Spectrum-X or InfiniBand: choosing an AI cluster fabric in 2026

On 20 May 2026 NVIDIA reported networking revenue of $14.8 billion for the quarter to 26 April, up 199 per cent year on year. On the earnings call the chief financial officer said Spectrum-X, the company's Ethernet platform for AI, was now larger than all Ethernet networking peers combined, while InfiniBand revenue had more than quadrupled year on year on the adoption of XDR. Dell'Oro Group's first-quarter data, published on 2 June, pointed the same way for the market as a whole: Ethernet switch sales in AI back-end networks more than doubled and accounted for about two-thirds of data-centre switch sales in AI clusters, despite what Dell'Oro called a strong InfiniBand rebound.

What Spectrum-X actually is

Spectrum-X is not a single switch but a pairing of NVIDIA Ethernet switches with NVIDIA SuperNICs. NVIDIA's current documentation lists the SN6600 with 128 ports of 800G OSFP and the SN6800 with 512 ports of 800G in 5U, both built on the 102.4 Tbps Spectrum-6 ASIC with 200G SerDes. On the host side, ConnectX-8 provides 800 Gb/s for Blackwell systems and ConnectX-9 provides 1.6 Tb/s per GPU for Vera Rubin NVL72. The switch and SuperNIC cooperate on adaptive routing and telemetry-based congestion control, and the platform runs standards-based Ethernet with SONiC or Cumulus. NVIDIA claims 1.6 times the AI network performance of off-the-shelf Ethernet, without defining the baseline. Spectrum-XGS extends the same approach to links between data centres hundreds of kilometres apart.

The trade-off is straightforward. Spectrum-X keeps the cluster on Ethernet, which most operations teams already know and which integrates with existing data-centre and storage networks, but its headline features depend on using NVIDIA at both ends of the link. InfiniBand remains a separate fabric from a single vendor, and the XDR rebound shows that many large buyers still choose it for training clusters. Generic Ethernet from Arista, Cisco, HPE Juniper and the ODMs is the third option; Dell'Oro reported that Celestica led AI-cluster Ethernet switch sales in the first quarter.

Dell'Oro also noted that 800G switches made up the vast majority of AI back-end Ethernet shipments, with 1.6T switches sampling and expected to ramp in the second half of 2026. That timing matters for anyone ordering now.

What it means for buyers

Decide the fabric before the GPU order is final, because the NIC choice follows from it. If you choose Spectrum-X, specify switch model, SuperNIC generation and optic type together and confirm that the GPU platform's reference design supports them. If you choose InfiniBand, plan for separate management and spares. In both cases, ask for lead times on switches, NICs and 800G optics separately; NVIDIA's growth means allocation, not list price, is usually the constraint. InTechnologies can source NVIDIA networking alongside GPU servers.