WAYNE, PA — Cornelis raised $205 million to expand into scale-up networking and bring a new generation of AI infrastructure products to market, backing a push to make networking hardware perform compute functions that the company argues can reduce costly accelerator downtime.
The funding will support production expansion, customer partnerships and go-to-market efforts for the company’s next generation of scale-up and scale-out networking products.
Cornelis also introduced Active Compute Fabric, an open architecture that combines scale-up and scale-out networking with programmable compute embedded in the fabric. The architecture marks the company’s entry into scale-up networking and extends technology already used for artificial intelligence and high-performance computing workloads in hundreds of data centers.
The approach is designed to allow networks to perform operations on data as it moves through a system rather than function solely as a transport layer. Cornelis combines lossless transport, in-fabric acceleration and programmable compute to offload collective operations and adapt networking functions as workloads change.
“AI infrastructure is reaching a point where faster endpoints alone are not enough,” Chief Executive Officer Lisa Spelman stated. “The fabric has to become an active part of the compute system.”
Cornelis is targeting a growing infrastructure bottleneck as AI clusters become larger and communication and synchronization demands consume more accelerator capacity.
The company estimates, based on its modeling of public data, that roughly half of GPU hours in a hypothetical 100,000-GPU system could be spent waiting for data. Cornelis values that unused capacity at about $1.68 billion annually and estimates it would consume about 500 gigawatt-hours of power, equivalent to the annual electricity use of nearly 48,000 U.S. homes.
Those figures are company estimates rather than measured performance results for Active Compute Fabric.
The architecture uses industry standards including UALink and ESUN for scale-up networking and Ultra Ethernet specifications for scale-out networking. Cornelis also designed it to support multiple types of accelerators rather than a single vendor architecture.
Qualcomm Technologies is supporting the broader push toward networking as a more integrated part of rack-scale AI systems. Tony Pialis, executive vice president and general manager of data center at Qualcomm Technologies, is scheduled to join Spelman during her keynote at AI Infra Summit.
“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,” Pialis stated.
Investor IAG Capital Partners estimates open-standard scale-up and scale-out networking for AI represents more than a $55 billion market opportunity by 2030.
Cornelis’ CN5000 networking product is currently shipping. Its CN6000 is sampling with customers, with expanded availability expected in the fourth quarter of 2026.
Support the local news that supports Chester County. MyChesCo delivers reliable, fact-based reporting and essential community resources—free for everyone. If you value that, click here to become a patron today.