At Hot Chips 2026, NVIDIA unveiled its Spectrum-X multiplane Ethernet architecture designed to scale AI factory networks beyond current interconnect limits.

NVIDIA presented the Spectrum-X Ethernet Multiplane Network Architecture at Hot Chips 2026, detailing plans to scale AI factory networking from thousands of GPUs toward half a million. The session repeatedly returned to the AI Factory platform framework, underscoring that this architecture reflects NVIDIA’s active buildout. NVIDIA divides the AI factory into five purpose-built networks: scale-across, scale-in, scale-out, scale-up, and AI context scale. Scale-in was newly emphasized during the presentation, arriving shortly after Broadcom showcased its Thor Ultra. The core argument asserts that a single general-purpose fabric cannot effectively serve all five domains, establishing the foundational premise for the entire architecture.

NVIDIA confirmed that its co-packaged optics solutions are now in production and outlined the strategic rationale behind Astra, scale-in, and BlueField-4. Scale-up networking operates over NVLink, with the NVL72 rack pairing an NVLink spine with switch trays to form a 72-GPU scale-up domain that NVIDIA credits with leading tokens per megawatt. While enterprise, hyperscale, and service provider networks each employ distinct spine designs, NVIDIA argues that AI factories require purpose-built Ethernet rather than repurposing general-purpose topologies.

Spectrum-X serves as NVIDIA’s giga-scale solution for scale-out Ethernet. The company claims 1.6x higher RDMA bandwidth, 2.2x better multi-tenancy, and 1.3x lower bandwidth jitter across 102.4T switch systems and 1.6T SuperNICs. Building on earlier analysis of the MRC RDMA transport protocol, NVIDIA emphasizes resolving jitter through an end-to-end system approach rather than isolated devices. Headline metrics show a 1.9x increase in training performance within a multi-tenant AI factory, drawn from a DeepSeek V3 multi-job training run. In noisy shared environments, step times for off-the-shelf Ethernet exceed Spectrum-X curves, with the performance gap widening across more than 80 steps. NVIDIA attributes this separation to low-jitter communications and tenant noise isolation.

Extreme co-design is most evident in NCCL performance. During Nemotron Ultra pre-training, gradient all-reduce achieved a 14x lead, token dispatch reached 3.5x, and both gradient reduce-scatter and parallel matrix multiply hit 2x, all relative to off-the-shelf Ethernet. Scale-out AI relies heavily on optics, unlike traditional cloud architectures. NVIDIA estimates that optical components account for 10% of compute power in an AI factory, significantly higher than conventional cloud data centers. Spectrum-X Ethernet photonics represents the production-ready implementation of this strategy. NVIDIA has placed a co-packaged optics chip featuring micro ring modulators into production and developed a 3D-stacked silicon photonics engine on the TSMC Coupe process. The company claims these innovations deliver four fewer lasers, lower power consumption, and a tenfold reduction in mean time between interruptions. NVIDIA remarked, “100G SerDes is last year’s stuff,” immediately following Broadcom’s showcase of the 100G SerDes Thor Ultra at Hot Chips.

Today’s hyperscale cloud typically begins with a top-of-rack switch topology. NVIDIA identifies this 2k-scale baseline as the conventional layout it aims to surpass for AI factories. Multi-rail represents the first evolutionary step, demonstrating 8k Rubin GPUs operating at 1.6T per-GPU scale-out bandwidth. The architecture utilizes 100T switches with 64 ports of 1.6T each to distribute traffic across four rails, ensuring the full bandwidth of the NIC connects directly to the switch port. NVIDIA then introduced the multiplane topology. Unlike multi-rail, which dedicates full rails, multiplane separates the fabric into planes that share switch hardware. This allows a NIC to connect to multiple switches simultaneously. Instead of routing 1.6Tbps to a single switch, traffic splits into eight 200Gbps links, each directed to a different switch.

Multiplane extends AI factory scale by 64x compared to multi-rail. NVIDIA projects support for 512k Rubin GPUs while maintaining 1.6T scale-out bandwidth per GPU. This architecture uses 100T switches with 512 ports of 200G, organized into eight planes across four rails. The multiplane design also reduces physical footprint, requiring 1.7x fewer scale-out switches than traditional multi-tier single-rail topologies, thereby lowering power consumption, rack space requirements, and overall costs. Fault tolerance is a primary advantage of multiplane. Under partial bandwidth loss, the Spectrum-X multiplane topology maintains 90% bandwidth with a 2.68ms detection time, whereas a traditional multi-tier topology drops to 0% bandwidth. NVIDIA translates this into a 1.6x goodput advantage over off-the-shelf Ethernet multiplane implementations. Spectrum-X detects faults in 2.68ms—roughly 400x faster—and recovers in 100ms, approximately 11x faster, sustaining 90% bandwidth. In contrast, off-the-shelf topologies experience complete bandwidth loss, requiring 1080ms for detection and recovery.

Spectrum-XGS extends connectivity across multiple AI factories. Featuring 800Gb/s per port on 102.4Tb/s switches and ConnectX-9 SuperNICs, NVIDIA claims up to 1.9x lower multi-site latency and doubled scale-across performance for distributed AI operations. NVLink Fusion integrates third-party XPUs into the NVIDIA AI platform, delivering 3.6 TB/s of all-to-all bandwidth per XPU to connect 72 XPUs within a single domain. This is housed in a cableless MGX rack rated for 45°C inlet temperatures and 100% liquid cooling, achieving 3x lower latency.

Software must evolve alongside this hardware. DSX serves as NVIDIA’s AI factory platform, encompassing power optimization, infrastructure software, and platform software across DSX OS, DSX Sim, DSX MaxLPS, and DSX Flex. NVIDIA emphasizes the need for large-scale simulation and validation prior to construction, which becomes increasingly critical as network topologies grow and cabling requirements become more complex. DSX Air compresses deployment timelines from months to days: infrastructure bring-up drops from six months to one week, infrastructure software development falls from three months to one week, and full deployment shortens from two weeks to one day.

NVIDIA concluded by reiterating the five networking infrastructures of the AI factory: scale-in, scale-up, scale-out, scale-across, and context scale. Each receives purpose-built silicon, with performance multipliers ranging from 18x down across bandwidth, packet rate, latency, and jitter. Every networking dimension presented is purpose-built and co-designed with the platform rather than borrowed from general-purpose fabrics, representing the central thesis of the presentation. NVIDIA argues that Ethernet can support AI factory scale-out, but only when the hardware is specifically engineered for AI workloads. The multiplane topology is the centerpiece, expanding addressable scale from 8k to 512k GPUs while claiming reduced switch counts and superior fault resilience. Whether these metrics hold in production deployments remains the definitive test, with upcoming Hot Chips sessions expected to clarify how the Vera Rubin software-and-optics ecosystem integrates.