Cornelis Networks Raises $205 Million to Put Computing Power Inside the Network Itself

Cornelis Networks, a maker of networking hardware for artificial intelligence data centers, said this week that it has raised $205 million and unveiled a new networking architecture designed to reclaim processing capacity that the company says is currently going to waste inside AI clusters. The announcement, made at the AI Infra Summit and detailed in a press release published by the company, positions the Wayne, Pennsylvania-based firm as one of a handful of challengers trying to loosen Nvidia's hold on the networking layer that links thousands of AI accelerator chips together inside a data center.

The new architecture, called Active Compute Fabric, embeds programmable processing directly into network switches and network interface cards rather than treating the network purely as a pipe for moving data between chips. Cornelis said the design performs operations on data as it travels between accelerators, including offloading collective computing tasks and adapting network behavior in real time to a given workload, functions that today are typically handled by the GPUs themselves or left undone.

The Numbers Behind a Costly Bottleneck

Cornelis said its own modeling indicates that in a data center running 100,000 graphics processing units, roughly half of the available GPU hours are lost because chips sit idle waiting for data to move across the network rather than actively computing. The company estimates that gap costs data center operators about $1.68 billion a year in wasted capacity and consumes roughly 500 gigawatt-hours of electricity that produces no useful work.

The $205 million round was led by IAG Capital Partners. Joel Whitley of IAG Capital Partners said in the press release that open-standard scale-up and scale-out networking for AI represents more than $55 billion of opportunity by 2030. The capital will fund production of Cornelis's two current switch lines: the 400-gigabit-per-second CN5000, which is already shipping to customers, and the second-generation, 800-gigabit-per-second CN6000, which is sampling with customers now and is expected to reach broader availability in the fourth quarter of 2026.

From an Intel Cast-Off to a Nvidia Challenger

Cornelis traces its roots to Intel, where a team of engineers built the Omni-Path interconnect technology used in high-performance computing. When Intel decided not to continue developing Omni-Path internally, that team spun the technology out as an independent company in 2020, raising an initial $20 million from Intel Capital and other investors to keep the architecture alive. The CN6000 chip can run in either Omni-Path mode, preserving features such as congestion management and credit-based flow control, or in Ethernet mode using the RoCEv2 protocol, allowing it to pair with third-party switches, including Broadcom's Tomahawk line, depending on what a customer's data center already has in place.

That flexibility is central to how Cornelis positions itself against Nvidia, whose NVLink and InfiniBand technologies dominate the market for linking accelerator chips but generally work best, and are most tightly integrated, within Nvidia's own hardware and software stack. Active Compute Fabric instead relies on open industry standards for the two types of connections used in AI clusters: UALink and a protocol called ESUN for linking chips within a single rack, and the Ultra Ethernet specification for linking racks together across a data center. Cornelis argues that reliance on open standards lets customers mix accelerators from multiple vendors rather than being locked into one supplier's ecosystem, an argument that has drawn coverage from TechCrunch and other outlets tracking efforts to chip away at Nvidia's dominance of AI infrastructure.

Data Center Operators Are the Intended Buyers

The customers Cornelis is courting are the operators of the large GPU clusters that train and run AI models: cloud computing providers, enterprises building their own AI infrastructure, and government and academic research centers, according to comments from the company reported by Network World. Alongside the funding, Cornelis announced a strategic collaboration with Qualcomm Technologies focused on evaluating how Active Compute Fabric might connect to Qualcomm's forthcoming data center accelerator chips, an alliance that signals interest from chipmakers seeking alternatives to Nvidia's networking stack as they try to compete for a share of the AI infrastructure market.

Chief Executive Lisa Spelman framed the pitch around the gap between what customers pay for GPUs and what they get out of them.

"We're giving you your GPUs back. You've bought all these GPUs, and they're being used at about a 50% rate. At that utilization, with a better, higher performing active network that has compute in the network, you can drive that GPU utilization up five points, 10 points," Spelman said, according to Network World's account of the announcement.

In materials distributed alongside the funding announcement, Spelman said the industry has reached a point where faster endpoints alone are no longer enough and that the network has to become an active part of the compute system. Cornelis Chief Marketing Officer Brandon Draeger, in comments carried by SiliconANGLE, said the payload does not arrive the way it left, a reference to the delays and inefficiencies the company says its in-network processing is designed to eliminate. Cornelis has also described early simulations showing the architecture can cut network traffic by as much as half in pre-production testing, though those figures have not yet been validated in large-scale commercial deployments.

What Happens Next

The near-term test for Cornelis will be converting sampling relationships for the CN6000 into paying customers ahead of the chip's expected fourth-quarter release, and demonstrating that Active Compute Fabric delivers the utilization gains it is promising once it is running production AI workloads rather than simulations. The Qualcomm collaboration remains an early-stage technology evaluation rather than a committed product roadmap, and Cornelis continues to compete not only with Nvidia's proprietary networking but also with established data-center networking vendors such as Cisco Systems and Arista Networks, both of which have been expanding their own AI-focused product lines. Cornelis said it expects to publish additional technical detail on Active Compute Fabric's scale-out capabilities, which extend beyond a single server rack to link multiple racks of accelerators together, in the coming months as the CN6000 nears general availability.