NVIDIA launches 102.4Tbps Spectrum-6 Ethernet switch for gigascale AI factories

NVIDIA is aggressively defending its infrastructure moat as AI clusters shift from specialized InfiniBand to open Ethernet fabrics. By vertically integrating the Spectrum-6 switch with Rubin GPUs and ConnectX-9 SuperNICs, NVIDIA aims to offer InfiniBand-like deterministic performance within a standard Ethernet environment. This architecture allows streaming and cloud giants to scale generative AI training without the interoperability hurdles of proprietary protocols. For the industry, this signals a transition where the 'unit of competition' has moved from individual silicon to the end-to-end network fabric. Watch for third-party benchmarks comparing Spectrum-6 against Broadcom’s Tomahawk 6 to determine if proprietary software offsets the cost advantages of open-white-box switching.

The launch of Spectrum-6 occurs as Ethernet challenges the long-standing dominance of InfiniBand in high-performance AI clusters. Per Dell'Oro Group (March 2026), Ethernet switch sales in AI back-end networks more than tripled in 2025, now accounting for over two-thirds of the market. This shift is driven by hyperscalers like Amazon, Meta, and Microsoft seeking vendor diversity and cost-effective scaling for trillion-parameter models. While InfiniBand led with 80% share in late 2023, the rise of the Ultra Ethernet Consortium (UEC) has accelerated the adoption of AI-optimized Ethernet fabrics.

Competition in the 100Tbps tier is intensifying. Broadcom began shipping its Tomahawk 6 (TH6) switch series in June 2025, which also supports 102.4Tbps bandwidth. Per industry analysis from TrendForce (October 2025), NVIDIA's Spectrum-6 was initially perceived as trailing Broadcom’s development cycle by roughly one year. However, NVIDIA’s strategy relies on 'extreme codesign,' bundling the switch with its own ARM-based Vera CPUs and Rubin GPUs to optimize power efficiency. Per NVIDIA (January 2026), the Rubin platform aims to reduce inference token costs by 10x compared to the previous Blackwell generation.

Technological focus has shifted toward co-packaged optics (CPO) and liquid cooling to manage the thermal demands of gigascale factories. NVIDIA's Spectrum-X Ethernet Photonics is reported to achieve 5x higher power efficiency than traditional pluggable transceivers. As individual data center sites hit power capacity limits, the industry is moving toward 'Scale-Across' architectures. Per Futuriom (June 2025), these next-generation fabrics are designed to interconnect multiple distributed data centers into a single unified computing resource, a requirement for the frontier models expected in 2027 and beyond.