The Network Defines the AI Data Center
Next-generation networking for high-performance, scalable, and secure gigascale AI factories.
Introduction
Networking Purpose-Built for the Era of AI Factories
AI factories now span tens of thousands of GPUs—and soon, millions—operating as a single distributed computing engine. To keep accelerators fully utilized, AI workloads demand deterministic latency, lossless throughput, stable iteration times, and the ability to scale not only within a data center, but across multiple sites.
NVIDIA networking delivers the full-stack fabric that makes this possible, combining NVIDIA NVLink™ scale-up, NVIDIA Quantum InfiniBand and Spectrum-X™ Ethernet scale-out, Spectrum-XGS multi-data-center scale-across, NVIDIA BlueField™ DPU and DOCA™ for infrastructure services, and next-generation silicon-photonics platforms to power the world’s most advanced AI factories.
A Trusted Foundation for AI Agents
NVIDIA Sentry and BlueField™ bring in-silicon, out-of-band agent security running on NVIDIA OpenShell reference architecture.
Technology
The Innovations Defining NVIDIA AI Networking
NVIDIA builds networking as part of an integrated AI platform—combining GPUs, SuperNIC™s, DPUs, switches, optics, and software into an extreme codesigned architecture that delivers uncompromised, predictable performance at scale.
Benefits
Accelerated Networking for AI Factories
News
Stay Up to Date on the Latest in Networking
NVIDIA Networking FAQs
NVIDIA networking is a full-stack, accelerator-optimized fabric purpose-built for AI factories—combining NVLink scale-up, Spectrum-X Ethernet and Quantum InfiniBand scale-out, Spectrum-XGS scale-across, BlueField DPUs for infrastructure services (including secure scale-in access and context scale for KV cache), and silicon photonics platforms. Together, these technologies deliver the innovative features and performance characteristics that AI workloads require across deployments spanning tens of thousands to millions of GPUs.
AI factories now span tens of thousands of GPUs operating as a single distributed computing engine—making the network the critical path of every AI workload. NVIDIA networking delivers deterministic latency, lossless throughput, and stable iteration times that keep GPUs fully utilized across training, inference, and RAG workloads at any scale.
NVIDIA offers four primary networking platforms for AI factories: Spectrum-X Ethernet for hyperscale AI cloud networking, delivering 1.6x higher performance than off-the-shelf Ethernet; Quantum-X800 InfiniBand for low-latency AI training and HPC with in-network computing; BlueField DPU for infrastructure services, security, and storage offload; and silicon photonics co-packaged optics for power-efficient switching at million-GPU scale. Spectrum-XGS Ethernet extends all of these across multiple sites into a unified AI super-factory.
NVIDIA Spectrum-X Ethernet is a networking platform purpose-built for AI, delivering 1.6x higher network performance than off-the-shelf Ethernet. It provides accelerated RoCE performance, deterministic latency, and performance isolation for multi-tenant AI workloads—and extends across multiple data centers through Spectrum-XGS scale-across technology, unifying distributed facilities into a single high-performance AI fabric.
NVIDIA Quantum InfiniBand is a high-performance networking platform delivering ultra-low latency, in-network computing, and next-generation silicon photonics co-packaged optics systems for AI and HPC workloads. With SHARP in-network computing that accelerates collective operations and self-healing fabrics that recover from link failures 1,000x faster than traditional networking, Quantum InfiniBand provides the industry’s fastest and most scalable networking for AI training clusters and scientific supercomputers.
The NVIDIA BlueField platform ignites unprecedented performance, efficiency, and innovation to power AI factories, the new class of purpose-built data centers for AI. With its robust compute power and integrated software-defined hardware accelerators for networking, storage, and security, BlueField creates a secure, accelerated infrastructure for any workload in any environment, ushering in a new era of agentic AI, where autonomous agents reason, collaborate, and act at scale.
NVIDIA Spectrum-XGS Ethernet is a scale-across networking technology that connects distributed data centers into a single unified AI super-factory. Spectrum-XGS delivers 1.9x higher NCCL performance for multi-site AI deployments compared to off-the-shelf inter-site Ethernet—enabling AI factories to scale beyond any single building’s power and physical capacity limits while maintaining the same high-performance fabric characteristics as within-facility Spectrum-X deployments.
NVIDIA silicon photonics networking integrates optical components directly with the switch ASIC—eliminating pluggable transceivers to deliver 5x better power efficiency and uptime compared to traditional pluggable designs. Co-packaged silicon photonics is the physical layer technology that enables networking at million-GPU scale and multi-site AI factories, reducing insertion loss for significantly better signal integrity and reliability.
NVIDIA DOCA is the unified software platform that unlocks the full potential of NVIDIA BlueField and NVIDIA ConnectX™ for agentic AI. It provides SDKs, libraries, enterprise-ready microservices, and orchestration frameworks—spanning networking, AI-native storage, runtime security, telemetry, and lifecycle management—to accelerate, secure, and scale production AI workloads. DOCA is the software foundation for building and operating intelligent, high-performance AI factory infrastructure.