Nvidia booked $25.74 billion in InfiniBand sales during the trailing twelve fiscal months ending July, nearly matching the $25.67 billion generated by its Ethernet and NVSwitch interconnects combined. The figures underscore how the GPU vendor has positioned NVSwitch as the scale-up counterpart to InfiniBand's scale-out dominance in AI clusters.

NVSwitch currently supports 72 accelerators in a coherent memory domain, expandable to 576 devices through two switching tiers. Nvidia does not sell NVSwitch ASICs separately; access requires a Grace or Vera CPU, a Blackwell or Rubin GPU, or — via the NVLink Fusion stack — custom silicon carrying licensed NVLink ports.

What is NVSwitch's role in AI clusters?

Scale-up networking ties GPUs and XPUs into a single coherent shared-memory fabric inside a cluster. NVSwitch performs this function for Nvidia-based systems, while InfiniBand historically handled scale-out communication between nodes. Nvidia owns both technologies after acquiring Mellanox, the InfiniBand supplier.

InfiniBand port-hop latency measures 100 to 120 nanoseconds, roughly one-third to one-half that of typical Ethernet switches. The fabric remained the undisputed low-latency leader for two and a half decades and retains an advantage today.

Who is challenging Nvidia's fabric position?

Two consortium efforts are taking aim at Nvidia's networking dominance:

• UALink — Founded May 2024 by AMD, Broadcom, Cisco Systems, Google, Hewlett Packard Enterprise, Intel, Meta Platforms, and Microsoft. The UALink 1.0 specification launched April 2025 and scales to 1,024 XPUs or GPUs in a single switching tier, nearly double NVSwitch's reach.
• ESUN/SUE-T — Backed by Meta Platforms, Microsoft, AMD, Arista Networks, Arm, Cisco, Hewlett Packard Enterprise, Marvell, Nvidia, OpenAI, and Oracle. Broadcom contributes only to ESUN/SUE-T and has left UALink.

AMD already operates UALink on Broadcom Tomahawk 6 Ultra Ethernet switches inside its Helios rack design. The protocol is a memory atomic standard for accelerator-to-accelerator coherence.

Why does Nvidia's MediaTek deal matter?

Nvidia purchased $3.5 billion in convertible bonds from Taiwanese SoC maker MediaTek. Co-founder and CEO Jensen Huang told Bloomberg that MediaTek is "the biggest and best maker of SoCs in the world." The arrangement grants MediaTek access to NVLink Fusion, letting it integrate NVLink ports into custom chips that participate in NVSwitch memory fabrics.

What does the market look like today?

Nvidia controls an estimated 95 percent of GPU revenues and roughly 75 percent of combined GPU and XPU revenues. That share translates directly into scale-up networking control through NVLink and NVSwitch. Hyperscalers, cloud builders, and AI model builders — including Meta Platforms, which earlier deployed InfiniBand — increasingly prefer open standards with multi-vendor sourcing.

Will Ethernet displace InfiniBand at scale-out?

The Ultra Ethernet Consortium, formed July 2023, targets scale-out networks that combine InfiniBand-class latency with Ethernet's scalability. InfiniBand will remain commercially relevant through Nvidia and certain licensed Chinese HPC deployments, but Ethernet's broad ecosystem — campus, edge, and datacenter — continues to compress InfiniBand into a narrower niche.

What happens next?

A unified UALink 2.0 standard could consolidate the rival camps and challenge NVSwitch directly. If UALink vendors match NVSwitch port-hop latencies while delivering InfiniBand-class figures atop commodity Ethernet, Nvidia will likely build the best UALink switches and ports available. Nvidia may also license NVSwitch more broadly if NVLink-equipped CPU and XPU customers push for it — though the company's stated policy currently prohibits such sales.