Nvidia generated an estimated $25.74 billion from InfiniBand products in the trailing twelve fiscal months ending in July. Combined Ethernet and NVSwitch interconnect sales reached $25.67 billion over the same window, according to industry estimates.

The $70 million gap between those two lines marks a historic narrowing. InfiniBand, the switched-fabric protocol that survived the dot-com bust, has long anchored Nvidia-led AI clusters. NVSwitch, the company's own coherent memory fabric, now threatens to pass it.

What is NVSwitch doing that InfiniBand cannot?

NVSwitch ties GPU, XPU and CPU memory into a single coherent shared space. One tier carries 72 Nvidia accelerators; two tiers reach 576 devices sharing memory. InfiniBand remains the lower-latency fabric, with port hops at 100 to 120 nanoseconds, roughly one-third to one-half that of comparable Ethernet switches.

The commercial trade-off drives customer choice. Nvidia sells NVSwitch only inside its own rack configurations. Buyers must include "Grace," "Vera," "Blackwell" or "Rubin" silicon in any NVSwitch-equipped rack.

Who else is competing for the scale-up slot?

The Unified Accelerator Link consortium was founded in May 2024 by AMD, Broadcom, Cisco Systems, Google, Hewlett Packard Enterprise, Intel, Meta Platforms and Microsoft. UALink 1.0 shipped in April 2025 and supports 1,024 accelerators in one switch tier, nearly twice the reach of NVSwitch. AMD pairs Broadcom Tomahawk 6 Ultra Ethernet switches with the UALink protocol inside its "Helios" racks.

• UALink 1.0 port-to-port hops target roughly 100 nanoseconds per the specification
• NVSwitch reaches 72 devices per tier and 576 across two tiers
• Ultra Ethernet Consortium was founded July 2023 to graft InfiniBand-style latency onto Ethernet
• Broadcom left UALink and now backs the rival ESUN group and its SUE-T load-balancing layer

A second ESUN camp includes Meta Platforms, Microsoft, AMD, Arista Networks, Arm, Cisco, Hewlett Packard Enterprise, Marvell, Nvidia, OpenAI and Oracle.

What is InfiniBand's near-term ceiling?

Industry estimates peg Nvidia at roughly 95 percent of standalone GPU revenue and 75 percent of combined GPU plus XPU revenue. InfiniBand likely stays near its current $25.74 billion run rate through calendar 2027. Beyond that, hyperscaler adoption of Ethernet scale-out networks narrows the runway.

Mellanox Technologies consolidated InfiniBand after reviving the protocol from the dot-com era. Mellanox paid $218 million for Voltaire in November 2010 to broaden switching reach. Intel answered with the $125 million QLogic purchase in January 2012, before spinning that team out as Cornelis Networks in September 2020.

Why did Nvidia buy MediaTek convertibles?

Nvidia purchased $3.5 billion of convertible bonds issued by Taiwanese SoC maker MediaTek. The deal grants MediaTek access to the NVLink Fusion stack, which lets third parties place NVLink ports on their own accelerators. Nvidia chief executive officer Jensen Huang told Bloomberg MediaTek is "the biggest and best maker of SoCs in the world."

The arrangement extends NVSwitch into silicon Nvidia does not make. Licensees can join a coherent memory pool with Nvidia parts, although the company still refuses to sell NVSwitch ASICs separately to non-Nvidia rack designs.

How large is the opportunity if standards shift?

Industry estimates put peak NVLink and NVSwitch sales near $100 billion if Nvidia holds its GPU and XPU share through the end of the decade. UALink vendors reaching InfiniBand-grade latency on Ethernet switches would sharpen that threat. Nvidia, on past form, would still build the best UALink switches if the standard prevails, just as it pivoted from InfiniBand marketing into Spectrum-X Ethernet.