Cornelis has raised $205 million and unveiled Active Compute Fabric, an open networking architecture that adds programmable computing capabilities directly into scale-up and scale-out networks as the company expands deeper into infrastructure for large AI and high-performance computing systems.
The financing will support Cornelis’ entry into scale-up networking, development of its next-generation product roadmap, increased production, customer deployments and go-to-market expansion.
Cornelis is also collaborating with Qualcomm Technologies around the role networking can play in future rack-scale AI data center designs.
The Active Compute Fabric architecture is designed to make networking an active part of the compute system rather than simply a mechanism for transporting data between accelerators.
As AI clusters become larger, GPUs and other accelerators can spend significant amounts of time waiting for data or for other processors to synchronize.
Cornelis’ architecture combines lossless transport, in-fabric acceleration and programmable compute, allowing the network to perform operations on data while it moves through the system.
The platform can adapt to changing workloads, offload collective operations and take on new functions as AI algorithms and software evolve.
Cornelis believes that performing more work inside the fabric can improve utilization of expensive accelerators and reduce wasted computing capacity.
Based on company modeling of public industry data, Cornelis estimates that in a 100,000-GPU system roughly half of GPU hours can be spent waiting for data.
At an assumed $4 per GPU-hour and 8,400 annual operating hours, the company estimates that could represent approximately $1.68 billion of wasted annual compute capacity and around 500 GWh of power.
Those figures are based on modeling and pre-production assumptions rather than measured performance of the company’s next-generation systems.
Active Compute Fabric is being designed around open standards, including UALink and ESUN for scale-up networking and Ultra Ethernet specifications for scale-out infrastructure.
That approach is intended to preserve customer choice across different accelerator platforms rather than tying the architecture to a single compute vendor.
Qualcomm Technologies is working with Cornelis as both companies examine how networking architecture can improve utilization and economics as AI infrastructure moves toward rack-scale systems.
Cornelis said the open-standard scale-up and scale-out AI networking market could represent more than $55 billion of opportunity by 2030.
The company already has products entering deployment.
Its CN5000 platform is shipping, while CN6000 is sampling with customers ahead of expanded availability expected during the fourth quarter of 2026.
Cornelis technology is already used for AI and HPC workloads across hundreds of data centers spanning commercial, academic, government and cloud environments.
The new funding will help the company extend that scale-out foundation closer to accelerators inside AI racks as it competes for a larger role in next-generation computing infrastructure.
KEY QUOTES:
“AI infrastructure is reaching a point where faster endpoints alone are not enough. The fabric has to become an active part of the compute system.”
“Customers want complete rack-scale solutions without being locked into a single vendor or architecture.”
Lisa Spelman, CEO of Cornelis
“Improving utilization and AI economics will require a more integrated approach across compute, memory, and networking, and Cornelis’ vision for an open, programmable fabric aligns with that industry direction.”
Tony Pialis, Executive Vice President and General Manager, Data Center at Qualcomm Technologies
“Open-standard scale-up and scale-out networking for AI represents more than $55 billion of opportunity by 2030, and we believe the network will decide how much of the total AI build-out delivers real return.”
Joel Whitley, Partner at IAG Capital Partners
