⚠ The Ethernet ConsortiumModerate threat

The networking grip depends on Nvidia's interconnect being proprietary and superior, and the danger is a coordinated industry effort aimed directly at replacing it with open standards. The Ultra Ethernet and UALink consortia — backed by essentially every major rival at once¹ — exist specifically to build an open, high-performance interconnect that would let buyers tie accelerators together without Nvidia's proprietary fabric, dissolving one of the stickiest threads in the systems moat.

The threat is pointed because networking is the part of the systems advantage a consortium can most plausibly standardize. Unlike CUDA's decade of accumulated habit, an interconnect is a technical specification that a well-resourced alliance can define and implement, and if an open standard reaches parity, buyers could combine rival chips with standard networking, reclaiming the freedom to mix vendors that Nvidia's proprietary fabric denies them.

The shield here: Nvidia's networking works today, at enormous scale, in production, and keeps advancing, while the open standards are still maturing and must catch a moving target. Committee-designed standards tend to trail a single vendor's tightly integrated stack, and the operational risk of betting a supercomputer on an immature open fabric keeps cautious buyers on Nvidia's proven interconnect for now.

File it at moderate. The open-interconnect consortia are the most credible and best-coordinated assault on the networking thread, aimed precisely where the systems moat is most catchable — but Nvidia's fabric is ahead, proven, and still improving, and the standards must not only match it once but keep pace as it advances. The grip is genuinely contested here, more than anywhere else in the systems story.