NVIDIA unveiled its new Spectrum-XGS Ethernet solution over the weekend, designed to link distributed data centers and enable them to function as a single, unified computing fabric—a vision the company calls “AI Super-Factories.”

The launch marks a strategic expansion of NVIDIA’s networking portfolio and its first serious move into long-haul data-center interconnects, bridging geographically dispersed server farms rather than just wiring machines within a single site. Developed in Israel and rooted in Mellanox technology—acquired by NVIDIA in 2019—the system represents a leap beyond Mellanox’s traditional focus.

Where Mellanox primarily delivered InfiniBand and Ethernet solutions for communication within a single data center, Spectrum-XGS addresses the new challenge of connecting multiple sites. The result introduces a “third dimension” of scaling: alongside scale-up (linking processors within a single server) and scale-out (connecting thousands of servers within a cluster), Spectrum-XGS now enables scale-across, turning separate data centers into one virtual compute unit.

Built on NVIDIA’s Spectrum-X Ethernet platform, the system combines next-generation switches, ConnectX-8 smart network adapters, and specialized software algorithms that optimize traffic over long geographic distances. It also integrates full end-to-end monitoring and latency controls—boosting the performance of NVIDIA’s NCCL (collective communications library) by nearly 2x in distributed AI workloads. NCCL is a critical software layer that synchronizes and transfers data at high speed across thousands of GPUs working in parallel.

The Broadcom Rivalry

Industry watchers view Spectrum-XGS as a direct response to Broadcom’s Jericho 4, a next-generation networking chip built for hyperscale cloud backbones and designed to stitch together massive server farms. Both solutions share the same ambition: transforming isolated compute “islands” into unified, cloud-scale supercomputers powered by Ethernet.

The differences, however, are telling. Broadcom supplies general-purpose chips that cloud providers integrate into their own custom networks. NVIDIA, by contrast, is delivering a turnkey solution—hardware, software, and algorithms tightly engineered for AI workloads. Jericho 4 is flexible and open-ended, while Spectrum-XGS is purpose-built: a closed, vertically integrated environment optimized for NVIDIA’s own AI ecosystem and global-scale compute.

The bottom line: with Spectrum-XGS, NVIDIA is pushing its networking capabilities far beyond the Mellanox core, moving from intra-data-center fabrics to inter-data-center scale. In doing so, it positions itself as a direct competitor to Broadcom in the battle over the connective tissue of the AI era.