Cornelis Networks Launches Active Compute Fabric, Raises $205M and Partners With Qualcomm

The startup develops specialized, congestion-free data center networking systems for AI and high-performance computing and competes with Cisco Systems Inc. and Arista Networks Inc. The Active Compute Fabric is an open architecture that integrates programmable compute directly into the network fabric. Cornelis said rapid growth in AI model compute cluster sizes has created data center network bottlenecks, leaving expensive AI accelerators idle while they wait for data. CEO Lisa Spelman said compute, memory and storage have become more workload-aware, but traditional networks have not made the same change.

According to Cornelis, the architecture combines in-fabric acceleration with lossless transport and programmable compute so the network can work with data as it traverses the system. It can adapt in real time to shifting workloads, offload complex collective operations and take on new functions. It is built on open standards including Ethernet and UALink for scale-up and Ultra Ethernet for scale-out, allowing organizations to deploy it with existing compute architectures.

Chief Marketing Officer Brandon Draeger told SiliconANGLE that Active Compute Fabric performs workload-specific operations as data moves through the network rather than merely acting as a messenger. He said it can assemble KV cache data for disaggregated inference, coordinate expert dispatch for mixture-of-experts models, accelerate collective operations such as AllReduce and compress gradients in transit before data reaches its destination.

“The payload does not arrive the way it left,” Draeger said. “In a collective operation, partial results from many endpoints are combined inside the fabric, so a single reduced result lands at the destination instead of thousands of separate contributions. Compressed gradients move as a fraction of their original size. The work happens once, in the path the data was already taking, instead of consuming accelerator cycles at both ends.”

Cornelis said those capabilities offer two advantages: less data must traverse the network, and GPUs spend less time waiting on communications and synchronizing arriving information. According to Draeger, pre-production simulations show the approach can reduce overall network traffic by up to 50%. He said accelerator utilization in large AI deployments commonly sits near half of installed capacity, and the architecture is designed to return a meaningful share of what is currently wasted, though improvements will vary by model, cluster size and customer stack.

Spelman said Cornelis is expanding into scale-up and scale-out so it can bring its network architecture closer to the AI accelerators that power AI workloads, feed them data faster and improve performance. “AI infrastructure is reaching a point where faster endpoints alone are not enough,” she said. “The fabric has to become an active part of the compute system. We’re seeing growing demand from customers for an open alternative that gives them more choice in how they build their AI infrastructure, and that approach is creating real momentum for Cornelis.”

Qualcomm will join Cornelis on stage at the AI Infra Summit in Santa Clara to announce a strategic collaboration aimed at shaping networking for rack-scale AI data centers. Both companies see existing passive network fabrics that only move data as a problem that causes AI accelerators to be underutilized, and they believe networks should receive more priority in the design of AI data center infrastructure.