Intel spin off Cornelis Networks has a plan to take on Infiniband, Ethernet

  • Cornelis Networks was formed when Intel spun out its Omni-Path business and is now led by former Intel exec Lisa Spelman
  • Spelman told Fierce the company is targeting three multi-billion-dollar market segments with its high-performance data center networking technology
  • Dell'Oro Group estimates the AI back-end networking market will be worth $20 billion by 2028

Taking on the likes of Nvidia and Broadcom is certain to be a challenge. But former Intel exec and current Cornelis Networks CEO Lisa Spelman thinks it has the goods to give the tech titans a run for their money in the data center networking space.

You may not have heard of the company before, but it has big name roots. Cornelis Networks was founded four years ago when Intel spun out its Omni-Path business, and it has since been working to grow adoption beyond its traditional cohort of enterprise and research high-performance compute customers. With artificial intelligence (AI) demand now exposing key flaws in existing technologies like Infiniband and Ethernet, Spelman thinks it now has a prime opportunity to do just that.

The CEO told Fierce it has big ambitions. Specifically, she said it is targeting customers in three multi-billion-dollar markets: hyperscale cloud players, next-gen cloud players like GPU specialists, and enterprise AI.

“We’re building pipeline right now…We’re pursuing a 10X-ing of our revenue year over year,” Spelman told Fierce, adding it is hustling to hire more sales staff. “Our next year from my perspective is limited by the amount of sales coverage we have. It is not limited by the number of workloads we will be able to address.”

Its targets? Hyperscalers, next-generation cloud providers and GPU specialists, and those looking to deploy high-performance AI for enterprise and research applications.

So, what exactly does Cornelis do? Well, Spelman said the company is narrowly focused on providing network connectivity within data centers for HPC use cases via its line of Omni-Path switches, adapters, gateways, cables and software. Its lineup currently includes a 100-gig product, but it is launching a 400-gig product in the first half of 2025. It plans to follow that up with an 800-gig option in the first half of 2026 and a 1.6 Tbps product in early 2027, she said.

Open door

Spelman said the market shift toward AI has exposed and exacerbated key issues with both Infiniband and Ethernet. She criticized Ethernet as a “lossy network” that has issues operating at scale and pointed to the formation of the Ultra Ethernet Consortium last year as the industry’s recognition of these problems.

Infiniband, meanwhile, has issues with connections over certain distances. And while better than Ethernet on loss, Spelman said “it wasn’t inherently built for scale and highly parallelled workloads.”

“There’s a market opportunity because of those challenges that the two established solutions provide,” she explained. “What we have done is build a network that has all of the characteristics that are required for scale out AI.” That is, a lossless, high performance network featuring congestion management and dynamic lane scaling. Put another way, Cornelis already offers all the features that the Ultra Ethernet Consortium is trying to achieve to deliver an ideal network, she said.

“We’re not an existing Ethernet provider trying to incrementally improve and add on top of Ethernet. We’re defining the whole network and figuring out how to build in compatibility” with existing tech, Spelman said.

Asked about the market Cornelis is chasing, Dell’Oro Group VP Sameh Boujelbene told Fierce “AI back-end networks is a net new market that emerged to connect GPUs/accelerators for large AI Clusters across multiple racks and sometimes even across multiple data centers.”

She added that the analyst firm is projecting that AI back-end networks revenue will grow at a more than 50% compound annual growth rate to exceed $20 billion by 2028.

“AI back-end networks represent a rapidly growing and highly attractive net new market opportunity. Additionally, they are driving new requirements, which in turn pave the way for significant innovation and disruption,” Boujelbene said. “This dynamic is further amplified by the substantial funding and investment flowing into the sector.”

Gartner Distinguished VP Andrew Lerner agreed that AI has created new opportunities in the market. Why? "It is a combination primarily two factors: #1 that the requirements for an AI fabric are dramatically different from traditional x86-compute oriented workloads" and #2 "the increasing and immediate demand with which enterprises want to deploy GPU clusters locally."

"It is critical to address these requirements in order to maximize GPU utilization, which is by far the most expensive component of AI. A GPU cluster can often easily be 5-20 and even 100 times the cost of requisite network switches. Simply put: In a GPU-intensive environment, the network’s job is to keep the GPUs busy," he added. "Thus, we are seeing organizations explore new vendors and approaches to address this challenge."

Spelman said she knows taking on the likes of Nvidia (which provides not only Infiniband but also Ethernet-based Spectrum-X products) and Broadcom will be tough. But she noted “we are small, but because we’re small we’re fast and we’re nimble and we focus on simplification.” That extends to pushing out custom products that integrate a customer’s unique IP where necessary.

“We are not pursuing the need to be everything to everyone,” she concluded. “We need to be something really important to a couple really important someones.”