' Lowering this cost is the holy grail for foundation model developers and hyperscalers. DriveNets attacks this problem through full-stack optimization. This isn't just about moving packets quickly; it involves fine-tuning the entire data path, from Network Interface Card (NIC) drivers and the operating system kernel up to the Collective Communications Libraries (CCL) that manage how XPUs exchange information during training. By optimizing these software layers, the company aims to minimize latency and synchronization overhead, ensuring XPUs spend more time computing and less time waiting.

"What they're doing is acknowledging that hardware speed alone is insufficient," noted one independent network architect. "The performance gains are found by treating the AI cluster as a single, cohesive system. Optimizing CCLs to be aware of the network topology, for instance, allows for smarter data routing that avoids congestion and dramatically speeds up the 'all-reduce' operations that are fundamental to distributed training."

This holistic optimization is managed by the DriveNets AI Cluster Orchestrator, a software layer designed to automate deployment, benchmarking, and operations, further reducing the complexity and cost of running AI infrastructure at scale.

The Physics of AI: Cooling the Revolution

The relentless march of AI is also forcing a physical transformation of the data center. The sheer power density of modern AI servers, packed with multiple power-hungry XPUs, has pushed traditional air cooling to its limits. DriveNets' announcement directly addresses this physical reality with the DriveNets 2600SL, a platform featuring a "true 100% liquid-cooled design."

This is not a cosmetic feature but a critical enabler for the future of AI. Liquid cooling is vastly more efficient at heat removal than air, allowing data centers to pack more computational power into each rack without the risk of thermal throttling, which can cripple performance. This leads to direct operational savings through improved Power Usage Effectiveness (PUE) and allows for the deployment of the most powerful next-generation accelerators. By offering both air-cooled and liquid-cooled options, DriveNets is catering to the diverse and evolving thermal landscapes of modern data centers, from legacy facilities to cutting-edge AI factories.

A New Market, A New Playbook

This move comes as the AI infrastructure market enters a period of explosive growth. "This is no longer an incremental networking opportunity — it's a brand-new market that didn't exist two years ago and is now heading toward a TAM of more than $100 billion," stated Alan Weckel, Founder and Technology Analyst at 650 Group. He sees DriveNets' offering as reinforcement that "Ethernet is the dominant fabric for AI infrastructure... overtaking the proprietary alternatives that came before it."

DriveNets has built its credibility on its ability to execute large-scale, disaggregated networks for some of the world's most demanding service providers. Now, it is applying that same playbook—separating hardware from software to deliver flexibility and cost savings—to the burgeoning AI market. With shipments of the new platforms slated for the third quarter of 2026, the industry will be watching closely to see if this open, economically-driven approach can successfully challenge the established order and become the new standard for powering the AI revolution.