Networking disruptor Cornelis Networks laid out its Ethernet vision while touting the AI-readiness of its open architecture, which has seen recent Top500 success in the supercomputing realm.

In an exclusive interview with SDxCentral, CEO Lisa Spelman said the company was ready for AI’s impact on networking through solving traffic pattern issues with fine-grain adaptive routing, claiming Cornelis has “the world's most advanced congestion management and congestion avoidance features.”

These advantages for high-performance computing (HPC) see Cornelis’s Omni-Path Architecture-powered interconnect in multiple TOP500 systems at Sandia National Laboratories and the Lawrence Livermore National Laboratory. Spinning out of Intel in 2020, Cornelis Networks is synonymous with the high-speed interconnect rival to InfiniBand.

Developed in 2015, the system connects servers, processors, and storage clusters for intensive and supercomputing workloads. Under the hood is an open-source framework that Cornelis touts as providing simplified differentiation in the market. Omni-Path core software is defined by Linux Kernel Drivers; a reliance on the OpenFabrics Alliance (OFA) stack; and capacity for open-source Message Passing Interface (MPI) implementations, lending to compatibility with AMD’s Radeon Open Compute (ROCm) platform and Nvidia’s CUDA environment.

With its recently launched CN5000 platform, Cornelis claimed to carry more data with less latency compared to the InfiniBand Next Data Rate (NDR) 400G data center networking platform, which is made up of network interface cards (NICs), switches, software, and cabling, and support for 400 gigabit per second (Gbps) networking speeds.

Spelman touted Cornelis’s native framework has developed over the last decade into an “extremely high-performance architecture that is built directly and exactly for HPC workloads.”

Cornelis has found extra vindication for these claims, with CN5000 debuting on the latest Top500 supercomputer index a year after its release. The platform powers Lynx, the supercomputer deployed by Lawrence Livermore in June. The 952-node cluster is powered by Dell PowerEdge servers and Intel Xeon processors, supporting production modeling, simulation, and analysis for the National Nuclear Security Administration (NNSA).

The latest Top500 list sees Lynx sit at number 277. While the vendor battle between network fabrics and interconnects is not solely defined by supercomputing, it is worth noting that InfiniBand continued to increase its showing on the list, outnumbering Omni-Path by a ratio of 18:1.

Omni-Path, for which Cornelis is the primary exclusive vendor, saw an almost 47 percent drop year-on-year in representation in the list, with it now powering 16 of the 500 systems. But according to its CEO, the fresh-faced Cornelis continues to lay tracks ahead of the train.

Into the ether

The path ahead when it comes to further all-around Omni-Path adoption may revolve around Ethernet. According to Spelman, the new stage of Cornelis’s evolution centers around the technology, with Cornelis preparing its 1.6 terabit per second (Tbps) CN7000 offering, which is set to integrate the Omni-Path layer with Ultra Ethernet Compliant (UEC) standards.

Ethernet will also define this year’s follow-up to the CN5000 in the form of the CN6000 SuperNic. A preview from November revealed the Ultra Ethernet-compliant 800G Ethernet adapter offers double the performance of the CN5000 with support up to 1.6 billion messages per second, according to Cornelis.

November also revealed that Cornelis is poised to help power the Genesis Mission project with its Ethernet networking tech.

According to the CEO, competitors touting Ethernet are disadvantaged by not having the fabric capabilities to handle AI-driven traffic and congestion.

“We're adding the Ethernet wire protocol, or the transport layer, to actually solve the AI congestion and data movement challenge. Whereas our competitors have to come at it from the fact that they have the transport layer and the wire protocol, and they have to add these complicated features on top,” Spelman said.

The Cornelis nod adds another blessing to the Ethernet use case. 650 Group argued AI-driven adoption of 800G and 1.6T system capacity would push Ethernet adoption past InfiniBand as “the dominant technology for scale-out,” generating more than $8 billion in revenues. This comes after AI led a 57 percent increase in global data center capex in 2025, according to Dell’Oro Group.

“The industry is attracted to what Ethernet represents for the long haul, which is industry standards that multiple people can build on top,” Spelman said. “But the definition of what Ethernet means, as we move to Ultra Ethernet, starts to adapt a bit. You're seeing more capabilities like traffic management and not just congestion awareness, but congestion avoidance. Different styles and capabilities of features are evolving a lot.”

Spelman stressed this was why Cornelis would retain its open advantage as more open-source systems emerge revolving around Ethernet, such as OpenAI’s Multipath Reliable Connection (MRC) system for AI model training.

“Whether it's been the Ultra Ethernet Consortium, the growing push into UA Link, the MRC announcement, you can see that networking is having a renaissance. It is important in the system of delivery of the AI application, and we think that's fantastic. It's good for us, and it's good for customers as well.”