Competing with giants like Nvidia and Broadcom will undoubtedly be a major challenge. However, former Intel executive and current CEO of Cornelis Networks, Lisa Spelman, believes that Intel has the ability to compete with tech giants in the field of data center networking.
–After Intel divested its Omni Path business, Cornelis Networks was established and led by former Intel executive Lisa Spelman.
–Spelman told Fierce that the company is targeting three market segments worth billions of dollars with its high-performance data center network technology.
–Dell’Oro Group estimates that the value of the artificial intelligence (AI) backend network market will reach $20 billion by 2028.
You may not have heard of this company before, but it has a prominent background. Cornelis Networks was founded four years ago when Intel spun off its Omni Path business. Since then, it has been striving to expand its customer base beyond traditional enterprise and research high-performance computing customers. With the increasing demand for artificial intelligence (AI) exposing key flaws in existing technologies such as InfiniBand and Ethernet, Spelman believes this is a great opportunity for it to expand its influence.
The CEO told Fierce that they have ambitious goals. She specifically mentioned that they are targeting customers in three billion dollar markets: hyper scale cloud service providers, next-generation cloud service providers such as GPU experts, and enterprise level AI.
“We are currently establishing sales channels… What we are pursuing is a tenfold increase in annual revenue,” Spelman told Fierce, adding that they are intensifying the recruitment of more sales personnel. “From my perspective, our development in the next year is limited by our sales coverage rather than the number of workloads we can handle.”
Their goal? Ultra large scale cloud service providers, next-generation cloud service providers, GPU experts, and users who wish to deploy high-performance AI for enterprise and research applications.
So, what exactly does Cornelis do? Spelman stated that the company focuses on providing network connectivity for high-performance computing (HPC) use cases within data centers through its Omni Path series of switches, adapters, gateways, cables, and software. At present, the product line includes a 100G product, but a 400G product will be launched in the first half of 2025. She also revealed plans to launch 800G in the first half of 2026 and a 1.6Tbps product in early 2027.
Spelman stated that the market shift towards AI has exposed and exacerbated fundamental issues with both InfiniBand and Ethernet. She criticized Ethernet as a ‘lossy network’ with issues during large-scale operation, and pointed out that the establishment of the Ultra Ethernet Alliance last year was a sign of industry recognition of these problems.
At the same time, InfiniBand has connectivity issues at certain distances. Although superior to Ethernet in terms of loss, Spelman said, “It was not designed inherently for large-scale and highly parallel workloads.”
Due to the challenges posed by these two established solutions, there is a market opportunity here, “she explained. What we have done is to build a network with all the necessary features to achieve scalable AI. That is to say, a lossless, high-performance network with congestion management and dynamic channel expansion capabilities. In other words, Cornelis has provided all the features that the Ultra Ethernet Alliance is attempting to achieve to provide an ideal network.
We are not an existing Ethernet provider, trying to gradually improve and add new features to Ethernet. We are defining the entire network and finding ways to be compatible with existing technologies, “Spelman said.
When asked about the market that Cornelis is pursuing, Sameh Boujelbene, Vice President of Dell’Oro Group, told Fierce, “The AI backend network is a brand new market that emerged to connect GPUs/accelerators across multiple racks and sometimes even data centers to form large AI clusters
She added that the analysis company predicts that the revenue of AI backend networks will grow at a compound annual growth rate of over 50%, exceeding $20 billion by 2028.
“The AI backend network represents a rapidly growing and highly attractive new market opportunity. In addition, they are driving new requirements, paving the way for significant innovation and disruption,” said Boujelbene. “The influx of large amounts of capital and investment into the industry further amplifies this trend.”
Gartner Honorary Vice President Andrew Lerner also believes that AI has created new opportunities in the market. Why? “Mainly a combination of two factors: first, the requirements of AI architecture are vastly different from traditional x86 compute oriented workloads; Secondly, the demand for enterprises to deploy GPU clusters locally is becoming increasingly urgent and urgent.”
“It is crucial to address these requirements in order to maximize GPU utilization, as GPUs are by far the most expensive component in AI. The cost of a GPU cluster can easily reach 5 to 20 times, or even hundreds of times, the required network switch cost. Simply put, in GPU intensive environments, the task of the network is to keep GPUs busy,” he added. “Therefore, we see that organizations are exploring new suppliers and methods to address this challenge.”
Spelman said she knows it will be very difficult to compete with companies like Nvidia (which not only offers InfiniBand, but also Ethernet based Spectrum-X products) and Broadcom. But she pointed out, ‘Although we are small in scale, it is precisely because of our small scale that we can respond quickly and be flexible, while we focus on simplification.’ This also includes launching customized products that integrate customers’ unique intellectual property as needed.
“We do not pursue meeting all the needs of all our customers,” she concluded. “We need to be the truly indispensable choice in the hearts of a few key customers.”