NVIDIA (NVIDIA) achieved a remarkable $11 billion in networking revenue last quarter, primarily fueled by the robust growth of its NVLink, Spectrum-X Ethernet, and InfiniBand products. This significant performance underscores a structural shift in AI data centers, evolving towards an 'AI factory' design paradigm.
This new design philosophy implies that compute, networking fabric, and input/output (I/O) are procured as an integrated system, rather than the previous fragmented approach of purchasing individual switches or network interface cards (NICs). According to analysis, NVIDIA's customers are now procuring network products at a rate approaching 90%. NVIDIA CEO Jensen Huang also highlighted this strategic shift, describing Spectrum-X as a 'blockbuster product' that marks the company's transition from a standalone switch vendor to a leader in AI-optimized interconnects.
This transformation directly impacts the purchasing decisions of traditional network equipment vendors like Cisco and Arista, as well as their customers. Market spending is consolidating towards AI-optimized network interconnects and rack-scale systems, influencing procurement priorities, interoperability testing, and product lifecycle planning. The ability to integrate compute, networking fabric, and acceleration capabilities has become a critical selection criterion. Cisco has deepened its collaboration with NVIDIA, planning to integrate Silicon One with Spectrum-X and SuperNIC technologies to build a unified, AI-oriented network architecture.
For procurement entities, future architectural choices will be increasingly based on workload patterns, network interconnect integration, and operational models, rather than solely focusing on independent equipment performance metrics. Nevertheless, supply timeliness and product certification cycles remain practical constraints for short-term deployments.
Within AI data centers, dedicated network interconnects are playing an increasingly vital role alongside compute cores like GPUs. NVIDIA's Spectrum-X is central to its Ethernet solutions, while the Quantum series represents its InfiniBand platform. NVLink, on the other hand, is responsible for connecting accelerators within the GB200/GB300 designs in a rack. Furthermore, BlueField Data Processing Units (DPUs) and SuperNICs work in conjunction with these network interconnect components to build server and rack systems. The ultimate goal is to achieve a more integrated, 'factory-like' deployment model, effectively alleviating performance bottlenecks between training and inference stages.
In terms of deployment choices, Spectrum-X Ethernet is the ideal option for organizations already familiar with and possessing mature Ethernet operational systems, tools, and skills, allowing for better alignment with their existing data center practices and L2/L3 network ecosystems. Environments pursuing highly integrated AI network interconnects will typically continue to opt for InfiniBand clusters based on NVIDIA's Quantum platform, especially when existing workflows and infrastructure are already established.
NVLink technology plays a crucial role in conjunction with GB200/GB300 and BlueField DPUs, further strengthening NVIDIA's integration capabilities in AI compute and networking.
A key point in the frequently asked questions regarding Spectrum-X Ethernet is that NVIDIA has surpassed Cisco in specific data center Ethernet sales, particularly in this quarter's data center Ethernet sales figures, outperforming Cisco and other competitors like Arista.